In this conversation, we explore how European companies are actually approaching AI implementation. Rather than racing to deploy the latest models, many are focused on organizational design, employee adoption, process changes, and measurable returns on investment. Alex explains why European firms tend to be more cautious than their Chinese counterparts, how concerns around AI sovereignty shape technology decisions, and why companies increasingly find themselves balancing U.S. frontier models, Chinese cost-efficient models, and European alternatives such as Mistral AI.

We also discuss the economics of AI adoption, including the emerging concept of “tokenmaxxing” or rather if that is even the wise path forward, whether AI is truly replacing jobs, how companies should think about ROI when AI introduces variable costs, and why the future may involve token budgets becoming as commonplace as mobile data plans. Finally, we explore Europe’s position in robotics, industrial AI, and regulation, and whether Europe’s strength may ultimately lie not in building the largest and best-performing models, but in defining how AI is deployed responsibly at scale.

Subscribe now

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, early adopters, and product managers. For more information on the podcast series, see here.

AI-generated transcript (for reference only)

Grace Shao (00:01)

Hi Sheng Yun. Thank you so much for joining us today. Really excited to have you.

Alex Lu (00:05)

Yeah, thanks very thanks for inviting me. I’m also very excited to have this conversation with you.

Grace Shao (00:11)

Yeah, awesome. So tell us about your journey. I think you’re in a pretty unique position. You know, like I said in the intro, you know, a lot of the conversation about AI right now is often positioned between China versus US, But you actually work predominantly with European companies in adopting AI and their digital transformation. So tell us about your your background and how you got into this.

Alex Lu (00:31)

Yeah. so thanks a lot. So actually, I went to France. I spent more than 10 years in France. I went to France in 2004 and I studied in a school called Ecole Polytechnique. and then when I graduated from the school, I started my work in in Europe, mainly for automotive industry and afterwards for the consulting industry. And still when I was in the consulting industry, I worked mainly for for the auto sector. So

I have a very traditional background of automotive. That’s why some of the work I’m doing currently in the in the AI, we can come back on that, is in the automotive manufacturing sector and mainly for European companies. Because I started my career in Europe, so I know I don’t I know them pretty better, pretty good. And the the the other thing point I want to mention is the school I started actually the Ecole Polytechnique was

Let’s say it it was a famous school in France or in Europe, but it it’s not so famous in in the world. actually this is in France they have a different educational system. but still with the with the rising of of AI in Europe, especially the French large language model called Mistral AI, the school becomes famous because the founder of the of of of Mistral AI comes from the the same school. So basically it’s also a a little bit like

Tsinghua university in China is like the the Tsinghua in in France, having the best talents for for the AI. So nowadays, when I continue my work in the AI transformation for companies or AI implementation for the companies, I work a lot with European companies. Firstly, I know that my I as I said before, and secondly, is when we look into the global competition between China, US, and Europe.

In the AI landscape, it’s pretty clear. It’s like China and and US or US China being the tier one or first ranked models. And Europe is kind of lagged behind. So most of the European European companies, they have this kind of attitude of being a little bit complex, I would say. on one hand, they are kind of seeking for, of course, for the best technology in the in the world to enhance their company’s competitiveness.

There comes the question, how I can define my AI strategy for next year’s between Chinese and US tech stack in AI. And the second question they raised often is while we are European companies, we want to keep keep our AI sovereignty, which is a very important topic in AI. again, we can come back on that. So their question is: okay, between this US and China tech race.

Is there any place for European companies regarding the foundation model companies or application companies or even corporate clients? What could be the playground for European companies? So these are major two questions are often received from European companies and you will you can see the thinking angle is they European companies want to at the same time keep it keep the AI sovereignty and at the same time keeping their competitiveness. That makes the question a little bit complex. Yeah.

Grace Shao (03:49)

Actually why don’t we just double click on the unpack that a little bit? What’s your view on it? Like what what do you advise your clients to do then if if they are kind of cut caught in a pickle or unsure how to build out the next stage of their infrastructure kind of being caught in between China and the US?

Alex Lu (04:08)

Yeah. So the the first thing I I always shared is in in in this tag race actually China and US we are not I want to twist twist a little bit the angle saying this is a competition between China and US. Actually, if we look into details, actually China and US are taking different directions in terms of the AI development, if I can say, because let’s say if we look into the US,

AI ecosystem or the AI development. I think a lot of efforts are put on the foundation model or kind of foundational research regarding how AI can be become AGI can be bring beneficial benefits to the humanity, or how we can guide Rails AI so that okay, one day we will not go into the direction of science fiction movies. So this is a little bit the the push from the US AI companies. While in China, actually

the ecosystem or from the national perspective, China’s AI is more about applications and more about how we can have the s beneficial from the whole society from the AI and how I can combine AI with my traditional technologies or traditional business to to to to to grab more values. So if we think in this angle, actually it will give us two different pictures. One is we cannot say that it’s kind of from

front to front front competition, because these two nations are just take different angles. The second thing is if we look into details based on these assumptions, we will say one nation is pursuing having the most advanced AI technology and one nation is pursuing most kind of most beneficial AI for the society regarding cost effectiveness, et cetera, et cetera. So then it comes to the question that you raised for European companies is

We always brainstorm and conclude on the simple question is what kind of AI are we looking for for European companies? Are we looking for, let’s make it simple, I take some an analogy. Are we looking for kind of you need all the employees to be the PhD employees having the most intelligence in the world? Then that will be the US foundation models. Or if we want to say we have the most cost efficient and best performing employees, virtual employees in your company.

Then we might consider Chinese models, foundation models. Then this is the trade-off. I think the companies should figure out. And the answer will not be so simple like that, saying, tomorrow I will switch to all US tech stack or Chinese tech stack. I would say the two ecosystem, as in the past in the digital area, will still continue for European companies, meaning that they need to juggle with Chinese tech stack in certain markets.

maybe in Chinese market for sure, but for other markets, developing markets where the Chinese foundation model are taking influence and as well as with US models. So this is the thing. And I think the other angle answer to to to their question is I I usually take the statement from Jensen Jensen Huang saying the AI is kind of five layer cake.

So what we are talking about is only one layer, which is the foundation model. And if we go deeper, then we will have infrastructure like data centers, like powers, chips, and electricities. And if we go upper, we will have the applications. So I would tell European companies or I told European companies often is I think the use cases in Europe makes a lot of sense because the cost is there and the employee was pretty

much expensive than Chinese employees. So if we deploy the same model, let’s say, and it of cost the ROI return on investment, you make the business case very easily in Europe than in China because the labor cost is kind of lower. And the the advantage of Europe, one I would say one of the advantages is about power and electricity. I study in France and in France y you would see they have the most advanced nuclear nuclear power technology in the world, at least in the past. And

I think the French government is also think about how we can build more power plants in the in the country to support Mistro’s AI development. And I listened to the founder of Mistral AI, Arthur Mensch. he explained to European Commissions how we can keep the AI development in the Europe is just he make a very simple analogy, meaning that intelligence equals to token. So we all know that, and he said token equals to electricity.

So if we want to make our society more intelligent in Europe, then we need to build infrastructure and more efficiently and more sustainably. And and my last point is I looked into the report released by Stanford, the HAI index. And very interesting because US is far away beh in advance compared to other countries in terms of the number of data centers, it’s around

2000 or more than 2000. I I didn’t remember the exact numbers. And the second and third, it’s not China in terms of the number of data centers. Based on the data, it’s United Kingdom and Germany. So Europe, Europe has the capability to build power plants, but I I I tend to believe these power plants are not currently used to train US models, in my opinion. So again,

This is the European competitiveness if we want to talk about AI. So if we enlarge a little bit picture, we say that okay, it’s five or five layers cake, then again, China and maybe is better performing better in terms of electricity, maybe a little bit less performing regarding the chips. Same situation for Europe. So it’s not only about the most performing models, right?

Grace Shao (10:13)

That’s an a really interesting take. So help me understand like what do you actually advise on these companies for? So you gave me a big high level picture, right? Well give me some examples on the kind of work you’re working on. It’s because I think on this podcast, we often invite people who are builders, founders, investors, and they give us a lot of high level views, which is great, right? But but I want to hear from you, how are you actually helping companies go through this AI transition? And

What are the bottlenecks? Maybe further down we can talk about that. What are the challenges? What are the exciting areas? But just help us understand what are the day to day tasks that you’re working on.

Alex Lu (10:49)

Yeah, thanks. again, I I might share two different perspectives from my experience when it’s again with the European companies. It’s very interesting example because actually I build a product doing this kind of market intelligence, market research for European companies and the value proposition at the time it at that time is we can save time for your employees and they a and and we make your organization more efficient.

And basically we find when we when we s when we sell this kind of value proposition to different companies, I see very interesting different answers. One is on the European side, he would say, this is very interesting, but before implement implementing, we need to think about a kind of tomorrow’s process process, meaning that if we put your AI product into our organization, how our employees will work with the product together, what’s the process look like?

And how many new skills would my employees need to perform or to better use your products? So I think the European company’s mindset is they will need some time to conceptualize the AI products or AI use cases. And then they will need to conceptualize and project, especially once the product is in place, what my company will look like. So they will spend

A little bit more time than Chinese companies to figure out the the regarding the talents, regarding the organization, regarding the process. And in my opinion, it might be a right right approach, in my opinion, meaning that they put human before the techno technology. And this is what I observed when I implement AI for companies saying that, okay, I bring you the best technology, but often we might have improved efficiency by 10 times or five times.

In one single process, but actually there will be some bottlenecks in other organizations, in the in the rest of the organization, then you cannot you cannot increase the efficiency of the whole workflow, let’s say. So that’s why a lot of people in US they talk about AI native organization to kind of remove the bottlenecks in the in the organization. And I worked another example, I worked for a European company, and it’s very interesting. He said, I receive

high level management. He said, I received so many reports from my employees, I I don’t have enough time to review and to approve them. That’s the case because we increase efficiency of the working level of the people, then the bottleneck becomes suddenly the a the leadership. And then we need maybe an empowered leadership by AI in the future to make the whole organization more efficient.

Or we need to think about a new organization where we include AI agents and human beings together because the two natures are producing things on a different scale. So this is mo most of the time, this is the European companies. And for the Chinese companies, the mindset is totally different. if we implement the AI solution, the same product to a to a Chinese client.

The the the answer would be, that’s very interesting. You save 20 or 30% of my employee’s time, but you know I cannot let’s say lay off the employee and to make some savings. So just tell me for the time we saved where he can work to produce more. So it’s always in the mindset, okay, we have some some time saving, but you I cannot pay 80% of the salary to the same guy. So

In in order to so I I need to pay him hundred percent salary, so I y your business case doesn’t work for me. So where where we can grab more values. Yeah, so you see a di

Grace Shao (14:37)

That’s really interesting. That’s a really interesting approach.

Yeah, because it’s like the company is reflecting actually a broader, I think, social, even cultural and perspective on how they’re perceiving AI. And in the China and US, often the conversation is so fixated on improving efficiency and people who are utilizing AI are actually more burnt out because they are like 10Xing themselves or whatever these days.

Alex Lu (14:49)

Exactly.

Grace Shao (15:03)

But you know, in Europe, that conversation is so different. And you can say maybe much more humane. However, like you said, the bottleneck right now is then how do these companies become the next generation? Like still relevant in the future, once this becomes normalized. So it’s interesting. I wanna go back to that a little bit later as well. I I wanna touch on something before we get further into I guess the comparison of you know, the adoption and everything is

Alex Lu (15:17)

Yes.

Grace Shao (15:30)

You and I met each other essentially online because I found out about your work that you helped a lot of European companies adopt Chinese model. I found that was very, very fascinating, right? you advised them on how to basically integrate, say, the Minimax and C AI of the world. Now, a lot of these model companies, when I speak to them, they say their priority right now is to basically sell globally. And of course, the Western markets are some of the most lucrative markets. the US.

headwinds are mostly in geopolitics, compliance, Europe. How do you view that as a market for them or opportunity for them? Like, is it equally challenging for these companies to sell to enterprises in Europe or do you think there are more opportunities for them right now and they’re they’re kind of taking off a bit more?

Alex Lu (16:16)

Yeah. I I I I if I think the conclusion if I I can state at the very beginning is kind of in Europe definitely there are more opportunities than for Chinese large language model companies than in in Europe, than in US, sorry. I have two proofs for that. Firstly is I discussed with a a CTO who is also a schoolmate from my school Polytechnique.

And actually he was very interested also by my newsletters on LinkedIn and then when they he asked me the question so apart apart from Deep Seek, what are dip other models and Chinese models that we we that we can use to improve our efficiency? Because when we meet all of he said he told me when we meet most of the IT implementation companies, they came with the solution like Anthropic or ChatGPT or OpenAI or or Google Gemini.

So we don’t see so many options. He mentioned the word options of Chinese models. And we know that Chinese models are more cost efficient. And then we can talk about token mapping. I think it’s kind of related topic. So this is one thing. I think in Europe, actually for the companies from the business perspective, they are also looking for different variety of different models so that they can bring what I said before, a best cost performance ratio models in the in the organization.

So this is one thing. And then I I told him that most of the Chinese large link model companies, firstly they started their business in China and then they tried to inf have the global influence. Like the most advanced one is zero dot zero one dot AI, but the other ones they are trying to catch up, like that AI you mentioned also Minimax. so I I think the thing is what I see today is the ecosystem of Chinese models.

Are not currently penetrating into the European markets. But definitely there’s a a room for Chinese players. The second thing is I always take the comparison with the other industries like EV industries, like car industries. because you you will see Europe put a lot of tariffs on Chinese vehicles. because okay, you you see a lot of Chinese vehicles because of

Europe wants to protect their own industries, et cetera, et cetera. But at the end of the day, they are not putting hundred percent tariffs. They are putting somehow reasonable tariffs on the Chinese vehicles. So the bottom line I want to mention is I lived in Europe before and I know the mindset of Europe European people. The mainstream, of course, we have different views. I think the mainstream for European people, most open ones.

Are saying okay, we need fair competition. The EV cars is just because okay, the European commissions are claiming that okay, you produce in China, but we in Europe we produce in a more sustainable way, so our cost is higher, blah blah blah. So if we take this comparison, I think definitely there will be some places for Chinese companies in condition that we play fairly in the European market.

And then we might come back to the third point I mentioned before, of course, there’s a a point of AI sovereignty. the biggest, the biggest player of European AI ecosystem is still Mistrol, so it’s the biggest player in the foundation model. And of course, Mistrol should be one of the choices options when we suggest to European clients as the large language models.

So I would see that if tomorrow the Chinese model enter these European markets, they will face a fierce competition with Mistro because Mistro basically they have a government back, let’s say, from France, and they have a very good positioning in the ecosystem. you would see in in two or three weeks you’ll there will be a VivaTech in France and Mistrol for sure they will be on the stage and for sure they will be

French or German presidents, French president and German chan chancellors. And with their unique positioning, I think most of the European companies they were firstly considered Mistrol, but still if Chinese companies can bring something on the table, business wise, the European companies will not only limit to only one model, there will be some balance between different models. And today, the balance I see is Mistral versus other US models.

Grace Shao (21:11)

No, I just think it’s really interesting ‘cause I think it also totally makes sense when I to talk to people who are in the Korean market or, you know, covering the Middle Eastern markets. sovereign AI is just such a top of mind like conversation for companies, whether it’s for compliance reasons, or regulatory reasons, whatnot. So it makes a lot of sense that Mistral’s position very well in Europe. However, are there any other players that maybe we’re overlooking outside because we’re not that familiar with the European market? Any other

foundational model labs that we should know of coming out of Europe.

Alex Lu (21:45)

Yeah, there will be apart from Mistro there’s another large language model whose name is H, but it’s less famous. And then you’ll have it’s not if we can say it is kind of word model by Yen Laquen, the ex researcher in Metafair, and he just came back to France and lab raised raised a large amount of money for the for for his model. It’s called AMI, yeah, AMI.

Grace Shao (22:01)

Mm.

Interesting. Okay, so let’s talk about the token maxing thing you touched on just now. So offline we talked about this a little bit recently. There’s been getting some buzz. It’s quite funny, you know, whether I’m it’s like big tech in the US or big tech in China. When I talk to them, people are saying, Okay, our managers are pushing us to token max. If we don’t basically use AI in our job and figure out ways to essentially replace ourselves, we get replaced, which is the irony in all of this. It’s it’s all kind of sci fi. but

Grace Shao (22:41)

Then the joke’s kind of been played now on the companies because you know there was just headlines coming out saying, this one guy basically spent like more than half a million dollars on tokens in a month, and that’s obviously more than his salary. And then companies are realizing, wait, this token maxing strategy is not cost efficient at all. So from an operational standpoint, I know you are someone who work a lot with companies to implement AI and find

Alex Lu (22:54)

Yeah.

Grace Shao (23:10)

the most cost efficient way for their for their operations, right? And not just costs, like you mentioned, it’s like a balance of costs, you know, and and operational sustainability as well as obviously company morale and everything. So how do we view this trend? Where is this going? Is this sustainable? Like just just give us some high level views on this.

Alex Lu (23:32)

Yeah, there there’s a a lot to talk about this, because the token is becoming really a trendy topic for individuals and as well for companies. so to answer your first firstly to answer your questions, I don’t think that’s sustainable. My view is the token mapping is kind of marketing for infrastructure companies. and of course, as you say, there’s a lot of people burn a lot of tokens and more than their salaries.

Then the question would be if I pay your salary or if if I pay your tokens. We’ll come back to this point afterwards. I discussed with some some Chinese companies. Very cost cons cautious. I think the the the thing is today when we actually for for the tech companies in China, there’s also some ranking of token consumed. but it’s kind of indicator of how people are use AI. But

It’s not is it the right indicator? I don’t think so. basically I think in the in the in current status we didn’t we didn’t find a very good metric to measure the performance of a human being empowered by AI. that’s the thing. So we take a kind of proxy indicator, which is the token for and of course there’s a lot of waste of token in in in in in the usage and I’m I’m not sure that every single employees

would be the master of AI if we don’t provide the sufficient upscaling in terms of the AI. Because from individual perspective, sometimes we use by coding, but if we don’t master the basics of coding, then we might waste some time and as well as some money and tokens in the by coding. So this is my view. So the token maxing is kind of marketing stuff and and the the day when we find out

Again, for the AI organization or for the organization, how we can measure the performance of individuals with AI, then we might have a clear picture and no longer token max. And the other interesting thing you you mentioned already, but I read also is Microsoft they are kind of switched to their copilot because ever s if everyone used used the entropy cloud model then become too expensive for the whole organization. It’s just not just not cost efficient.

And brings me to my point is when discussed with some Chinese companies. So, you know, Chinese companies are very cost conscious. And they are thinking is I think that it was a joking, but this is right angle of thinking is show we in the salary of our employees to allocate a part of the tokens monthly for our employees. meaning that okay, if the

If in the in the in the past situations hundred percent of the salary tomorrow might be eighty-five percent of yesterday’s salary plus fifty percent by tokens. And the tokens you can you can use and if you don’t use tokens efficiently then it’s the savings for the company. So this is it’s

Grace Shao (26:43)

That’s really crazy. But I kind of see what you mean.

Like so essentially it helps you with your job. So that’s why it’s on you. But then what if you just don’t w but what if you don’t want to use AI? What if you just like I can do my job perfectly fine the way I did it before and I don’t want to token max and I want to keep my hundred percent?

Alex Lu (26:50)

Exactly. That’s the question that the Chinese company needs to answer, but you reflect on your point mentioned that the token consumption is sometimes much more expensive than the salary. So it causes Chinese company companies to think that okay, I spent salary, I spend tokens for the intelligence, I spent two times to hire employee. So why not combine them together and doing kind of tomorrow’s package is your basic salary plus tokens?

Grace Shao (27:32)

So actually on on that,how should companies think about it then? Because, you know, it’s really easy to say, okay, this is an AI native company. There’s 20 people in this company. Everyone’s token maxing because it does bring the 20 people’s efficiency to say like 400 people, whatever it is, right? However, what about the traditional companies, especially the ones that you work with? Like a lot of them are OEMs, manufacturers, you know.

It it doesn’t make that much sense for them to really jumping on this AI bandwagon as well then. Or how do you advise them then? Or how do you think how should they think about it?

Alex Lu (28:06)

Yeah. I I think for the for the traditional companies or European companies, it doesn’t make sense for everyone to give the token maxim because as I said, I’m pretty aligned with the European approach saying that okay, in order to release or unlash the value of AI, we need at least to upskill a little bit our employees. We cannot expect employees like with thirty ex years experience in the industry and tomorrow he switched to a kind of AI expert in the

in the in the in his company. So I I just want to combine our question with my previous comment saying that today if you look into the Chinese market today there are some big big traditional telecommunication companies like China Mobile they are proposing the token plan for individuals it’s like your smartphone monthly monthly plan yeah

Grace Shao (29:02)

Wow. Like data plan.

Alex Lu (29:05)

It’s a kind of data plan, exactly. So the token is becoming kind of infrastructure like electricity, like water, or like your smartphone, monthly subscription. So this might be the way the companies might pursue, saying that, okay, yesterday I might give you a kind of monthly plan for your telephone. So I can reach out to you and you can read the emails and you can use the telephone to walk with emails, work teams or with Zoom, etc. etc. And tomorrow it might be a com kind of monthly subscription.

For different employees, then you have a monthly token plan you can use for your personal, not for professional work in AI. I guess that might be the way that the China might be moving for individuals and for companies. And again, for European companies, they are not there yet, but when I discuss this vision and this kind of trend, and they are pretty interested, they might be moving in the same direction.

And for the companies, at least not at the national level, but at the company level, to provide kind of a monthly subscription to a limited number of people who master AI, and the first wave of people adopting AI is their coding team, their IT team, their digital team. So they will be the first employees to use this kind of concept of monthly subs subscription to tokens.

And of course, for manufacturing companies, there’s a lot of people working in the factories, in the plants, or or in the on the production lines, and they are not be impacted, they will not be impacted by this kind of AI wave. But still, I think the things are are moving slowly and it’s it’s changing so quickly. but this is currently my discussion with European companies.

Grace Shao (30:48)

That’s actually very interesting. I it makes a lot of sense actually to build it in in as like a infrastructure like 5G data. And then it’s really, it’s really like there’s a cap on how much the company will pay for, but then how you utilize it should be, and you’re more mindful of how you’re utilizing this, right? And not wasting the tokens and and buy the that thus you know, wasting your energy, compute everything. So

Grace Shao (31:13)

I want to bring it back to the Chinese pricing models really quickly. I know you work with a lot of European companies, they are the buyers essentially. You also help them connecting with the Chinese vendors, essentially, which are like the Chinese LLM labs, Minimax, Drupal, Moonshot, etc. Now, how should we understand the pricing model of these companies right now? Because it’s obvious that they are pricing themselves much cheaper to US peers.

Some might you know, obviously argue that their performance might not be as on par like on par or as at the frontier. however, even when they do play catch up, you know, the reflection of it is it just s seems like a complete different cost structure. Help us ex understand that, like they’re thinking, why they’re pricing it much lower and how that plays out in the long run.

Alex Lu (31:46)

Mm-hmm.

Yeah. Actually, there are two perspectives on that. maybe I will firstly talk from the client’s perspective and then I I might conclude with the recent price decrease by Deep Seek. maybe you you have already read about it. so f from the European companies as I said before, the thinking is w that the that that’s that’s the statement for the companies I met. we do not need entropic models for all the time.

That’s for sure. Because this is very expensive even for a company. So for sure they will need a kind of different options from different models, like the best U US models and the cost-performing, the best cost-performing models from Chinese models and the AI sovereignty models like Mistro. So basically there will be three combinations and then there will be engineering of technical issue that meaning that how to manage these models to

Perform the right tasks. So, meaning use cloud to perform the most complex tasks and use Chinese models to perform kind of less complex tasks. And I think European companies they understand this. And they, of course, they are looking for Chinese models for the cost effectiveness. And I would say this is also one of the bottleneck of US models because they are very

In a relative way, very expensive. Therefore, it it’s the bottleneck of the massive adoptions. Only the European big, big companies can afford like continuous use of US models. well there are a lot of SMEs in Europe. So this is this is the thing. And for the Chinese model suppliers, I think the the way I I see the the the price issue is if you ask me

Can Chinese companies increase their token prices? I would say surely, because if you look into the financial report of ZAI or Minimax, actually they are not they invest a lot in the research to develop these models. And the expectation from the industry AI industry is if you want to train or pre-train a next model, you will cut it will be more costly than the previous pre-trainings. so for sure.

Chinese companies can increase their token prices. And w that that’s what they are doing actually after the open cloak, if you read into the news. And the thing is, compared to the US model, still the Chinese model are very cheap. I think there’s one very strategic thinking thinking angle is if you think about the Chinese models, most of them they are open source models. And the the the the thinking angle is, I think, for the Chinese model players is

We want we open source these models because we want people to use these models. Because they can deploy it on their own infrastructure, they will have more freedom, or they can use our open source model to train their own models. and maybe they they will use our our tokens by or they will they will understand or know our models better by open sourcing. So if if we combine this thinking angle, I would say.

The Chinese model strategy might be to increase the influence in the world, maybe in the developing markets, where people are more cost conscious, and to help people to use this AI to adopt AI in a cheaper way. And then in the long term, in the future, that’s very Chinese, maybe again to increase the prices once we take the market positioning.

It’s like the price competition for the last decade regarding this digital sharing economy or digital era. Nothing has changed. So a very aggressive c pricing strategy to at least to to have the market share and then once we have the market share then we can establish our our our our position in the market and ca kinda do a lot of monetization stuff.

That’s the one thing regarding the increased influence globally and taking the lead in the AI industry for the developing countries, in my opinion. Of course, go going to Europe is is part of the their strategy. So this is from the Chinese model’s perspective perspective, and it’s a very special case, of course. It did this is Deep Seek. DeepSeek released just the before and right after the release, during one month, I think for the developers we enjoy the

75% of discount regarding the token price. So it’s very deep discount. And recently, I think one or two weeks ago, DeepSeague announced that they will keep this 75% discount for for for forever. So it’s kind of they they just discount their token prices by such huge amount of discount. I’m pretty surprised. and

Again it di it it launched a price war in the market and you see recently Xiang Mi decrease also their token prices and I don’t know if other players will will follow in Chinese market at least. But if we think about Deep Seek cases, it’s a very special case because Deep Seat this year it doesn’t create a lot of buzz in the AI community in the US. I think so. I I’m not living in US but I read some newses. news, sorry. I think the

nowadays DeepSeek, I’m not saying that we have the best performing model. and and and in terms of of the tok coding performance, DeepSeak is is not at the top top level compared to other models. But the interesting thing is DeepSeag this time is trained on the Huawei ASEAN chips. so again, I think the price decrease of DeepSeag combining with their recent news of raising money and hiring some harness engineering

Across the world, I would suspect that DeepSeek by decreasing their prices, they just want to break through the ecosystem established by NVIDIA. This is my thinking, and and that’s why after the President Trump visit to Beijing, there are 10 Chinese companies are not authorized to buy Nvidia chips, but up to now you see few others.

Grace Shao (38:09)

That’s interesting.

Alex Lu (38:23)

I think there’s a thinking from the national wise from from the nation thing that okay with Dipsy can we break through the Nvidia chips plus CUDA? And if because that’s so cheap, so most of people they might use Deepsi in the future and they might be used Huawei as ASN chips because Deepsi got trained on these chips and it’s best support DeepSeak’s performance. So this is another angle. Yeah, so you would see

Grace Shao (38:23)

Mm-hmm.

So the open source strategy. Sorry, go on. It’s basically a strategy

to get people in to get the developer into its ecosystem, its own community first, which is what Jensen’s been saying the whole time. Yeah. no, I I agree with you on that. I actually I I wanna and steer away from the chips today because I I am quite fascinated. So you work with companies, adopt AI, but how does that actually what are companies really using AI for? Like we hear about stories.

Alex Lu (38:54)

Exactly.

Grace Shao (39:16)

you know, companies are token maxing, whatnot. And obvious the obvious one, like you mentioned, is in coding capacity in IT, but no again, not every company is in tech, you know, not every company needs coping co coding capacity. sorry, let me just say that. Not every company needs coding capacity. So like what are we seeing actually on the ground, especially for maybe more brick and mortar stores or old school traditional industries? Why would people all want to adopt AI right now?

Alex Lu (39:47)

the the the adoption rate actually for European companies is pretty low, to be honest. most of companies, if we say at a large scale, they don’t adopt sufficiently AI and they just are afraid of missing out something. So this is a FOMO. they are just feared of missing out some opportunities, and if they don’t use AI today, they might be less competitive in the future. So the the f the most common use cases I see in

companies for coding and for it and sometimes it’s easier to measure the effective effective sorry effectiveness of ai that’s in the most most of the time in the sales marketing department so meaning that if you use ai you can produce produce more contents and with more contents you have more impressions with more impressions you might have more conversion rate you might have more conversions and you might have more sales revenues so

This chain is actually well formed. So by using AI, you can track the individual metrics on the chain, and then you can kind of monitor the results by using AI. And most of the time, I get a very simple question of European companies, and very difficult question actually to answer is: what’s the ROI of implementing AI? What’s my return? then it’s a very difficult question because in the

digital, 10 years ago in the digital era, I can tell the ROI, I can estimate why, because the incremental cost of using digital products is kind of almost zero. You just need your digital products and then it makes more efficient, it makes more automate. Well, in AI, that’s very difficult because if you think about it, if you use more AI, you will consume, as you say, more tokens. So, meaning that

An employees, you need to pay the salary. If he is a heavy AI user to produce more content, then you will need to pay his tokens bill. And then the ROI might not be so immediate. Or there might not be ROI actually for the individual use cases. Then we come back to the question: is okay, by using this AI, how we can make the whole organization more efficient and how we can generate more revenues for the whole organization.

While for the individual users, maybe there’s no business case. So I think again, the the the the the difference compared to 10 years ago is the people who use AI and who use heavily AI, then he will have a bill to to pay. That’s a variable cost. That’s very important. And secondly, is the variable cost will be really

The beneficial of the variable cost will really depend on the skills of each individual. You may pay $100 for employee A or employee B. If B master better AI, then you will have 10 times more results, financial results, compared to the first case. So again, I think you asked the right question. the ROI question is definitely a very good question. and most European companies they seek about ROI before investing. So they are very cautious.

While again, if we compare to the Chinese companies, we are more pragmatic. So let’s implement a POC. it costs a little bit, but let’s implement it. If it doesn’t work, never mind. We waste some money, but we we we continue, we iterate or we continue with another use cases.

Grace Shao (43:17)

So you think the Europeans are taking a more cautious approach, but actually more cautious on what the potential ROI is. Then I bring it to the question that is a bit more philosophical and like a societal, not so businessy, is then isn’t the headline or the mainstream discussion on AI is replacing our jobs completely overblown then? If companies are not even investing in the like, you know, buying tokens, I don’t think they’re replacing people and comp just replacing roles with. Like AI, are are they? How do I understand this?

Alex Lu (43:50)

For the tech companies, I think your statement or the statement is true for the tech companies because they’re traditionally there are a lot of coders, there are a lot of programmers, and and actually I see a lot of developers, individual developers in the market because they work for tech companies and now with the with AI. That that would be very challenging. And again, currently for European companies, if I would say

They’re still at very, very early stage compared to to China. the cost is one thing, and we can take at the other angle, causes equals to conservative. So they are a little bit conservative and they care a little bit more about their employees. So actually I I I will not see in European market AI replace a lot of human workers. It’s not happening today. Will will that happen tomorrow? I think so.

Grace Shao (44:46)

Mm-hmm.

Alex Lu (44:49)

but again we need to find another society structure or we need to find other job opportunities for the human beings when AI comes to the companies and replaces some of them. It we’re not like very aggressive like at the tech companies like Meta or other tech companies. it will happen slowly, but of course AI has impact on the on the employment on employment, even for European companies. and

Grace Shao (45:13)

Mm. The economy itself will evolve and and jobs will look different.

Alex Lu (45:21)

Exactly. it that that’s exactly what I I was in Europe ten years ago. It’s exactly the discussion around industry four point zero if people remember. We say that okay tomorrow we’ll have some automated machines in the plant. So it’s kept it’s not it’s happening currently in China. We call it a dark light factory. So it’s very automated. you can run the factory without turning the light on.

so basically at that time in Europe we had a very big debate on where the employees employ employers should go once industry four point zero is in place. And the answer was there were sorry, the answer was there will be some upscaling and new job opp opportunities created with industry four point zero, and we need more skilled people to master these machines. And that’s that’s the same thing for the AI.

Tomorrow we will need people who can orchestra, who can manage the agents, AI agents, instead of doing the same job as a simple agent.

Grace Shao (46:23)

Yeah, I see. So so on that, I wanna ask, you know, given Europe’s strength in industrial, like industrial strength manufacturing, where do we see opportunities for companies to really couple that with the development evolution of AI right now?

Alex Lu (46:41)

You mean the the use cases, right, for the companies?

Grace Shao (46:44)

Use cases, new opportunities, new potential businesses. where could we see p like, you know, new businesses come out or, you know, new business revenues for current industrial companies?

Alex Lu (46:55)

Yeah. for European companies currently the use cases we’re discussing is more around kind of efficiency use cases. So for example, they want AI to help them to do some root cost analysis because if you run a a plant and if the machine is kind of done, the production line is kind of stopped, and then you you you lose basically a lot of money because you missed up.

opportunity of producing X unit units of of your products of your cars. So basically people care a lot a lot about how I can analyze the root causes of of a machine being done. And this traditionally was a very heavy task. We need we need a lot of experts to be involved and because there’s a whole system of different machines in the same plant. And the machine is kind of the product production line is kind of

made in a industrial sequential. So every parameter on different machines might have an impact on the chain. So we need to involve a lot of experts and by using AI actually we can we can understand better. We can do some causality analysis and do some root cause analysis and find the root causes more easily and in the future to do some predictive predictive maintenance and to improve the efficiency of the companies. So this is currently happening.

People are asking for that. And some companies are also asking for these kind of knowledge management platforms. Like we we need knowledge management for new enrollment of employees, for HR policies, for reimbursement policies, for new employees onboarding, etc. etc. So a lot of around that. And if we look into the vision and into the future, I think European companies are start to think about it.

I’m talking a lot a lot about European companies, but that’s the same thing for for the companies in China, it’s just kind of more advanced. So sorry, I I’ll come back. So if we take into the vision of European companies, actually they are also thinking about the future, which is how I can use AI to increase my revenues and to make the pie a little bit bigger. And then it comes to the discussion of agentic economy.

Meaning that can I use my agent to kind of sourcing, to kind of sourcing for my company? Can I use my agent to do some business development, to write emails, to do some code calls, to reach out to potential clients? So these are the things that people will come to think in the next wave, saying that okay, if we have a very good engineering of our agents, guidelines of our agents, what an agent can say, what he cannot say.

what he should say in which context. So once this is done, again it’s very European, they need to use everything kind of under control. Then I think we are ready to to go for the athentic economy so meaning that agent can do business in in the place of the companies.

Grace Shao (50:04)

I see. And if I were to say I’m the founder of AI native company, how would you advise me other w because it would be very different from what you’ve been saying about advising more traditional industries?

Alex Lu (50:10)

Yeah, it it i if you are a AI native founder, I think I’m I’m doing the currently the same position. there are a lot of things to consider. For example, in terms of the technology, the foundation model is evolving very pretty quickly. So how I make sure that my AI agent idea or concept or business model will not be revolutionized or disrupted by this

Foundation models. This is something we need to think about. The second thing is I always tell myself and also people in the say same AI community is we we don’t start to build our products from scratch without discussing with the clients. So why not in in a more safer way, why not discuss with the clients, build products for certain clients, and then kind of

Conceptualize the products and build more standardized products that we can sell, we can say, we can sell to market and we can scale in the future. It means the build of the product comes always from a specific demand of the clients. And once if there’s a demand, then we can do something, we can build things. Why this? Because, in my opinion, all the AI native funders, I think we are pretty aligned is produce.

something or build a product in the future will be much easier in the past. And if we compete with AI in terms of the intelligence, there’s no way a human being can catch up with AI. And we should place our time where the AI cannot compete and where we still need a human being. I I I make very simple analogy to some friends of mine saying emotional intelligence, meaning that how we can establish relationships with the people, how we can build a trust.

So still I think if I’m a founder or if AI native founder, he should go out to meet clients, discuss with clients, build a trust and have some demands from the clients because building the process will be pretty easy and the cost of failing is pretty low. So build fast, fail fast, scale fast and it works even more in the in the future.

Grace Shao (52:36)

And then my question on that is how do we actually understand how to build guardrails and safety around this? Because you talked about how Chinese companies you work with are often a bit more like gung ho, let’s go, we’ll t we’ll fix it if after it’s broken, kind of mentality. Whereas the European companies maybe are seen as a bit of a slow adopter in many ways, you can say more cautious, more humane, and protecting their concurrent employees. But, right, like

End of day, if this is the future evolution of our economy, how do we go forward with this? And then how do we actually build more intentionally?

Alex Lu (53:13)

Yeah. technic technically, actually there are a lot of skills, there are a lot of technical stuff in the area to build the guardrails for the agents, like Anthropic, I they are doing doing a very great job, and also some Chinese foundation model companies and also agentic companies. So all of all of that they call that the harness engineering. So they put every concept into the harness saying that okay, we need to build a harness and to make the guardrails.

So this is the technical perspective. But still, this technical perspective is very from the developers or programmers. And if we bring the case into a real company case, then it really depends on each use cases on each company. I would say for any new human employees which is who is a new hire in the company, at least when I join European companies, there’s always a code of conduct.

You see, it’s it’s simply a a document that we need to learn. We need to we need to we need to be compliant in the future in in our work or professional work within the company. So I would say for the AI agents that the same thing. they are very important in the future, a kind of infrastructure to evaluate the performance of the AI agents, meaning that if the AI agents is delivering the performance as we wished before, so there’s a kind of benchmark evaluation.

And also the evaluation should include also is the AI agent performing correctly as we wished in terms of the code of conduct. And the code of conduct should in my opinion, be written by human human human beings. It’s like an extra bic team, they have they have written a a hundred-page of constitutional constitution for for for for cloud. And then each company should write their code of conduct for.

every agent in every department. And a lot of Chinese founders then they are entrepreneurs, they are also joking at okay, we develop an AI agent today for companies but the next question will come shortly is when should we retire our AI agent it if it doesn’t perform correctly or why when we should replace them. So you see the evaluation or benchmark of of the AI agents would

shortly become a a a pro a p a problem in the market when we adopt massively the agents.

Grace Shao (55:45)

So then each organization will have to institutionalize this, essentially you’re saying, and have their own standards of code of conduct, whatnot. That makes a lot of sense. Yeah. And right just like how companies right now regulate data usage, even company devices, whatnot, right? Like this will all just be part of the compliance that employees will have to learn. I want to ask you one last question, which is what’s one differentiative view you hold?

Alex Lu (55:54)

I think so. In terms of the AI?

Grace Shao (56:16)

In terms of everything, it’s a question I like to just kinda throw throw it at people when they come to the podcast. It’s a it’s a wild card.

Alex Lu (56:24)

Okay. I think one of the points I always mention, it comes back to my background, is today the AI race is between US and China. So we say that European is kind of lagged behind. but do not forget that actually technology is one thing and the usage of technology is another thing. And again, if we come back to our

my my statement saying that implementing AI is not about technology. It’s not it’s about process culture and organization and human being. So I think the placard of the Europe is they’re pretty good at regulations. And if you think about they issued GDPR before the Chinese PIPO, which is protection of personal data. And they have this kind of European AI Act. And then if I think about how anthropic

They penetrated these enterprise solutions versus ChatGPT and generate today more revenues than open AI in terms of AR, because of the simple concept of responsible AI. Then I would say tomorrow, if the AI comes to the enterprise level, enterprise implementation, and if everyone should be responsible in the company with their own agent or with their own developed AI, maybe Europe has a part to play in that.

in the in the AI in the in the world of AI, because their initial statement is kind of we want AI to be regulated, we want AI to be responsible. So this is my point of view.

Grace Shao (58:06)

Thank you so much. You know, today you’ve been really generous just explaining to me and and the audience just how AI is really being implemented into these big companies and the more European perspective. is there anything else you think we’re missing or any misconceptions we might have about the relationship between European companies and Chinese companies or how Europe is perceiving AI? Is there anything you think we’re missing or do you think we covered it all mostly?

Alex Lu (58:36)

Yeah, I I I think we covered most of them, but I just want to mention one thing is even though we say that okay, there’s two different nations in the world, US and China, competing AI, or in we we we take different directions of AI. And still I received a lot of recently questions from European companies, and they are really, really interested by Chinese tech companies. So you would see

they are pretty open and they come frequently nowadays to China and they have the mindset of learning what Chinese companies are doing, what Chinese foundation models are doing, and especially seeking their use cases. so one thing I would say is when I receive them, we show some very advanced Chinese use cases. They would say, you are in a different environment because we have different laws, we have different regulations compared to you Europe.

but they are quite interested about what’s happening in Hong Kong because the regulations in Hong Kong is pretty closer to European markets. So still, I I see we might have a lot of potential collaborations between China and Europe in terms of the AI, in terms of the physical OI. We didn’t mention the robotics, and definitely it’s an area where European can have more playground, not only

About the humanoid robots, they want also to have their places in the hardware value chain for the robot robots. Like a lot of

Grace Shao (1:00:10)

I’m sorry, it’s I know we’ve hit our time, but what what is your view on that? Because you know, European companies traditionally been the leaders in robotics, right? Industrial robotics, like machinery. where do they stand now in the world? You know, are are the Germans and the Japanese still leading the space or or how how are they gonna be kind of presenting themselves or positioning themselves on the supply chain right now?

Alex Lu (1:00:15)

No worries. Yeah. So for the very traditional industry robot robots that let’s say it’s like KUKA, you have a lot of robotic arms. So they are still kind of leading the world, so you have a lot of robotic solutions implemented in the in different car makers’ plans. but for the humanoid robots, actually Europe Europe is lagged behind again because it’s not all only about the value

About the not only about the supply chain of the robots itself, it’s also about again the software and the large language models behind the robots. So the mindset of European companies today is: okay, we understand China again has the most advanced humanoid robotic companies in the world. US has maybe advanced in software in large language models or word models. China is pretty good at the supply chain.

So again, the same question they ask themselves. But the the recent demands I receive from European companies are are two. The first one is as a traditional European companies, we know that they know that the value chain of making a car is quite similar. Let’s say it’s not hundred percent the same thing, but there’s sixty or fifty percent are common of making a car and making humanoid robots.

So their thinking is okay, can we participate in the wave of these kind of robots with the development of China? So like motors, like electric motors, like actuators. Yeah, German, German guys are pretty good at at this apply. So that’s the first thing. The second thing is demand is a lot of European companies saying that okay, we have the real use cases in Europe because we are lack of workforces in our plants.

It could be an aging population, it could be some strike of labor unions. So they in order to keep the plant working, as we said before, about the predictive maintenance, they are very welcome, the Chinese robotics in the European markets. Again, the robots need to be compliant with European regulations, conditions, and they are very welcome. So the most common demand I receive is hey, hey, I I want to do a kind of analysis about

how I can be part of the supply chain in China and how I can leverage Chinese supply chain to be more competitive. The second one is okay, I have a use cases, then we need to think about how I can implement the humanoid robots in the European markets. And then we we can discuss about the business model of the robotic companies like Unitree of AJ Boss, because it’s not only about putting their robots in the factory, it’s about calibrating the robots, it’s about capturing the data, it’s about think about a closed loop of robust training. It’s about

the again, the guardrails how make sure robots will not harm a human being if they cross each other in the plant. So yeah, this is quite common nowadays for physical AI for European companies also, yeah.

Grace Shao (1:03:35)

Mm-hmm.

But in fact, actually you mentioned CUKA and it was bought out by Matee, right, a couple of years ago. So you’re also seeing a lot of Chinese companies like in the embodied AI, physical AI space actually actively buying out traditional brands in in Europe. How is that received actually locally?

Alex Lu (1:03:47)

Yes. actually for the for the embedded robots humanoid robots, there are not so many MA of Chinese players acquiring European companies. So basically I think for the humanoid robots, let’s say the robots like AJ Bot or like Uni3, China is much more advanced. And there was one robotic company in France, but they are kind of in financial difficulty. And another robotic company

They were in they are invested by Renault in France, but still their technology if you look into that is not as advanced as Unitree or AJ Rob A Gi bots, for example.

Grace Shao (1:04:42)

I see. one last question is just do you think it’s fair that we’re overgeneralizing all the European companies into just one EU right now? Or do you think actually a lot of different com countries have different goals, ambitions, or even, you know, future tracks for them laid out?

Alex Lu (1:05:03)

very good question. So I can only when I see European companies, sorry, actually I’m thinking about French and German companies. So actually I cannot represent all the European countries and for different countries like Spa Spain, Italy. I’m I’m not familiar familiar with the country. I didn’t live there. I I didn’t receive enough clients from from these countries. So actually you are right.

when I talk European companies, I’m more thinking about French and German companies. And of course, they are pr pretty different.

Grace Shao (1:05:35)

Okay. Well, thank you so much. Yeah, thank you. I just think it’s such a unique perspective because, you know, it it’s it’s more it’s easy for me to find someone who tells me the pure European perspective. It’s easy for me to find someone in the China US, but it’s harder for someone to for me to find someone f you know, who straddle between Europe and the Chinese market. You know, it’s obviously not as mainstream. So I’m really appreciative of your time and your insights and your sharing. Thank you so much, Alex.

Alex Lu (1:06:04)

Thanks,

Grace. Yeah, thanks a lot again for i inviting me and accepting me for the podcast. And thanks a lot for your audience. And yeah, let’s keep in touch if any chance happens. we can have another talk if needed.

Grace Shao (1:06:17)

Definitely.

For years, on Western social media, China seemed like an exotic (dangerous) location, and then the COVID lockdown happened, which obviously made everything worse. Social media about China felt quiet for a while and then suddenly over the following few years, social media was flooded with videos of cyberpunk cities from Chongqing to even the ancient town of Fenghuang being completely lit, literally lit. In 2025, IShowSpeed did a grand tour of everything from the metropolis that we all know of to the Shaolin Temple. Now, while all this was happening, the AI boom and robotics hype were also taking off.

After the 2026 Spring Gala demos of humanoids and quadrupeds, and the series of robot marathons, the average non-tech professional's intrigue about China finally hit its peak. All of this coincided with China then lifting its visa requirements for much of Europe, Canada, and the dominant anglosphere, making it much easier to obtain tourism visas for others.

Which brings us to now: the West tourist-ing China's hardware in flocks. Reasons may come in many different forms, and I encourage people who are interested in China tech and beyond to visit. What most tech-tourists walk away with are wow robots, wow autonomous cars and EVs, and wow consumer wearables. China feels futuristic, and in many ways it is. And as much as there is a “but at what cost” narrative that people love to slap on, the reality is that, in the narrow track of physical AI, China is ahead of the rest of the world. Here we've written extensively on how, when, where, why.

Subscribe now

Today's post is about the hype, the gaps, the supply chain and the future.

1. Physical data remains the bottleneck

Robotics has progressed enormously over the past decade. And yet, to the layman, it may still not feel that impressive. Nothing feels Matrix-enough yet. We've spent years closing the distance between what the technology can do and what an ordinary person perceives it can do, and those two curves still haven't met. But the demos look incredible to people who follow the space.

Despite its progress, part of that gap is real, but not because the hardware is not there yet; it’s actually in the data that trains the software that empowers the shells.

As I wrote on the bottlenecks for embodied AI a year ago, the binding constraint isn't ambition; it's physical data that hasn't allowed robots to complete physical tasks the way we want them to, or at least not effortlessly. LLMs are trained on text, images, and videos lying around the internet, but robots need sensor and actuation data that simply doesn't yet exist at that scale.

When I visited Nvidia's HQ Voyager earlier this spring, I was fascinated by its simulation program, Isaac Sim put on display in its showroom in the basement. It is the program being used to build the Omniverse platform. On the screen in front of me, thousands of robots were just walking, falling, climbing, rolling, doing backflips, then face-planting on a repeat loop. It felt eerie and wildly futuristic at once. It gives developers a simulation tool that can assemble simulation scenes by assigning complex materials, enabling physics, and configuring robot and sensor models. As we know, the data is scarce, so Isaac actually provides synthetic data for training too. And out of desperation, we’ve also seen some pretty hilarious/dystopian requests come through. One, companies built around collecting data of people wearing wearables to do daily chores (the ScaleAI model?). Two, a New York company, Shift, offering to pay you to clean your room in exchange for collecting that footage. No better way to put it: sh*t is getting weird. (Could you imagine the other kinds of physical data companies will want to collect……)

The point under the weirdness is that physical data is the unsexy gating factor for the whole field. A humanoid built for full-body tasks carries 50-plus precision actuators; a human body has roughly 360 joints. Teaching a machine the angle, force, and height to lift a single pen is genuinely hard. So the layman's instinct that “this isn't quite there” is, in a sense, correct. What I predict we'll see is even more companies finding innovative ways to collect physical data.

Then I want to move to the next point: in one of my recent writings about Qwen, I discussed how the app aims to showcase a new way of agentic commerce to users, and I emphasized that you often have to educate consumer behavior before you can change it. Consumers and the market don't reward the product that's technically ahead. Rather, they often reward the one who arrives after the behavior has already been primed, and I'm seeing that play out in wearables as well.

2. Hardware cannot keep up with software

First is the battery lifetime.

You know those god-awful looking headsets - Apple Vision Pros *sorry not sorry*. I have a personal hate for them because my husband now wears them every evening to “manage his agents.” Anyway, those headsets don’t really last more than 2-3 hours; I mean, thank god for that, for me. I dont really want to see him pinching the air and flailing his arms around for longer than that, but the technical bottleneck many cite is actually the battery life, especially for those I've heard now even take them on flights for a “personal theatre experience.”

Subscribe now

So, some innovative solutions I've seen at exhibitions are the battery “capsules.” They look like a standard supplement pill that attaches to the back of these devices, often the AI glasses, and because they're redesigned to be so lightweight, they do not affect the user experience too much. However, even with that, the additional battery time it buys is maybe another ~2 hours. And I've seen these for the iFlyTech lens as well as the Rokid ones. For how much battery these devices eat to run the software that powers them to be AI-ified, the current solutions just don’t match.

Second is the device form.

And thus brings me to the next point for the builders in wearables: was the challenge ever to “replace the phone?” Who decided that carrying a brick in your pocket or purse was intuitive in the first place? The phone was a natural extension to build on, as we had already normalized the idea of a portable communication device, even back to the walkie-talkie days. However, for the next generation of everyday device, an AI native device, is it to replace the phone, or be an extension of the phone? That is a question I've found many founders in this space pondering and maybe disagreeing on. So maybe, the more likely future, I think, isn't one in which one device dethrones another. It's the phone staying the core screen for maximum-effort tasks while the experience fragments across form factors: visuals through a lens, audio through headphones, navigation through a pin. The phone becomes the operating system or controller, if you must, and the body may, in a way, become the interface.

Third is the speed of hardware iteration

But no matter what you're building- say a pin, a headset, an earpiece, a pair of glasses, a watch, or the wildest ones- an AI-enabled belt, everyone cites the biggest issue to mitigate right now is the hardware-software integration and battery. Software is often iterating so fast that hardware simply cannot keep up, and that mismatch, not a lack of ambition, is what's holding the category back.

And if you want to watch that mismatch get worked out in real time, with all its fragmented interfaces, battery hacks, and agents bolted onto everything, there's really only one place to go see it at scale right now, which is, increasingly, where everyone is going.

3. The rise of “tech tourism” in China

All of this leads to a phenomenon that we’re seeing across headlines: the surge in so-called tech tourism in China. Everyone's rushing to see the robots and trying to be part of it, or, at the other extreme, what calls the “galapagosization” of Chinese tech.

Nathan Lambert has talked about this too. After the SAIL team toured Chinese labs and tech firms, many people reached out. I get the inbound myself: requests for intros from people visiting China for the first time who want to meet the companies. Just last week, Rest of World ran a piece on the droves of Silicon Valley founders and investors reverse-visiting China. Though this has been happening quietly for a few years, it’s really gaining momentum now.

Now, anecdotally, just this past weekend I met up with a friend, an ex-Uber employee who moved from New York to Singapore, then built a successful startup across Korea and Southeast Asia. He's since sold the business and now spends 30% of his year in China, simply because, in his words, “things are happening here and people in the West don't realize, and it's just fun and important to see it first-hand.”

Not everyone can swing a 30%-China, rest-wherever lifestyle. But the underlying idea, that to be in tech is to be near China and visiting regularly, is not new. I've written before, in my EV-to-AI piece, about how nearly every robotics company has some China tie, whether through sourcing, a relationship with a manufacturer, or a collaboration. It's structural, not incidental. China controls roughly 63% of the key companies in the global humanoid-component supply chain, which is why a Western robot still tends to have a Chinese part somewhere inside it. What's newer is the wave of interest in software. Nathan noted on our podcast that more people are asking to visit the labs themselves, and more are trying to understand the industrial policy, something the FT has covered as well.

The timing isn't all a coincidence. As mentioned, it lines up with China lifting visa requirements and extending unilateral visa-free access to 43 countries, mostly European, as well as visa-free or easy 10-year travel arrangements for Canada, Australia, the United States, and others.

And to be clear, I'm not doing the “China is oh-so-good, look at the cyberpunk rock cities” thing. The neon light shows are whatever. I never understood the appeal, just as I never understood my mother exclaiming how pretty a building was because it was covered in lights. For me, the beauty is in the groomed willow trees draping over the alleys in Shanghai, the way the old lane houses blend into the new architecture.

Anyway, the trend itself is real and accelerating, and quite telling of a bit of an awakening, the West catching up to a shift that's been underway for years. Even just a few months ago, Bill Gurley said that his observations were that founders and executives in China followed US developments closely, but that reverse interest was low. And only in a few months’’time, it seems that people are packaging these trips now, from Tech Buzz China to consumer-branding consultants like Ashley Dudarenok and more. Prices vary, access varies. I wonder what happened?

And if I may, I may ride this wave myself and hit some factories in July, so if you have suggestions of companies worth visiting, let me know.

4. Macau, BEYOND, and the humanoid meh

Speaking of roaming around Greater China and seeing robots, I went to Macau this past week to see the physical AI displays at BEYOND, a relatively new conference trying to become China's CES. (Beyond 2025 coverage by yours truly)

Beyond the ~700k RMB-priced Galbot convenience-store shop robot (what, wild pricing, and sorry, slow execution), I saw many, many tired robots and robots being controlled by a remote control (not so subtly).

I dont think I’ll go into the humanoids, but in a world pushing back on automation, I think we've been blinded by the glitzy humanoids when the real labor relief is going to come from industrial robots, which, again, I've been writing about for over a year, both in the rise of China's robotics industry and in the real-life use cases already deployed. This isn't speculative. China already has roughly 1.75 million industrial robots on factory floors, more than any other country, and about 51% of global annual installations as of 2023. It installs more factory robots each year than the next four countries combined. The cost of a six-axis arm has fallen more than 60% in a decade, and a finished machine rolls off the line at 30–50% less than a comparable Western build. That's not a demo but actually an installed base.

And the use cases that matter the most are the unglamorous ones. Watch a few textile factory videos, and you'll be astonished by how much is already automated. Cobots weld and inspect on assembly lines, and agritech drones spray and monitor crops. DJI alone had over 210,000 agricultural drones in operation globally as of 2022, and lychee in southern China has been getting picked by robots since 2022. Tractors replaced horses; this is the same story, one rung up.

The labor math is the quiet driver because China's working-age population is shrinking, and the younger generation increasingly doesn't want the monotonous, often unpleasant work on the factory floor. That disinterest isn't unique to Gen Z in China; it's everywhere. But demand for goods isn't dipping, so industrial robots at scale simply pencil out.

I still think humanoids are weirdly egotistical for humans, and we can't yet justify their utility. They're expensive, a little creepy, and nowhere near reliable enough for the use cases people fantasize about, like childcare or senior care. And I'm not alone in thinking so. Yet we keep seeing headlines about BrainCo, Unitree, BYD-based PaXini, and the like all prepping for IPOs that keep the hype churning, from the brains to the arm makers to the full-on humanoid shapes that are winning the attention war even as the industrial arm wins the actual one.

5. What I was most inspired by

So the humanoids took the spotlight, even though most of these shells looked somewhat the same. For me, the most interesting thing I took from the exhibition halls wasn't the robots. It was a conversation about the future of wearables, kinda looping back to our point 2, which is where that fragmented-interface idea from earlier stops being a thought experiment.

Cars are becoming a wearables platform. Li Auto launched a pair of glasses (Livis) that interact with its vehicles. And there's now a pair of Rokid where one of the guests essentially hacked it and created his own glasses that can control his Tesla, because there's an open agent ecosystem behind it. Rokid leaned into infrastructure and agent support, opening their platform to third-party agents and letting users find ways to make them useful. So maybe in the hardware-software space, the open-platform play may matter more than the hardware itself?

Share

Software is outrunning hardware. We lightly touched on this already. The pace of evolution is so fast that software changes faster than hardware can absorb. That's the integration problem, stated plainly, and the fix is mostly architectural. Right now most cloud providers take a common approach: they use APIs and SDKs, basically standardized plugs or connectors, to link up with a hardware maker just once. After that connection is set, any model update or new cloud feature can be pushed to the device instantly through those connectors, so the hardware can update and improve on the fly without the clunky, laggy experience. I've felt this gap as a consumer myself. I've bought several smart devices that came with built-in AI models, and you could even pick between different ones, but the response time was just too slow, and our patience these days isn't that great. Btw Wi-Fi is totally another issue; if you do not have a consistently good Wi-Fi connection or 5G data, the user experience diminishes noticeably.

Battery is still the wall. As mentioned, this was noted by, like, everyone. The Rokid glasses ship with a small battery, and the experience isn't smooth. A clip-on “capsule” adds a couple of hours; it's very light, a minor physical obstacle. Small touches change everything. Being able to change the prompt for image-to-text or audio explanation on the fly is what makes the experience actually usable. The architecture matters too. 云端分工, splitting heavy compute between the cloud and the device, is part of what keeps the frames light and the battery alive in the first place. But you also need an offline mode for when the connection drops, and latency over WiFi is its own tax.

The “tech for good” use case that stood out. A good product should feel intuitive and solve real problems, not what one of the speakers called "made-up problems by tech people.” And one set of glasses really made me feel that. I saw an older model of assistive glasses designed for the blind. The wearer uses them to detect unsafe situations; the glasses warn through audio, and the front-facing lens can capture images to send to emergency contacts. This may be niche and not yet fully developed, but it is something I see real use for and that can enhance the quality of life for some.

The adoption gap. Dai Yusen (戴雨森) of ZhenFund (真格) recently joined Zhang Xiaojun's podcast, and at the end of the conversation, he touches on hardware. He puts it bluntly: the hurdle for wearables is very high, because people don't actually need another gadget. A new device only earns a place on your body if it adds value beyond the phone, not if it just duplicates what's already in your pocket. The other main friction is installation and maintenance; most of these tools- like those Rokid glasses I mentioned are actually not that useful unless you're a developer yourself and can make something useful out of it based on your needs, as that guy did. The barrier to deployment is that, right now, the market is limited to what they call “geeks” or the high-incomers.

6. Model is the engine, harness is the use case

That same conversation with Dai, managing partner at ZhenFund and an early backer of Manus, Kimi, and a bunch of agent startups, gave me another frame I keep coming back to. His argument is that the core capability still lives in the LLM, but the value is increasingly created in the harness, the agent scaffolding wrapped around the model. He thinks Claude Code is the best example of a harness done right, and he rates Kimi as the strongest open-source model coming out of China.

Why does that matter for everything above? A good harness frees up your attention so you can do more. You stop doing the task and start managing the thing that does it. It's the “one-person company” (OPC blowing up in China btw) idea, where having agents basically makes you the boss. And there's a flywheel underneath it. The better the harness, the better the data it captures. The better the data, the better the model. The better the model, the better the harness. Dai'spoint, which I agree with, is that the real innovation right now is horizontal, in that harness layer, not in another point increment on a benchmark. Incumbents are rarely the first to ship a genuinely new business model or interaction. Most of the harness-layer creativity so far has come from the likes of OpenAI and Anthropic, not from whoever has the biggest model.

Here's how it ties back to what I've been circling this whole piece. It's the same claim as my hardware thesis, just one layer up the stack. In software, the body around the model is a coding terminal; in the physical world, it's a robot or a pair of Rokid glasses. Either way, the value pools in the agent layer, not in the shell. The data flywheel is also why the physical data bottleneck at the top of this piece matters so much. In the physical world, the harness is the data-collection apparatus, which is exactly why those dystopian “pay you to clean your room on camera” schemes exist.

It also names a worry I share with Dai. Because of the physical-AI hype, more people are piling into world models, splitting them into pre-training and post-training, and trying to build them the way you'd build an LLM. But a model that has to predict physics isn't just a bigger language model, and pouring it into the same recipe at greater scale might be the wrong harness for the wrong problem. None of this resolves on a hype-cycle timeline, which brings me to the last thing.

7. Not all hype, but there is some hype

In that same interview, Yusen talks about how Steve Jobs's education of the market took years, not quarters. A lot of unglamorous infrastructure had to be laid first. Apple invested in ARM to secure access to the chips it would eventually need. The Newton OS failed due to its long lag and poor user experience, but the team that learned from it went on to lay the foundations for what became macOS. For a stretch, the macOS and Apple mobile teams were effectively competing over which OS would define mobile.

So, while putting all these thoughts together, my biggest insight, if you must, is that the next generation of physical AI will not be hardware plus AI, but rather AI plus hardware. Hardware being only a vessel to carry out the intentions.

This is the same insight Li Xiang of Li Auto has been pushing in his own industry. As I noted in the EV-to-AI piece a few months back, he argues “electrification” was only the first half of the game, and the true second half is “intelligence,” not traditional software bolted on but a complete evolution built around AI. The winning products of this cycle will be conceived AI-first, with the hardware in service of the intelligence rather than the other way around.

The wearables winner won't be whoever ships the best frames or whatever gadget. It'll be whoever owns the agent layer underneath them, the way, for a matter of fact, Xiaomi and Dreame are already capturing. The hardware will commoditize; the platform the agents run on won't. And the breakout moment for this category is further out than the hype implies, measured in years of infrastructure, not the next demo cycle.

More soon. A few interviews in this space are coming out too in Differentiated Understanding.

In other news: MiniMax launched its latest model, the M3, but, as always, you know me, I like to wait and see what the feedback is first before assessing the business implications. And as we predicted months ago, Tencent was destined to have to integrate an agent into WeChat. The FT just reported on this, though the compliance bit will be tricky.

Finally, see you all at SuperAI in Singapore next week. Cheerios~

Find out more about SuperAI and the speakers’ information here.

ANDANDAND DON’T MISS THIS: 20% discount through my link!

The conversation traces China’s technology evolution from the early internet era to the present. TP argues that China’s internet ecosystem was shaped by a combination of censorship, protectionism, local engineering talent, and intense competition. That created powerful domestic champions such as Tencent, Alibaba, Huawei, Baidu, and ByteDance, which later became the foundation for super apps, payments, e-commerce, cloud infrastructure, and AI.

The discussion then moves into China’s shift from software and internet platforms into hard tech: EVs, batteries, robotics, drones, semiconductor supply chains, and AI-enabled industrial systems. TP emphasizes that China’s technology companies are unusually willing to enter each other’s markets. Xiaomi moved from phones to chips and EVs; Huawei moved from telecom to semiconductors, AI chips, and autos; BYD moved from batteries to cars, solar, transit, chips, and potentially robotics.

A major theme of the episode is that China’s AI story is not only about large language models. It is also about the physical stack around AI: batteries, sensors, motors, chips, power systems, critical minerals, factories, and real-world deployment. TP argues that this manufacturing and supply-chain density may become a major advantage in embodied AI and robotics, especially as real-world robot data becomes more valuable.

Follow TP Huang here on X or Substack here

Share

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, and product managers. For more information on the podcast series, see here.

Chapters

00:00 The Evolution of China’s Tech Landscape

05:58 China’s Internet and Tech Sovereignty

09:01 Investment Trends in China’s Tech Sector

11:04 The Role of Government in AI Development

20:00 The Intersection of EVs and Robotics

26:07 China’s Competitive Edge in EVs and Robotics

36:18 Global Strategies of Chinese EV Companies

42:31 Advancements in AI and Robotics in China

48:31 China’s Digital Infrastructure and AI Adoption

57:38 Underappreciated Developments in China’s Tech Landscape

01:00:00 Non-Consensus Views on China’s Economic Health

AI Generated Transcript (for reference only)

Grace Shao (00:00)

Hello everyone, welcome back to another episode of Differentiated Understanding. I am your host, Grace Shao. As many of you know, I also write the newsletter AI Proem, which is AI PROEM on Substack, so do give that a follow.

Today we’re doing something special. We’re doing an audio-only version. I’m joined by TP Huang, an independent China tech analyst who writes about the intersection of fintech, EVs, batteries, AI, and broader China industrial policy. He has built a large following on X and Substack by combining data, supply-chain detail, and geopolitics to explain where China tech is actually heading.

In this conversation, I want to use TP’s lens to understand the bigger China tech landscape: how China moved from internet platforms and payments into EVs, batteries, robotics, and now AI-enabled industrial systems. And since he quite literally said, “I can talk about anything China tech,” when I reached out, this conversation may follow the themes that I prepared, or really just go anywhere it naturally takes us. Very excited to have him on. Welcome, TP.

Grace Shao (00:02)

Hi, TP. Thank you so much for joining us today. I just did your intro before talking to you. And I told everyone that when I emailed you and reached out, I said, here are some topics I want to talk about. Is that okay? And you quite literally said, “We can talk about anything China tech.” So the conversation today could cover quite a lot of bases. I’m so excited to hear from you and have you kind of dissect a lot of your knowledge for us. And, you know, I’ve been a big fan of following your Twitter, your X, for a long time. Anyhow, thank you so much for joining us today.

TP (00:33)

I’m just really glad to be here, Grace.

Grace Shao (00:37)

Yeah. So you’re a mysterious man. Give us some color on your background and why you are so knowledgeable about China’s tech ecosystem, because you’ve really been covering everything from robotics to LLMs to the internet era. You cover them all, including hardware and chips and everything.

TP (00:56)

Yeah, so it’s kind of interesting that my actual background is not very technical in that area because I’ve been working mostly in the finance sector, or fintech sector slash crypto, for most of my working life. And I did spend a year recently working in an AI firm, so that was something different. But now I’m back to doing more crypto kind of stuff. So my background, I guess now, is a lot more AI-related.

But a lot of the interest I had back in the day was in the renewable space and climate change and things like that. So that really got me started following solar panels, wind turbines, and then EVs. I first read about BYD back in 2008, like a lot of other people. And then as EVs were really taking off in China, that’s when I thought, okay, I really need to understand the full tech stack behind it. So that kind of got me into the entire battery supply chain, a lot of the upstream stuff, and then chips.

The chips part became such a big deal because of AI. So then we had the October surprise back in 2022. That’s when I decided, okay, I’m really going to try to understand how the semiconductor manufacturing part of it works also. And thankfully, I was able to be connected to a lot of people. That allowed me to really understand a lot more.

So I don’t profess to be an industry insider or anything like that. I’m just talking to other people who are working in the industry for some knowledge and writing about it. And then with AI, I actually worked on my own, no, not on my own. I worked with an AI startup, and one of the projects we did was actually for an AI toy. So I had experience running what I would consider to be AI robotics efforts. So I have a lot of real-time experience with embodied AI and also just using large language models. That’s kind of how I got into all this stuff in the first place.

Grace Shao (03:26)

It’s really cool because you have experience across the whole array. One personal question is: what drives you to really continue writing? Because you do write prolifically on Twitter. You have these hot takes, you put things together, and I think you’re quite widely followed by anyone who covers China tech. So what makes you want to share things publicly?

TP (03:49)

Yeah, I guess it’s more like a personality kind of thing, where I really just enjoy writing. And I think there’s something missing in the information space about what is going on in China.

Last summer I was in China for a month, and I plan to be in China again for a month this summer, and I just saw a lot of really cool stuff. I think it’s good for the world as a whole to understand what’s going on in China, for Americans and for all Westerners to understand what’s going on in China, so that we are better informed in understanding how people can work with China and what kind of things people who want to compete against China need to know. But as a whole, I think it’s better to get proper information out there.

And because China is a different language, and most people in China post in their own internet ecosystem on Weibo or WeChat, people don’t really read this stuff. So they get their sources from very bad sources on the English internet. A lot of them are just missing the nuance of what’s actually going on inside China. So because there is this vacuum, I just felt I’m obligated to actually do something about it, to help everyone understand better.

Grace Shao (05:31)

That’s awesome. It’s part of why I write AI Proem too. Well, okay, let’s get into the real stuff today. You’ve been following China’s tech for a while, like you said. Help us understand, just with the sentiment shift, how you view the early internet era to today’s success in hard tech and AI. What really has propelled China’s success in the tech sector in the last 10 to 20 years?

TP (05:58)

Yeah, so I think if we look back on things, China made a pretty big bet on developing its tech sovereignty back in the early 2000s and 2010s. It put a lot of policy in there under censorship reasons. It said, we’re blocking, we don’t want Google or whoever wants to enter China to actually censor the search results so that it fits our local law. And then what actually ended up happening was it became more of a protectionism kind of thing. So China was protecting the local tech champions at the same time that it was pouring a lot of money into these firms.

So it allowed firms like Tencent and obviously Huawei and Alibaba to grow up. Later on, China also developed ByteDance. And if you look at how things are around the world, most countries, most leading Western countries that could have possibly developed their own tech ecosystem, like European countries or Japan, didn’t do it. The only other country that has a pretty robust local tech ecosystem or tech champion is Korea with Naver.

And if you go to Korea, you notice that if you’re using Google Maps, it’s almost unusable. You kind of have to use Naver. So I think there’s a clear correlation between blocking US tech and some level of protectionism to having a local tech ecosystem being developed. And obviously it requires good local engineers also, so that they can take advantage of that. But China had all the ingredients for it.

So even though it started maybe a decade after the US in developing this ecosystem, it was able to develop it because it didn’t have to face this immense competition from the US right away.

And I think also there’s a lot of, you know, we talk about involution in China. I think there were stories of how when Uber tried to enter the Chinese market, because they had to face all these local Chinese companies that were working under 996-type hours, they were eventually pushed out of the market. So I think those are really the interesting parts of how the China tech scene developed in the 2010s.

Grace Shao (09:01)

Does that kind of feed into what we’re seeing now? Because right now it’s a completely different set of technology, yet in many ways it is building off the digital infrastructure that we just talked about, that got built out in the last 10 years or so.

TP (09:17)

Yeah. I think as a whole, if you go to China, even the internet ecosystem works entirely differently from America. In America, for the longest time, we had a search-oriented internet. You use Google, you use a lot of Google products, or you use social media. Whereas in China, because Baidu was never that great, people kind of advanced right away toward these mega apps like WeChat and Alipay.

And as part of the movement on these fronts, you have these giant ecosystems developing where they not only have their own super apps, they also have their own e-commerce networks, their own payment systems, and they all got enough resources to eventually build their own cloud infrastructure and now develop into the AI world. So some of the biggest players in China when it comes to AI are the usual tech giants like Alibaba and ByteDance.

Grace Shao (10:35)

Yeah. So okay, let’s move on from that, from that big holistic overview of China’s internet space and tech sector. So much of the investor focus right now is still through the old internet platforms, like we mentioned, because of the natural progression of how they also become the major players in AI.

But what kind of breakthroughs and capital moved from apps and payments into EVs, batteries, robotics, AI, and hardware? Are we seeing that these hyperscalers or big tech companies are also the major players in these other technologies that we’re talking about? Are they the main investors and backers, or is that a completely different ecosystem?

TP (11:17)

Yeah, China is kind of interesting to me in that a lot of the players are so uber-competitive that they are willing to get into other people’s spaces. So we saw Xiaomi move from the phone into developing their own pretty advanced AI team. They have their own chip design, and most notably they have their own EV division, which is doing really well.

We saw Huawei start off in telecom and then move into the entire semiconductor ecosystem, and also their AI chips, and also into the auto division. We saw BYD start off as this battery company, and then it got into all these areas. It got into cars, it got into solar panels, it got into public transit, it got into the chipmaking side of things, and now it’s also looking to get into robotics with humanoid robots.

Whereas you don’t really see that as much in America, where it’s mostly a typical thing I used to listen to on Wall Street, this entire idea of capacity discipline. Which is basically: how do we reduce competition so that we can get a higher margin? Whereas the Chinese marketplace seems to be one where everyone’s trying to squeeze in at the same time and just fight it out until whoever has the best cost controls ends up winning.

From that point of view, I think this is why for some time people saw that the Chinese stock market hadn’t been growing as much as the US stock market, because there’s just so much competition inside China. So a lot of the funding for these efforts inside China actually had to be backed by the government, these big funds and things like that. And also these things, they are willing to put money into areas of lower initial returns.

A lot of the car factories, maybe they’re not the best investment if you’re looking for a 100% return. Maybe it’s not the best for that. But because it provides local jobs and things like that, the government is willing to put some money into it. And we saw that right now with semiconductors also, and also the data center build-outs. So that is how, over time, the entire Chinese manufacturing ecosystem kind of got built out.

America is trying to do a little bit of that right now with AI data centers and trying to do that with the tariff wars. But fundamentally, the market in the US is about squeezing out competition and lowering capacity in order to charge more. Whereas the Chinese system is about how to scale up production and lower the cost of production in order to have higher margins. So it kind of works differently.

Grace Shao (14:43)

Yeah. So on top of government help and actually putting money into sectors that often have lower initial returns, sectors that are not so sexy in the beginning, let’s talk about DeepSeek.

I think it’s been interesting because we know DeepSeek and many of the other Chinese labs weren’t getting a lot of capital until maybe 2022 or 2023. However, now they’re obviously being pushed front and center as the main economic drivers. Not only are they being looked at as very sexy investments from the private side, but the government funds are also looking to put cash behind this.

How do you view the relationship between government policy, government mandate, and the AI labs in China? That’s part one of the question. Part two is, if DeepSeek and a lot of these Chinese labs permanently price their models at, say, one-thirtieth of the American labs’ prices, what’s the thinking on that? And what’s the sustainable business model for them looking forward?

TP (15:44)

Yeah, so I think it took a while for China to really catch on to this entire large language model thing, because a lot of the Chinese AI, when I looked at it back in the early 2020s, was aimed at embodied AI. So in terms of smart manufacturing, how to improve the grids, drones, robotics, and also EVs, things like that.

Whereas a lot of the US funding for AI was, I guess, kind of abstract. You want to develop the best models, and then we will find the use cases for them. But once it took off, I think there was kind of a light-bulb switch inside the Chinese sector that we can’t just let this go, we have to catch up. The way Chinese people think about things is like, we have to get in on these opportunities.

So in the beginning, with Chinese large language model development, I think it was mostly the big tech companies like Baidu that were kind of leading the efforts. But over time, more recently, I think you find that it’s the startups that have done some unique research that have done the best, like DeepSeek, obviously Kimi, and Z.AI.

And obviously some of the big tech companies are still quite successful, like ByteDance. They have a very good AI product. And Alibaba, with the Qwen stuff, is also very well developed. But you do see that the Chinese government, ever since the DeepSeek moment, has been investing more in funding to make sure that the domestic AI startups are able to get the funding they need to compete.

In the most recent example, DeepSeek, they actually got paired up with Huawei, or maybe they came together somehow. But you can see just in the V4 release recently that there was a lot of integration work between Huawei and DeepSeek. The DeepSeek models are deeply integrated, so that you can use the Ascend chips from Huawei to better train and run the models. And this is part of China’s overall strategy of being self-sufficient in both the hardware and software side of things for AI.

So even though it’s probably easier to just buy NVIDIA chips, the risk of getting cut off by the US government is pretty high. So it’s in China’s long-term interest to have its own ecosystem across the board.

No other country has that. China has not only the chips and the software, but also the entire AI data center build-out ecosystem. There has been a lot of investment or money put into AI build-out-related stocks recently, like optical modules, optical transceiver suppliers, fiber cable suppliers, PCBs, power chips, and things like that.

So there is a lot of investment across China, not just in the software part of it, but also in the hardware integration part of it. And at the end of it, it’s all supported by the Chinese government in some way because they want to make sure that they have a domestic supply chain, so they can’t just get cut off at any point.

Grace Shao (20:42)

So you’re basically in the camp of what Jensen was saying: export controls are not working. In effect, they are cutting American suppliers or vendors out of China, and in that case, actually pushing China to become more and more self-sufficient.

TP (20:57)

Yeah. I mean, for a long time there, Jensen and the good people behind SMCI were trying to get as many NVIDIA chips to China through backdoors, or through Asian and Southeast Asian data centers, as they could, right? So that the Chinese AI suppliers remain hooked onto the NVIDIA ecosystem. But you can see that by sometime late last year, the Chinese government was actively blocking these things from happening because they really wanted the domestic AI players to use the local ecosystem.

Grace Shao (21:39)

But is it actually being replaced right now? Or do you think in the short term, medium term, long term kind of thing? The long-term strategy is self-sufficiency. Short term, it doesn’t seem like it’s realistic yet, right?

TP (21:51)

Yeah, so this is the interesting part. For much of 2023 and 2024, what the Chinese players were doing was that a lot of them were importing the permitted versions, like H800 and H20s from NVIDIA, through official channels.

And then there was a lot of smuggling of chips into China at the same time, and the Chinese government was allowing this. So whenever they were building AI data centers, they would have the data centers that use domestic chips and ones that don’t use domestic chips.

So what would happen is, let’s say Alibaba was looking to access NVIDIA compute and it doesn’t want to get sanctioned by the US government. So what it would do is, it buys some NVIDIA H20s, puts them in a data center, and also leases compute that runs on NVIDIA from one of the state-built or local government-built data centers that smuggled in chips, because it didn’t want to get in trouble by buying them if it’s not allowed to.

Another thing that these firms started doing that’s entirely illegal, again, is actually just setting up companies offshore that would buy these NVIDIA chips and then build data centers in the rest of Asia, places like Japan, Thailand, and Malaysia. And then they would lease the compute for these NVIDIA chips from these data centers.

And that’s still going on right now. The Chinese government is allowing that because domestic firms like ByteDance would just say to the Chinese government, we need this ability to use American chips in order to not be left behind. Because if you talk to the AI developers in China, they don’t enjoy using Ascend libraries for training. They don’t mind using them to run inference, but for training, they still prefer to use NVIDIA chips.

So there is an effort right now to also get the training part of it up to par. And that’s kind of what the DeepSeek work with the Huawei team in recent months has been about. It’s kind of interesting to see how much better the integration has made the Ascend chips run training and inference on the DeepSeek models.

There has also recently been a Qwen model called 3.7 that came out. And they also released their own AI chip called Chengwu MA90.

TP (25:10)

And part of the interesting thing about that is not only did Alibaba have the self-designed chip, because it was designed internally, it used its own internal AI models to write the kernels for the chip. And it had some really good results. I think going forward, a lot more of these domestic chips will actually be able to at least do part of the training also.

Grace Shao (25:39)

That’s really interesting. And it actually echoes some of the stuff I’ve heard on the ground as well. So like I said in the beginning of our conversation, I don’t want today’s conversation to only focus on China’s LLM and model space. I want to double-click on something you mentioned at the very beginning of this answer. You said China actually started with its capital focus and technological focus on EVs and embodied AI.

What’s interesting is that that side of things didn’t really pick up in the US or in the West, per se, until more recently. So did the EVs come first, or did robotics come first? Or did they kind of converge and come at the same time, and there’s synergy there?

TP (26:23)

Yeah, so when it comes to the EV and robotics story, I tend to think of it as something that started because China was doing all the manufacturing of consumer electronics. And that’s how it was able to then develop these OEMs in the smartphone space, like Xiaomi, Huawei, Vivo, Oppo, and Honor. Basically, they developed this entire workforce inside China that was very good at dealing with supply chains and also integrating things together and doing manufacturing.

I personally had an experience with this about a year ago, where we were trying to make this AI toy, and I got on a call with a Chinese factory. I won’t say which one. But basically, about five minutes into it, I realized America was in trouble because we had all these American engineers who are decently smart people. And the sales lady at the Chinese factory just knew way more about how hardware works and should work than any of us did.

It was a very humbling experience just to see how there’s a lot of process knowledge involved in this. There’s a lot of experience involved in this stuff, right? And my cousin actually works in Shenzhen.

Grace Shao (27:46)

It was like learning from experience instead of PhDs, right?

TP (28:03)

They developed their own automated device that tests blood samples to see what kind of disease you might have, something like that. And what I realized talking to him was that this entire supply chain in China around Shenzhen or around Hangzhou is very deep.

Because of that, a lot of the modern tech that we see with embodied AI comes from this basic understanding of supply chain, software-hardware integration, and also electrical platforms. What are the commonalities between drones, robotics, cars, and EVs, right?

First, you need to have this battery underneath. You need to have electrical platforms. You need to have PCBs. You need to have cooling systems involved. You need to have control chips. You need to have power management chips. You need to have main control chips for the actual device. You need to have AI chips.

All this stuff, in the beginning, Chinese suppliers were sourcing from abroad. Over time, due to export controls, they started doing this domestic substitution. They’re still the biggest importer of chips globally, but a lot of that stuff is coming in-house now.

So if you do a teardown of a DJI drone, you’ll probably find memory chips from CXMT and YMTC. You’ll probably find CMOS chips for the camera modules from maybe OmniVision or something like that. And the battery is obviously going to be domestic. And all the stuff that we saw with drones and with EVs, we’re now seeing with humanoid robots and other kinds of robots, because at the end of it, a lot of the basic concept is very similar.

You need to have some kind of a brain for the embodied AI machinery. And then it needs to have some kind of battery source to actually do the functionalities. And then it needs to move using some kind of motors, and then it needs to be able to absorb information from its surroundings with these sensors.

That is why China has such a large ecosystem, because it has a good upstream supplier network and a lot of people working on this stuff. Whereas if you come to America, there’s just not a lot of that talent around.

So if you want to develop an AI robot, you have to do everything in-house and figure it out. Because if you can imagine, if you don’t develop in-house and you contact a supplier in China, you can’t really iterate things quickly because you’re working with someone over there who doesn’t speak English and also doesn’t work the same hours you do. So the turnaround time is just much slower.

Whereas if you have an idea in China for an AI robot that you want to build and sell to the market, you can get it produced in a month. That would be crazy for any kind of AI startup in America to do.

Grace Shao (31:59)

Yeah. In fact, I think there are a lot of robotics companies right now with founders who are literally tweeting about this thing: we must move to Shenzhen. Or I know of companies that actually get their hardware completely end-to-end, basically buying from OEMs from Shenzhen and slapping on a tag elsewhere.

But I want to ask, why did China ultimately come out on top in EVs? Because from what you just mentioned, technically wouldn’t countries like South Korea have an edge? They have car manufacturers, they have chips, they have memory chips, especially when you just talked about brains. It’s not like the brains that we’re talking about right now are AI brains yet.

So what made China actually come out on top with EVs and robots? Was it again this narrative around government push, because the country needs clean air? Was it because of innovation? Was it because of renewables and everything coming together? How do we understand this?

TP (33:01)

Well, I think Korea itself is actually a country with a lot of industrial policy also. So I wouldn’t necessarily say that the Koreans were less aggressive about government support than the Chinese were.

I would say that if you look at just the human capital side of things, we’re looking at a magnitude difference in the number of engineers coming out of South Korea and China. So that’s something not easily made up.

If you have 10,000 battery engineers from China every year, and let’s say you have 1,000 from Korea, the 10,000 are going to crush the 1,000 over time. And you can kind of see that. Back in the late 2010s, the Koreans were ahead of China in battery technology. But because Chinese industries were moving so fast and the supply chain was moving so fast, China has been ahead of Korean battery makers for several years now. And the gap is only expanding as we move toward more advanced solid-state batteries, or lower-cost sodium-ion batteries.

Batteries are such an important part of the modern electrical transition that it’s kind of mind-boggling that China controls so much of the entire ecosystem. People keep talking about TSMC, or Taiwan having some percentage of manufacturing for chips, which by the way is not true. But Taiwan only has a small part of the entire ecosystem. Korea only has a small part of the semiconductor ecosystem, right? America has a huge percentage of the semiconductor ecosystem.

But if you look at things like rare earths, critical minerals, and batteries, China actually probably controls 80% to 90% of these ecosystems. So even the Korean battery makers rely on the Chinese supply chain for a lot of their inputs now. And there’s just no way to get around it because the Chinese process knowledge, cost advantage, and engineering advantage are very hard for a smaller country like Korea to overcome.

Grace Shao (35:47)

Interesting. Yeah. So how should we understand these companies’ international strategies? Because I think you’ve written about it before. Like you said, they are major exporters. How do the battery companies and EV companies position themselves globally? Are they quite aggressive? Are they suppliers along the supply chain? Are they building up consumer brands? How do we understand that?

TP (36:19)

Well, it’s different with different people. I think because the domestic market is so aggressive and so competitive, companies like BYD had to go abroad to get higher margins on their products. That’s kind of forced a strategy where they’ve aggressively expanded. Things especially picked up in the past few months because of the Iran war, where there’s also a lot of demand for these EV products abroad.

And as a result of that, it helps what I call China Inc. As you see more of these high-tech EVs abroad, as you see more of these DJI drones and Chinese AI models abroad, there is a generally higher view of Chinese products now from much of the Global South. And as a result of that, Chinese firms are also having greater success selling their products.

I think one of the interesting things recently is just to see how much the Chinese automakers’ market share in Europe has already surpassed the Koreans and is catching up to the Japanese. Just looking at that, it gives me the impression that the Chinese automakers, and just China Inc. as a whole, have gained a reputation for quality in a very short period of time. And you can only do that if the automakers themselves are making a real effort to build their brands and promote their products in these markets.

And I think they’re getting paid off because my guess is that BYD’s automotive sales have much higher margins on stuff sold outside China than inside China.

Grace Shao (38:43)

I see. So it’s still like a pricing strategy, or winning on pricing, you’re saying.

TP (38:50)

I think in China it’s more of a pricing strategy, but abroad you see them actually marking things pretty high. So maybe there is a pricing part of it, but if you listen to Stella Li, Executive Vice President of BYD and President of BYD Americas, talk about the new models that they launched in Europe, they’re very much trying to frame it as a luxury brand, with the Denza model brands.

She would say that this is technology that does not have any competitor or equal in Europe. We’re just way ahead of the Europeans here. We’re going to build the fastest charging network that you’ve ever seen. You can charge your car in five minutes, for example.

It’s kind of interesting because BYD can sell its cars at a much higher price outside China than inside China. Inside China, it might have to sell its cars at a discount to Tesla cars. Outside China, it might sell them at the same price as a Tesla car. So yeah, I find that interesting.

Grace Shao (40:04)

That’s very interesting. And I’m kind of playing devil’s advocate purposely. Anecdotally, I’ve obviously been in a lot of BYD cars when traveling in China. They are actually really, really sleekly designed. And like you said, in China, for some reason, they’re positioned more as not a luxury car at all.

But even in Hong Kong, I’m seeing more and more Zeekr cars and BYD cars taking the roads, and they’re definitely replacing previous Audi and Volvo owners. It’s very interesting that that’s the trend. Outside of mainland China, the reputation of these Chinese EVs is almost more premium than they are in China.

TP (40:49)

Yeah. And one of the reasons BYD wanted to do well in Japan and Germany was that it thought that once it started selling well in Japan and Germany and got approved by those automotive nations, people inside China, especially suburbanites in Shanghai, would then accept BYD as quality products. It is kind of interesting that a lot of times the Chinese can’t really accept that we have quality products unless it’s also being accepted abroad. It is kind of interesting how that works.

Grace Shao (41:21)

Psychology, I guess.

TP (41:31)

Yeah.

Grace Shao (41:37)

I guess it’s a little bit of a psychological play on this as well. I do like your framing on China Inc. And I think recently we’ve seen that even in the consumer space. It was so interesting that Luckin Coffee bought Blue Bottle coffee, and you’re getting more and more of these kinds of purchases, like SHEIN buying out Everlane, etc.

But I want to bring it back. I want to bring it back to AI.

You said earlier that China’s mastery of hardware manufacturing has given it an edge in scaling humanoid and service robots. But how do we understand where we are with world models and the actual next stage of embodied AI and physical AI right now? Because like what we just discussed, China’s manufacturers are very experienced in building out the robots, drones, and various forms of robot mechanics. But where are we with actually injecting that with AI?

TP (43:02)

Yeah, so I’ve been in touch with the guys behind the China Research Collective, and they are actually inside China, so I’ve had some discussions with them about this. They’re telling me that because China has this hyper-competitive local market for jobs, a lot of young people are having trouble getting the jobs they wanted. So they’re willing to help these AI companies collect data on doing things to help these world models.

It’s kind of interesting because you need a certain amount of data so that the robots can simulate human movement and then do the tasks. But at a certain point, if you have a child, you know that it takes them a long time to be able to walk around and then run, because they need to first feel and touch everything and learn everything over a year or so. During this time, their muscles develop and their muscle memory develops so that at a certain point they no longer need to think about how they walk. They can just walk. They no longer need to think about what they can or cannot eat, because they already put that stuff in their mouths to test it out.

Longer term, I think once you have enough robots in China, they will just be able to improve exponentially in their capabilities because they will be able to fast-track all this, what I call reinforcement learning in the real world. If you try grabbing an object a million times, eventually you’ll figure out the best way to grab it. And once a robot learns how to grab it, that gets shared amongst all the robots of that family.

So I think as you see the Chinese robotics rollout speed up, this is when you see this decisive edge in the world models. We already saw this with drones, right? The Chinese drones are just so much better at moving around and doing stuff because they had so much more data than anyone else.

We’re seeing it now in EVs, where the Chinese self-driving cars are really good because they’ve had a lot of data out there, where people are just using autonomous features to do all the work. And you’re seeing that BYD today is having this entire unveiling where it’s talking about its path toward L3 and L4 autonomous driving.

The more data it has, the better it’s going to get. That data becomes an advantage going forward. In the future, whoever has the most robots out there in the real world, and has all that data, can then train their robots faster. That’s why it’s kind of a big deal right now that BYD says it’s going to have 20,000 robots in its factories this year, because then it has all this data on using robots in a factory setting. That’s going to improve the performance of the world models by leaps and bounds.

Grace Shao (46:48)

Mm-hmm. Because the biggest bottleneck right now is just not having enough 3D data. And collecting that kind of 3D data is extremely challenging without, like you said, real, actual physical deployment. That’s fascinating.

TP (47:10)

Yeah. I also want to point out one other big difference between the Chinese players and the foreign players outside China, which is that China has this entire critical mineral supply chain. That is foundational to the rare earth magnets, for example, needed for the different robots and EVs, and for the motors, and also the materials needed to build the humanoid robots themselves, like magnesium. It produces about 80% of the world’s magnesium, and magnesium alloy is considered to be the main material that you want to use for humanoid robots.

Grace Shao (47:57)

I want to tie it back to what we also talked about earlier. Does the very strong digital infrastructure layer, just from fintech, IoT, and 5G, now contribute to China’s very quick adoption and diffusion of AI in the real economy? And how do you view this kind of positive cycle versus in other economies, where sometimes the digital infrastructure maybe just isn’t there yet and seems to need time to build up as well?

TP (48:31)

Yeah, I actually think this is one area where America might have a leg up on China, because the American big tech companies tend to also be the biggest cloud service providers. The Chinese ones are a little smaller. So right now, you only see the competition between the US and China because they’re the only two countries that have this data center and AI infrastructure advantage over the rest of the world.

The biggest players in China, like ByteDance with their entire AI cloud infrastructure and their entire AI app ecosystem, are also the ones that are able to deploy their apps globally the fastest. In America, ChatGPT/OpenAI has this commanding position not because it has an ecosystem, but just because it was the first to do it. It had a first-mover advantage.

But if you look at the players outside of ChatGPT, it’s Google slash Gemini that probably has the largest market share, because it has this big data center hardware, this AI infrastructure advantage over other players. And also it has this app system that people can use the AI features in.

In China right now, personally, I don’t get to use the AI apps in China all that much, but I do have a Chinese phone, and I use ByteDance’s Doubao app, and it’s really good. So that has allowed ByteDance to have the best video generation model out there, called Seedance 2.0.

Grace Shao (50:29)

Mm-hmm. And they really leverage and lean into their data advantage as well. Obviously, if you own TikTok and Douyin, you have the most amount of video data in the world.

TP (50:43)

And not just that, they also have CapCut.

Grace Shao (50:58)

They do, which is the editing tool. I actually use it to edit our videos here on AI Proem. It’s great. I kind of want to wrap it up soon.

I want to ask you a forward-looking question. If we connect the dots from your fintech days covering the digital economy to what we just touched on, EVs, robotics, hardware, everything, where do you think China’s digital economy goes over the next five to 10 years? What are the biggest bottlenecks? Will that look very different from the rest of the world? Or do you think the evolution of technology will be organic and go in the same direction, no matter your geographical location or your domestic strengths or weaknesses?

TP (51:30)

Yeah, so I will first talk about where I think they can possibly see the most improvement, and that will be the semiconductor part of it. I do think they will have a fully domestic semiconductor supply chain pretty soon. And that, along with government support in terms of putting money into these high-capex, maybe lower-rate-of-return investments, will allow them to more aggressively build out the domestic semiconductor infrastructure.

Once you have that infrastructure, then you can produce all the AI chips, all the phone chips, and all the analog chips that you need for your various embodied AI products and EVs and all these other leading sectors. And once you have that, that means you’re no longer constrained. You’re no longer constrained by compute. You’re no longer constrained by possible Western tech export controls on you.

So then the AI players in China are equal in terms of AI infrastructure. And that allows them to compete a little bit better with their American counterparts. Now, they do have some obvious advantages over their American counterparts. We’ll have to see how this plays out, because China does have this entire grid build-out that is just unrivaled. And as we move to a more electrified global economy, being able to build not only data centers but the entire grid is actually a huge competitive advantage over the rest of the world.

I don’t really want to say who wins the AI race, because I feel like you can only lose the AI race by not participating and investing in it. But if you invest and put a lot of money into it, like both the US and China have, both of these countries will have a huge share of the global economy going forward.

TP (54:15)

I just don’t see how you can put this much effort into AI in America and not get something out of it.

Grace Shao (54:24)

I just feel like it’s not a zero-sum game.

TP (54:28)

It’s only bad if you don’t try to build your own AI industry, right? If you don’t invest, that’s a problem. But if you invest, something good will happen, I think.

Grace Shao (54:42)

What about the smaller countries where they don’t have that capital, and maybe they don’t have that much capital to deploy into this, or even frankly the talent to build their whole AI stack? Where do they fit into all this?

TP (54:53)

Yeah, so I think that’s one of the factors that might help the Chinese ecosystem over time, because a lot of the open-source stuff is coming out of China right now. So if you’re from one of the smaller countries, let’s say Singapore, and you want to develop your AI sector, you are more likely to use an existing open-source model and do reinforcement learning training on top of that, and then develop your AI product on top of that, than use something you don’t have any control over, like Claude, for example.

Grace Shao (55:41)

Interesting that you use Singapore, because I was just there last week and literally OpenAI just announced their satellite office. I think they said they would employ 200 people. Singapore is an interesting story because, if anything, they’re super gung-ho on AI, from top-level diplomats and ministers to companies. So it will be interesting to see how they play out this strategy.

My question for Singapore is: they can attract a lot of talent globally to go over. They can attract a lot of new companies to go over, which is what they did with the internet era too. ByteDance, Tencent, Facebook, everyone’s there. But then what is the value they propose for the locals? Or how do they plan to diffuse AI into the economy? I don’t know how they make themselves that relevant globally beyond being a hub for these companies.

TP (56:35)

That’s a very hard thing to say because I don’t see Singapore, just on its own population, actually developing anything unique. The people who would work in Singapore’s AI industry could work in any other country also. So I think Singapore has always put itself out there by being a country that attracts talent from all over Asia, right? And they attract a lot of capital also from the rest of Asia.

There have been a lot of issues in recent years where they say all this money coming in hasn’t really helped the local-born population in Singapore. So that is something interesting to watch out for.

Grace Shao (57:25)

Yeah. I don’t want to go on a tangent on Singapore too much. So, last two questions. One is: what is one underappreciated hard-tech development you think people are missing?

TP (57:38)

Yeah. Last year, I wrote a thread about a list of what I call sanction-breaking tech that was happening in China. A lot of these are not things you see in the media as much, because they are the zero-to-one steps in the upstream supply chain that need to be achieved in order for an end product to be built three or four years later.

So things like high-speed analog-to-digital converters and digital-to-analog converters, advanced diamond substrate for heat sinks and other purposes, high-end gallium chip designs, and a lot of the lower-level material science-related stuff that people don’t really see.

But once China develops these things, that’s when you see this really fast iteration afterward. Because everything in China is kind of built upon the idea of having the upstream supply chain and the process knowledge. And then it can iterate through the end product a lot faster.

So as fast as China has moved in the past 20 years, I don’t think the West is really prepared for what is to come out of China in the next 10 years. I really don’t.

Grace Shao (59:36)

Interesting. Okay. Well, I think that’s a topic that no one really has an answer to. No one really knows the future, right? But I appreciate your thoughtful answer.

My last question for you is a question I ask everyone who comes on the show. What is one differentiated view you hold that you think is non-consensus?

TP (1:00:00)

Interesting. Well, one thing that I’ve talked a lot about with people recently is that if you listen to mainstream media, when they talk about China, they always talk about the economy not doing well and that China has this housing bubble that’s apparently a real problem, right? And that China has this demographic problem going forward, and that’s why China might have problems going forward.

I’ve actually always held the opposite belief, in that I’m always under the impression that China grew overly rapidly for many years because it built up this real estate bubble, and all that money went to real estate instead of the tech sectors. And at a certain point, it decided that it could no longer blow up this real estate bubble because young people weren’t getting married and having kids because they couldn’t afford homes. So it deliberately deflated the real estate bubble in order to solve this problem.

And then it still claims to have grown at around 5% a year for the past few years. If you can deflate a bubble and grow at 5% a year, that is quite the accomplishment, actually. So I would say the Chinese economy is quite healthy.

You would rather have an economy that can grow strongly in the middle of an asset bubble deflation versus an economy that is growing just a little bit in the middle of a historically large asset bubble, like you have in the equity market in the US.

Grace Shao (1:02:05)

That’s a very interesting take, actually. I’ve never heard someone say that. But yeah, I kind of see where you’re coming from.

TP (1:02:14)

Yeah, that is my take.

Grace Shao (1:02:17)

I love it. TP, look, I’ve taken up an hour of your time. I really appreciate your insights. And you entertained my brain going in all directions as well. We’ve really talked about a lot of different topics today.

Is there anything else you think we didn’t cover that you would like to share with everyone? Or do you think we can always pick this up again another time?

TP (1:02:40)

The only thing I would say to everyone out there is, if you enjoy AI, try one of the cheap Chinese models and see how it works for you. I’ve tried it myself. It’s great for my work purposes. And I highly recommend everyone use Kimi.

Grace Shao (1:02:58)

There’s a plug. No, I’m kidding. They are good, actually. I think I use different models for different things, but ultimately I find that if you’re really using them for more basic writing and everything, the Western ones are better. But if you’re really hosting your own models and running your own agents, then a lot of the Chinese ones are a lot more cost-efficient.

So thanks again, TP. Thank you so much for your time.

TP (1:03:28)

I’m glad to be here. I’m glad to be on your show. And you can all follow me on X at TP Huang. I’m really glad to be on this show.

Grace Shao (1:03:38)

Definitely. And TP is on Substack too.

So let’s go back to our bread and butter: understanding Chinese big tech strategy. With Alibaba and Tencent both having just reported quarterly earnings in the past few weeks, this is a good moment to look at how the two are actually playing AI.

What’s becoming clearer with every quarter is that the hyperscalers are trying to integrate AI into their distribution and do all they can to enhance their offerings. As much as the discussion is around how labs are monetizing, for the big tech, the real money still gets made one layer down: providing infrastructure and service add-ons. That’s true for Alibaba, that’s true for ByteDance, and in some capacity it’s true for Tencent.

Same same but different

It’s super easy to overgeneralize and ask “how’s China monetizing AI?” Well, to start, China’s two largest internet platforms are both investing heavily in AI, but their strategies have diverged from the get-go. Despite both trying to sell to enterprise, their ways of pursuing that couldn’t be more different.

Alibaba says it’s rebuilding commerce around AI. But the Qwen app itself hasn’t really taken off, hasn’t changed consumer behavior at any fundamental level, and frankly the agentic activity loop isn’t mature enough yet. Tencent is trying to turn WeChat into an agentic operating system, but it’s still cautious, wary of data breaches and of any agent going rogue inside the most important app in Chinese life.

For both companies, AI isn’t a standalone product. It’s what they’re leveraging to elevate their existing businesses. Or, to be cynical about it, to remain relevant. On the consumer side, this probably isn’t something people will pay for as a SKU. It gets embedded into existing businesses, workflows, and user habits. The win is engagement and stickiness inside the ecosystem.

Share

Alibaba’s answer is cloud plus commerce. Tencent’s answer is WeChat plus agents (until they get a stronger proprietary model?).

You can argue Alibaba’s playbook (sell cloud first, then the Qwen APIs, then services on top) now looks like the cleanest and most lucrative strategy. Alibaba is building from the infrastructure layer upward: Qwen models, Alibaba Cloud, Model Studio, enterprise agents, and on top of that, AI merchant tools and shopping assistants inside Taobao and Tmall. It wants AI to be both a cloud revenue engine and a new shopping interface. The cloud business will likely continue to see notable growth in the coming months as well. As its flywheel takes off, the stronger Qwen models run, the more cloud sales too.

For more big tech coverage, see here.

Tencent is building from the distribution downward, maybe unwillingly, because its cloud has never outcompeted Alibaba’s, and its Hunyuan model continues to lag behind peers too. What it has is the ubiquitous WeChat: Mini Programs, Video Accounts, ads, search, payments, the social graph, and the ability to put agents on top of it all. Tencent doesn’t need users to open a new AI app every day. It needs AI to make WeChat more useful, ads more effective, creators more productive, merchants more efficient, and transactions more seamless.

Alibaba: AI as the new commerce and cloud OS

Kicking off the most recent March quarter earnings call, Alibaba CEO Eddie Wu said: “Alibaba’s full-stack AI investments have progressed from incubation to commercialization at scale. This quarter, we achieved accelerated breakthroughs across models, cloud infrastructure, and applications.” He flagged that Cloud Intelligence’s external revenue growth accelerated to 40%, with AI-related products accounting for 30% of that revenue.

Alibaba’s strategy has been clearer from the start, and some may say it is the more aggressive and more vertically integrated one. Models (Qwen). Infrastructure (Alibaba Cloud). Enterprise products (Model Studio, agent tools). And the commerce layer on top (Taobao, Tmall, Alibaba.com, merchant tools). It’s trying to wire all of this into a single AI-first stack.

In the March quarter remarks, Eddie Wu said: “We see massive potential for agentic AI; we launched multiple enterprise AI agents for office and coding use cases and fully integrated e-commerce capabilities into the consumer-facing Qwen app, deepening synergies between AI and our consumer ecosystem.”

On the other side of the world, Meta has been trying to do something similar, planning to integrate its agentic shopping tool into Instagram. But people underestimate how hard this actually is. An agentic shopping app is not a chatbot that can talk about products. It has to read messy human intent, query a live product catalog, maintain state across multiple steps, process the actual payment, assume liability when something goes wrong, and trigger a return or refund that flows back into real logistics. Each one of those is its own integration. And every integration you don’t own becomes a third-party dependency that can break, reprice, or simply say no.

That’s why when OpenClaw (the open-source AI agent framework OpenAI later acquired) first launched, it had a huge developer moment but never reached normal users. It was a personal AI assistant that ran locally on your own machine and controlled a browser by clicking through pages the way you would. Brilliant in concept, but a bit more brittle in practice, at least hard to scale. The moment a checkout page changed its layout, the agent broke. The moment you actually needed to pay, you were back to handling your own credentials and cards yourself. No matter how many people tried to download it and scale it up, it didn’t work, because the agent layer itself was fine, but everything underneath it (auth, payment, fulfillment, recovery when something went wrong) was still somebody else’s stack.

So credit where credit is due, as much as Qwen has not taken off or disrupted user behavior much yet. It was a genius move. Qwen has access and ownership across the layers, which is why it could actually ship an app that runs end-to-end. Unlike OpenAI needing a Stripe deal, or Meta still figuring out payment partnerships, Alibaba already has shopping, payment, post-purchase service, and logistics sitting in one integrated stack. No third party needs to make it seamless. There is no third party.

The strategically important thing isn’t that Qwen is improving. It’s that Qwen has Taobao and Tmall built within it, and it has Qwen wired into the apps; it becomes a recursive learning loop.

Traditional e-commerce starts at a search box. You roughly know what you want, type some keywords, scroll, compare reviews, ask the merchant a question, place an order, track delivery, deal with returns if it goes wrong.

Alibaba says Qwen Shopping Assistant can help with idea generation, product discovery, in-sale support, order management, and post-purchase service. So this isn’t a smarter search page. It’s an AI sales associate that runs across the whole journey. Lower friction for consumers, potentially higher conversion for merchants, and for Alibaba, a new control point between the user and the product result page.

That last bit is the strategic prize. If the assistant becomes the interface, Alibaba can then reshape how products are discovered and how merchants compete for visibility. Of course, it means how merchants market themselves will change too.

As the old interface was keyword search plus recommendation feed, the new one could be conversational intent plus agentic execution.

The risk is the flip side of the same point, though. If AI compresses the shopping journey, it disrupts how Alibaba actually monetizes discovery. The old model sold attention across search results, displays, and feeds. If the assistant just gives users fewer, more direct recommendations, the new interface has to lift conversion enough to make up for the lost ad inventory. So it’s powerful but not risk-free, as AI can improve shopping and also disrupt how the platform’s existing business models work. Ask Google about it…

Anyway, but the market reacted quite positively to it all. Alibaba shares surged 7% as the company doubles down on AI. Quietly, the company has achieved greater depth and breadth in AI integration than most assumed.

Ok, so you know how I kinda shat on the idea of agentic commerce? How it only made sense for utility needs, not discretionary spending? And the lack of engagement doesn’t make sense for clothing/food, whatnot. So I think that, in the next stage, to really ensure engagement, innovation might need to be part of the interface. Just like Brian Chesky recently said on TBPN, “I think the future is not apps. The future is agents, but I don’t think they’re going to be text-forward. I think they’re going to be really rich user interfaces”…” with e-commerce, you want a very rich user interface. It would be agentic. You can have a conversation with it, but the point is that it has to be more visual.

And on the other hand..well, yesterday I was having coffee with a friend who works in construction and real estate, and he completely challenged my thesis. For businesses like that, and for wholesalers, purchasing is routine. There’s a budget. The repeat purchases are clear, and the SKUs are clear. Agentic commerce makes a lot of sense for B2B buyers, and it can eliminate much of the corruption and handling fees that currently pile up in those flows. It can also optimize pricing for exact ingredients or raw materials based on data, rather than on the moods and preferences that drive consumer consumption. That actually makes sense when he pointed it out to me. Now the question becomes who can build that plug-in layer for companies to adopt.

Goldman Sachs’s Decoding the Agentic Economy report estimates enterprise AI agents could lift global token consumption 24x by 2030 and 55x by 2040. That includes enterprise adoption of agentic commerce.

Which connects to what I keep saying on public forums and panels: AI adoption isn’t actually happening at large scale yet, and it isn’t here to replace jobs. Jobs are many tasks bundled together, and AI can do some of those tasks well. What we’re in is early-stage work-enhancement technology, closer to email than to factory automation, as Goldman illustrates. That also means agent adoption within workflows is nowhere near where it could be. My bet is the next stage shows up in wholesale and B2B sales: better efficiency, fewer mistakes, less corruption, and the tedious middle-manager work compressed down to one or two people checking in between purchase and delivery.

The merchant side may matter more than the consumer side

Merchants already pay Alibaba for traffic, for tools, for services. If AI helps them run their business better, the reason to pay is even more obvious.

Wukong is Alibaba’s AI-native enterprise agent for merchants. Accio, which helps merchants with industry research and ad work, has already taken off. Adoption details are still thin, but the use cases write themselves: product listings, image and video generation, customer service, ad campaign optimization, inventory analysis, promotion design, review management, and figuring out what’s actually likely to sell.

This is where AI produces concrete ROI. A small merchant doesn’t have a full marketing team, design team, copywriting team, and data analyst. If AI helps that merchant create better product pages, better ad creatives, and better customer responses, Alibaba turns AI into a productivity layer for millions of sellers.

Same logic on Alibaba.com’s Accio and Accio Work. Cross-border sourcing is information-heavy, fragmented, and painful: find suppliers, compare prices, understand certifications, negotiate, translate, manage documents, coordinate logistics, handle follow-up orders. Which tbh is really a textbook agentic use case. If Accio Work moves from “AI search for suppliers” to “AI operating system for global trade,” that’s a cleaner monetization path than a consumer chatbot.

Tencent: AI as the invisible upgrade

So we’re all anticipating Tencent’s Penguin reveal next week, but beyond that hype, it’s no secret that Tencent’s management has felt pressure to justify why they aren’t doing that well on AI. A year ago it looked like a cautious approach. Now it kind of looks like being left behind. They’ve been on a hiring spree to build proprietary LLM capabilities, but the core dilemma (to integrate AI into WeChat, or not to) is still haunting them.

At the annual shareholder meeting, Pony Ma addressed the “is Tencent AI falling behind” question directly: “A year ago, we thought we were on board, but then we realized the ship was leaking. Now we feel like we’ve managed to get on, but we can’t even sit down properly. We’re still hoping the ship could go a bit faster.” Ma noted Tencent didn’t have particularly strong foundational AI capabilities early on, but is addressing it through talent, team management, and internal training and reassuring investors that they’re gradually getting on track.

Tencent’s AI strategy from the start was less obvious from the outside because it isn’t really selling AI as a standalone product. What we gave kudos to them for was embedding AI into the businesses that already make money and being open to their non-proprietary models. But that’s becoming simply not enough…

Anyway, its near-term justifications have all been about how AI has bolstered existing revenue streams. Ads are the cleanest example of them doing a good job at leveraging AI to enhance $. Marketing services revenue grew 20% YoY, helped by AI-driven ad recommendation upgrades and stronger closed-loop marketing inside Weixin. That’s the near-term ROI. If better models improve targeting, bidding, creative optimization, and conversion prediction, advertisers spend more. Tencent doesn’t need consumers to pay RMB20 a month for an AI assistant. It needs advertisers to see higher ROI.

Tencent also disclosed that AIM+, its automated campaign management tool, powered about 30% of total marketing services spending. AIM+ is essentially Tencent’s version of the AI-assisted ad-buying layer: instead of advertisers manually tuning every campaign, the system optimizes toward performance goals. Structurally similar to what Meta has been doing with Advantage+: the platform absorbs operational complexity, advertisers provide budget and objectives, AI does the optimization. And this brings us back to its superpower. Tencent has the unfair advantage that Weixin, Video Accounts, Official Accounts, Mini Programs, Mini Shops, and payments all live inside one ecosystem. The more closed-loop the transaction path, the better the AI can measure and optimize.

Tencent’s e-commerce path is social, not search-first

Alibaba’s e-commerce AI strategy starts from shopping intent. That’s always been Taobao’s DNA. Tencent’s starts from social attention, which is WeChat’s DNA, and its ecosystem remains its stronger differentiator in agentic AI, as it remains the most natural gateway for agentic tool access.

In recent months, they’ve launched several AI products: Hunyuan, Yuanbao, CodeBuddy, WorkBuddy, QClaw, and more. Nothing has been adopted widely yet.

In Alibaba’s world, users go to Taobao or Tmall because they want to buy something. AI helps them find the product and complete the transaction.

In Tencent’s world, users are already inside Weixin: chatting, reading, watching, following creators, searching, joining groups, scanning QR codes, using Mini Programs. Commerce emerges from that activity. Furthermore, Mini Programs can potentially evolve into modular ‘skills’ that AI agents invoke to execute real-world tasks in WeChat. Without installing separate apps- and we all know that mental hurdle.

The puzzle piece not to overlook is Mini Programs and Mini Shops. And we even signaled this last time we wrote about Tencent.

Mini Shops GMV continued to grow rapidly. Tencent added brand merchant incentives, coupon sharing for repeat buyers, and creator-to-merchant matchmaking. That’s social commerce infrastructure, and AI can improve every part of it: content recommendation, creator matching, product targeting, search ranking, ad optimization, customer service.

Video Accounts time spent grew more than 20% YoY, helped by recommendation model upgrades. Weixin Search query volume grew more than 25% YoY, helped by LLM-powered ranking and broader AI search coverage. These aren’t just engagement numbers; they can be interpreted as commerce infrastructure numbers. Because more time in Video Accounts means more distribution for creators and merchants. Better Weixin Search means easier discovery of products, services, content, and Mini Programs. Better AI ads mean more profitable advertiser spend. Growing Mini Shops mean more transaction-related service revenue.

Tencent’s e-commerce strategy is not to become Taobao, but to make Weixin/ WeChat a better place for (agentic) commerce to happen naturally.

The real Tencent strategy: agents inside WeChat

So we ended the last section on this. Tencent embraced OpenClaw; it pushed out its own version, blablabhlah. Yes, because Tencent’s AI direction is agentic AI.

Tencent has Hunyuan, Yuanbao, CodeBuddy, WorkBuddy, QClaw. It also has Weixin, QQ, WeCom, Mini Programs, QQ Browser, payments, and cloud infrastructure. The plan isn’t to build one chatbot. It’s to make AI agents useful across the whole ecosystem.

This is where Tencent’s distribution edge really helps because a standalone AI agent has to ask for everything: permissions, integrations, APIs, user trust. Tencent already has the user relationship, the chat interface, the payment rails, the enterprise messaging layer, and a massive Mini Program ecosystem.

The most important framing from Tencent is that Mini Programs can eventually be “skills” that agents call. A Mini Program is already a lightweight app inside Weixin: book services, sell products, process payments, manage memberships, handle deliveries, connect to offline merchants. If agents can call Mini Programs as tools, Weixin becomes an agent execution environment.

That means the user doesn’t just ask a chatbot for advice. The agent eventually helps complete the task. Book the appointment. Order the product. Apply the coupon. Track the delivery. Summarize the group chat. Create the Mini Program. Generate the ad campaign. Pull the document from WeCom. Send the follow-up.

That’s the agentic version of WeChat. (More on this in QClaw’s overseas push.)

Which is why Tencent shouldn’t be judged only on whether Hunyuan beats Qwen or DeepSeek on benchmarks. Benchmarks matter, and they’re trying to play catch-up, but it’s not the whole story. Tencent may need the world’s best model if it has the most useful execution layer; it needs the best, most efficient model that is made natively for its use cases.

And this brings me back to something we wrote nearly a year ago as Alibaba is trying to create a better AI shopping experience. If executed correctly, Tencent may actually create the more powerful AI operating environment.

In this conversation, Tiezhen explains why Hugging Face became the GitHub for models and why open source is not just a distribution mechanism but a different way of coordinating research. We discuss why Chinese AI labs have leaned so aggressively into open models, how DeepSeek changed the commercial logic of open source, and why Qwen, Kimi, GLM, MiniMax, and others are using openness as a way to win attention, recruit talent, and accelerate the whole ecosystem.

His core argument is that China’s open-source AI push has three layers. At the researcher level, open source preserves attribution and career mobility. At the company level, open models can become benchmark-led marketing, developer distribution, and a recruiting advantage. At the ecosystem level, government and university incentives are beginning to cultivate open-source culture among younger engineers.

We also discuss why US frontier labs have pulled back from openness as research and business have become more tightly coupled, why distillation is much murkier than the public debate suggests, and how DeepSeek’s releases increasingly function as shared R&D for the broader AI ecosystem. The conversation then turns to monetization: why open-weight labs can still make money through API tokens, base-model access, post-training services, and inference optimization.

Finally, he lays out his current thinking on AI bootstrapping: the idea that agents may eventually help improve their own harnesses, generate training data, and even improve the models they rely on. We close on a more philosophical question: if a handful of closed labs control access to frontier capability, open source becomes more than a technical preference. It becomes a check on the concentration of power.

Tiezhen/ Tom is based in Sydney, Australia. Feel free to reach out to him on X to chat.

Subscribe now

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, and product managers. For more information on the podcast series, see here.

Share

Chapters

04:07 The Philosophy of Open Source at Hugging Face

12:51 Challenges and Opportunities in Open Source

17:12 The Role of Collaboration in Research

21:50 The Future of Open Source and AI

33:58 What Constitutes Distillation in AI

37:18 Navigating Copyright and AI Distillation

37:43 The APAC AI Landscape: Insights Beyond China

43:08 Understanding the Ecosystem: Labs vs. Hyperscalers

46:21 Monetizing Open Source AI Models

52:02 The Future of AI: Bootstrapping and Self-Evolution

Transcript (AI- generated for reference only)

Grace Shao (00:00)

Tie Zhen thank you so much for joining us today. I’m really excited to have you on. We’ve been trying to make this happen for a while and just so glad the timing’s finally worked out. To start, can you tell us a bit about yourself, your journey, and where you’re at right now in your career and how you see the whole ecosystem? And also, just help us understand Hugging Face a little bit as well.

Tiezhen Wang (00:19)

Yeah, thanks, Grace, for inviting me. I know, sorry for the long delay. It has been a while, but I’m recently in transition because I just left Hugging Face. So to give you a quick information about very high-level overview, you can think of Hugging Face as the GitHub for AI. If you are not familiar with GitHub, you can think of Hugging Face as Amazon, where you can find all kinds of models in one store.

And we are helping, so my job is to help researchers to get their models, which is the open source models on Hugging Face. And they can use the best, like all the tools, all the services on Hugging Face to make their models more discoverable and available to everyone. We also offer all kinds of technologies. For example, we allow them to create demos so that developers do not need to download the whole models.

and they were able to try it out and see how it goes. And we also offer services so you can create your own agent using open source models. We do all kinds of scaffolding on top of open source models. another part of work that we do is to help them get more traction. We use LinkedIn.

I use Twitter mostly to help them getting well known by the public. And we write analysis on their models and letting people know what are the new inventions from the model, et cetera. we work with researchers across the world. Like myself, it’s focused on APAC, especially Chinese researchers. Yeah, that’s pretty much the goal.

quick overview of what I do. If you have any questions, just let me know.

Grace Shao (02:03)

And how did you get to this role? Because I understand you were with Google for quite a while as well.

Tiezhen Wang (02:07)

Yes, I was with Google as an engineer. work on ML frameworks. But then we had a bunch of reorg. And I was assigned to a project which is not open-source. But I really like talking to people in the open source world. It’s kind of very different. So when you are paid to work something versus you want to work on something yourself,

Like you have very different mentality and very different feelings. So when I was working on the open source machine learning framework, I talked to people outside Google. And I can see the stars in their eyes. They do want to work on something they want. And even though they may not get paid, et cetera, I really like this feeling. So after I was assigned to the non-open-source project, I want to try something like

new but also in open source and I was like talking to people in Hugging Face and I really liked them. At that time, like Hugging Face was not like part of the mainstream. It was like a niche product for researchers where researchers can upload models. But I do see there’s a huge potential for Hugging Face to grow up because first I believe in open source and the second like Hugging Face is going to be the entry point where like all people will come in and search for open source models. But the most important of all is that I feel that Hugging Face is a company who understands how open source works. Open source is a huge leverage. If you use it well, it’s going to be very powerful. And Hugging Face is like 200 people, like very small companies compared to other companies growing up from the same area. But they are able to use open source as a leverage.

and called for collaborations across the world and do very impactful things. a lot of people, a lot of big companies are doing open source, but they just don’t understand this age. That’s the essence of open source. And I do feel that Hugging Face is doing really well there. That’s one of the reasons why I want to join Hugging Face.

Grace Shao (04:06)

Yeah, I think that’s amazing. I think that’s something we definitely will double click on later, especially when we talk about why China’s labs seem to have been embracing open source. Just kind of one last question on just the whole ecosystem and how hugging face fit into it. What was the philosophy really held by the whole company? Because I actually listened to one of the founders interviews, Clem’s interview recently. And during the interview, he talked about how Chinese scientists have always been long term contributors to open source technology. And then he said it was really like kind of a pivotal moment around 2022 where American open source contributors kind of took a step back and then there was a sentimental shift in the ecosystem. Why is that and how does Hugging Face kind of view the whole ecosystem?

Tiezhen Wang (04:47)

Yeah, there are several questions. Let me try to address them one by one. The first one is the philosophy behind Hugging Face. I think it’s really the mindset. so anything that we see where we can have a collaboration, like Hugging Face will just reach out and see if we can collaborate. So if you go to see a lot of work released by researchers, they will have paper on arXiv.

and also their project on GitHub. And you’ll see me on all of these issue number one, which is the first issue after the repository has been released. And we just write something saying, offer blah, blah, blah. Do you want to collaborate on something? So for anything that we can collaborate on, we will just call for collaboration. And some we’ll go through, some we’ll not. But this collaborative mindset is very, very different from.

like a business point of view. From a business point of view, you will first think, what is my edge and how I win the market, how I compete with others, and what are the end areas. After the competition, what’s the end game, how it will go. So that’s the way of how you can justify the investment and everything. In open source world, it’s totally different. It’s like, I want to do something.

I just say it and I do it and there are developers who want to join in and we do it together and we grow the pie gradually. we do not have like, let me put it the other way. So if you see an open source model coming from one of the Chinese lab, for example, GLM 5.1 is released and you may think like Kimi or Minimax like other open source model provider.

in China would compete with them. But actually not. Like you will see they are commenting on the Twitter saying, congratulations, et cetera. This is a collaborative mindset where everyone is stepping up on each other. we can do a lot of, as a group, can continue to push the frontier forward. So I think this is very, very different.

Yeah, and talking about your second question, the Chinese, well, I wouldn’t say labs. Chinese researchers, labs, companies, et cetera, they all want open source. I think there are three different folds. The first one is on the researcher side. A researcher would always prefer if their work is open source. That’s coming from their academia background, because when you

Like on the CS world, when you write a paper, you have to show that it’s actually working. You have to show that all the numbers are real. Other people should be able to verify that. And you can only do that by releasing your code, releasing your models to the community so that other people can evaluate. So a researcher, after they graduate and they go to a company, they will bring this mindset forward. And by default, they are open source people.

And another perspective is for their self, for the career development of themselves. So as an engineer in big companies, it’s very often that you are working on some project and nobody knows that you are working on that project until you say that out on the game or on your resume. But open source is very different. We know precisely who has contributed to DeepSeek before.

And that’s very attractive for for researchers, because if I have done great work, I want the whole world to know that I’m doing excellent work. This will help me have better branding, help me to do more collaboration, help me in the future step in the career. So a researcher would always love open source, by default. So that’s the first part from a researcher’s level. The second one is from business level.

So well for individual is quite easy to embrace open source from manager level from the executive, they need to justify the investment on open source. I have to spend tens of millions in training a model and you want me to give it for free. That’s crazy, right? That’s how people think before DeepSeek. Although we have lot of open source models before DeepSeek, but the trend is completely changed.

Before DeepSeek, people were thinking, oh, maybe the model is not that good. Maybe I’ll just open source it. But if the model is good enough, maybe I’ll keep it for private. And that’s one of the reasons why you see a lot of people were saying open source is not that good, especially from Robyn. And lot of people do not understand how the open source works.

works. But then people do realize that if they do not open source, they do not even have a chance to stand on the market. Because their model first is not really good. If they just compete on the marketing level, on the business level, they do not stand a chance, not even a chance. So you spend tens of millions and you get nothing. But if you open source, at least you have some sharing and people will remember. And also you can have the market from.

for the researchers. I think Qwen team was one of the first team who understand it from a business level and start like open sourcing work. And as the result, it’s very, very good. Like they almost taken the ecosystem from Llama and now they are becoming the default for researchers to do research, which is like a huge branding for Alibaba. And like, I guess like if Alibaba wants to do any kind of business, like it’s quite easy for them.

to approach to researchers saying, we are not nobody, right? We are the provider of Qwen and everyone wants to talk with them. And another side for the business is that they find it really hard to attract top talent if they do not do open source, because all these talents want their name on papers, et cetera. if they can pay a lot of money.

but they still do not have the best talent. But on the other side, if they do open source and the researchers know that they come to this group and they can have their name marked on history, it’s going to be very attractive. So like this company, even not releasing the best models, they try to release something to make researchers happy. It’s kind of like their...company perk. So that’s another route. But after DeepSeek, everything changed. People know that if I do open source, I can have huge branding for my company. DeepSeek is not doing any kind of commercial stuff, like alteration to cusTiezhen Wangers. Yet they still have a huge evaluation of, I think the most recent number is [unclear: “14 million HKD” in transcript; confirm figure].

That’s a lot of money. So by doing open source alone, they can make money. And that changed the mindset for lot of people. so after DeepSeek, Kimi, GLM, Minimax, and StepFun, they all come into this open source world. actually, they have made a lot of success stories, like GLM and Kimi, by doing open source, lot more people understand them. And they kind of open up.

the global market, not just the market in China. for them, I feel that it’s not like losing a lot of money because they doing advertisement in a different way. Kimi was spending tens of millions RMB per year on advertisement. And the result is very short retention. People know them, come to their side, and they do not feel any different. And they just move away. Now, the researcher team, the manager, the executive means, knows that the best score on open source benchmark is the best advertisement. So they can concentrate all their power, not wasting them on advertisement, but concentrating all their money and resources on training the best model. But this best self, it’s the best marketing, and they can create great models and start earning money.

So I feel that on the business level, everything starts to make sense. But now there is a new challenge, which is how you can stop people from taking the free ride. It’s a longstanding problem for open source. I did something, for example, I made a database. I spent a ton of engineering hours. I open sourced it. But I’m not making any money, because the cloud provider is taking that for free and start making money and monetizing it.

it’s happening for open-source world as well. I open-source the model and all these inference providers and chipmakers and BDA-AMD are making money, but not the researcher who created the initial model. That’s why you see some licensing change and discussion on that. Kimi did the first non-commercial license, and then MiniMax made a more restrictive version.

Tiezhen Wang (13:40)

made a more restrictive version. But I don’t think that’s the final version. People are still trying different things. And I believe maybe in one or two years, we will have a more standard way of balancing open source and commercialization, et cetera. So that’s the second level. The third level is the third level. So the Chinese government is really encouraging people to do open source.

If you do open source, you have extra credits on your bachelor education, et cetera. And Shenzhen recently announced a very interesting policy. So you can have housing points if you do open source on GitHub. basically, they are categorizing.

Grace Shao (14:21)

So the incentive, yeah, go straight to the students, like even in academia, while they’re still in university.

Tiezhen Wang (14:27)

Yeah, so it’s kind of cultivating this open source culture when other researchers and developers are still in universities, which is really good. So I do feel that the culture of open source is, if they are winning the young students, we are going to see more open source projects. And to be honest, I do feel that that’s the right approach.

Because if you’re not thinking about open source, you are thinking like traditional way of collaborating with people, which is company or corporation. And I feel that the essence of why we had cooperation or company is not keeping peace with how we evolve now. I think about, you set up a company in Hong Kong 200 years ago. Why? Because you have a group of people. You want this group of people.

That’s why it’s called company. You have a group of people and you want them to work together. And how you can make sure that everyone had their benefits. Everyone is doing a lot of work. Obviously, they want to have a return. And you do that by setting up the shares and also the voting system. that’s how a group of people is working together. But now the word company has changed. It’s more like a

multi-international company where the worker in the company has no work in deciding how the company runs. Whereas open source work is more likely the original version of a company. You have GitHub, you know who has contributed what. Everyone knows your contribution, and you can have your name listed. the group of people coming from all around the world, can.

collaborate on something. They do not need to be part of a big company going through all the interview process. They can just collaborate. So I think that’s very, very interesting. And now with Zoom, Tencent meetings, and all the Google Docs, it’s much easier to collaborate internationally. I don’t need to know who is contributing to the PR, but I know someone is interested in my project, and we can work together. And I feel that.

That’s probably the future way of how people can collaborate. that’s to end the last point on society level. I think the society is advocating for open source. also open source is probably the way how the society will evolve.

Grace Shao (16:48)

Thank you. is like so insightful pack that I have to digest that. But you you mentioned quite a few different topics, which I can definitely take this straight, conversing different directions to start. have two questions and they’re actually unrelated. So one at a time. Number one is you really make a point about China being really, you know, strong advocate on open sourcing the LLMs. However, I think

Could you tell us the history of open source in China in general? Was there a tradition to want open source technology even pre-LMDs? That’s number one, first half of that question. Second half of that is you say there’s a lot of incentive for researchers to actually want to open source everything, right? And then therefore they can claim their contribution. Well, in the recent interview between Zhang Xiaojun and...

deep minds, Yao Shui Yu, I think maybe you’ve also listened to it. You know, one thing that really stood out to me was how he was saying people need to be like responsible. And like for someone who’s not technical, I actually really struggled to understand what he meant at first until like actually Jiang Xiaoxuan actually asked him to clarify as well. His whole point is that in academia, people are so used to only claiming a certain section of what they contribute. So for example, for a big piece of paper or research,

that you would take credit for what you contributed, right? And you want to make sure that it’s best optimized, known, heard, seen, whatever, right? Recognized. However, in terms of how LLM can work properly in terms of the long run, whether it’s like, you know, further in post-training and further, you know, know, usage, whatnot, it’s important that people don’t claim so much credit to their own part of the work. It’s more important that people work collaboratively. But kind of to your point on open source that, you know, they can work collaboratively and make sure that each piece works together better instead of each piece working best on their own. So it kind of contradicts your comment on why people want open source, because in that sense, wouldn’t it make sense for people to not want open source? I don’t know. That’s another question. And the third part of this is really if open source makes so much sense for tech companies and makes so much sense for academics.

then why are the American labs so anti open source right now? Like what is driving that? Is it purely because commercial reasons or philosophical reasons? This is very big, but you did throw a lot at me. So I’m going to throw these questions back at you.

Tiezhen Wang (19:07)

Yes, sorry for my very long answer. I think it’s probably by itself worth writing a blog post with enough content, and I can elaborate more. But great questions for the story. Can you remind me? I guess we can go through them one by one. Can you do mine? Yeah.

Grace Shao (19:25)

Just like in general, source China, China open source. What’s the sense on that? Beyond LLM, right? Like why did Chinese companies always contribute to open source technology? Clem talked about this in his interview, but he didn’t go into that about it, right? So number two was just about, yeah, number two was just about like, why do these academics want to claim their names, right? Is it better for the company in the end or is it just best for them, like the selfish reasons?

Tiezhen Wang (19:37)

Yeah, okay. Let’s try it. Yes. Mm-hmm. Yep.

Grace Shao (19:52)

And number three is why are American labs kind of anti open source right now?

Tiezhen Wang (19:56)

Yeah, so let’s try to address the first one. I think it’s a great question. And I do see the shift. So I feel that AI is probably one of the very few areas where Chinese open source contributors dominate. If you look back to, for example, I would say the initial days of modern open source comes from like an

Linux or Apache or database and everything. And where you do see a lot of individual contributors from China, but you are not seeing enough Chinese company creating a project. And then the project gets adopted globally. You are seeing that gradually when we move to the area of cloud-native, like when the Kubernetes comes out.

And a lot of Chinese cloud providers are trying to really pay attention to this whole open source world. And you will see that this grows. But now it’s like this. So it grows exponentially. So I think it comes from two folds. The first one is the Chinese participation in the global market. It needs time to warm up.

Like for example, lot of Chinese contributors, they can only contribute two projects in Chinese because of the language barrier. So that kind of limits how much they can actually do. And now with larger language models, with better education in the new generation of developers, the language barrier is not that strong. That’s why.

That’s how the Chinese open source contributors can make a better impact. And another one is, so in the traditional way of a company’s, like how a company’s structure itself, if you do open source project, it’s kind of hard to justify your credits because the open source by itself is not the core business of a company. There are very, very few companies who

had their core business made on open source. Like PingCAP could be one of them. PingCAP start with open source and then find monetization plan. But that’s so small. So few of them. And in the new areas, a lot of companies, their core business is open source and plus monetization. Even for IPO companies, for public list of companies, Minimax is basically one such example. They have their best models, open source.

and then trying to make money. So this is very different. If your core business is open source, of course you will put more resource on open source. And it’s more likely for your project to gain a lot of developers. And I feel that the third one is the international collaboration has never been easier before, apart from language barrier. So after the pandemic, I feel that all of a sudden everyone is used to

like Zoom and Hangout and collaborating with someone who you don’t see face to face. And this is a great chance for open source project to ramp up. Because before that, you have to meet face to face, and the bandwidth and the people you can meet is kind of limited. And now you have a huge, like, as long as your project is great, like you have a huge pool of potential developers.

And the last one is probably AI by itself, like coding agent itself. Although it does make code review much harder because there are probably a of AI scope. But it really lowers barrier of who can contribute to an open source project. Before, you want to contribute to a project. They are probably developing language you do not know. And also, the code base is pretty strong. A developer might not be able to.

contribute to the project until he has a very thorough understanding. And that’s probably like months of work. Now you can just ask AI how this part works. And I only need this feature. And what are the code I need to modify? And I can just give it a test locally, and it works. I contributed to some Rust project without being a Rust expert. So that’s how AI makes everything better.

So I think it has all these reasons. There are probably more, but I think someone at the end of the day, in two or three years, maybe starting to write some history about how every single aspect of technology, moment and everything, people’s mindset shift, how to cultivate the open source spirit. But I think, yeah, that’s the...

Top ones coming out of my mind.

Grace Shao (24:26)

And then the second question was just that would researchers focused on their own name and frankly ego in this sense actually be the best way to help cultivate the best LLM or whatever whatever product that’s the end product that’s to be shipped. Does that make sense? Because it kind of contradicts what Yao Shunyi was saying on Zhang Xiaoxuan’s podcast, right? He was saying in that sense a lot of researchers

will try so hard to only own what they are working on, but they have less of a sense of responsibility for the bigger project.

Tiezhen Wang (25:00)

Yeah, I haven’t really read the broadcast entirely. But talking to your point, feel that open source or not, or having researchers name these data on papers or not, it’s actually a game changer. If you are a researcher and you might want to stay in academia, why? Because all the papers you publish is very important to your career. Everyone sees.

Like you have published which paper with like who and the paper well like also mentioned that you contributed which part of work are you Look like my corresponding author. Are you the main person or you are just like contributing a part of it? And you there’s h-index to measure the impact of a researcher So like you have all that infrastructure working for you if you stay in academia But if all of a sudden you want to start working

for company, you lose off that because the company might be able to say, we have a policy where all your open source and all your paper publications, even writing a blog is controlled by the PR team and they have to decide if you can do certain level of things. So your exposure is reduced. You might be very happy because you earn much more, like 10x the salary compared to be an assistant professor.

Like after five years, if you ever want to go back to academia, that’s impossible because you lose all of your track record. And with open sourcing, things are very different. The top nailing app wants to hire the best talent from academia. And the people from academia wants to work for the app. But at the end of the day, the two system

It’s the same system because all the record is public. So it’s very easy for people to come in and come out, come in and come out. It’s kind of different from people coming from academia and then lose track in and get lost in the company world. So I do feel that having your name listed on the work you publish is very important. It is kind of the concept of [Chinese phrase unclear]. So you.

your branding grow with whatever you have done. So your reputation is based on whatever you contribute. So you need to pay extra attention on that.

Grace Shao (27:16)

I see what you mean. And the last bit of just now what we’re talking about was just why is that if you believe open source makes so much sense to these researchers, that so many researchers in the US or at least some of the companies, the entities right now are not willing to go open source.

Tiezhen Wang (27:31)

Yeah, there are lot of companies who change their position. Google used to be the company which impressed open source the most, like Google open sourced and like TensorFlow Kubernetes and bunch of other important open source projects. The open sourced transformer, which is the cornerstone of our modern AI system. So Google was really impressing open source. But I feel that

At some point in time, Google stopped doing all the open source work because they kind of lighting OpenAI and other companies taking the free ride. Google did a lot of fundamental work and then it’s kind of taken by other companies for free. And also, OpenAI and Anthropic, I feel that they are still contributing to open source, but that’s not their main project.

Their main project are hidden secrets that they do not want to share so that other people can catch up. And the research and business is getting so coupled. So for example, a researcher in OpenAI found a way to improve the intelligence level by 10%. Let’s take o1, for example. They managed to find out how to make model think. And by this chain-of-thought and thinking process, the model is

like way much smaller, way much smarter. But they do not want to share the gist.

As long as they do, other people will catch them up. So it’s kind of a very restricted environment. Although researchers may still want to open source some of their work and have their name listed there and sharing very detailed observation. But the business doesn’t just allow them to because you are trying to. Yeah. also, researchers can also talk to like in some conference. I know how open source, sorry, I know how o1 was roughly made by reading bunch of YouTube videos from the internet made by OpenAI researchers. But after that, you see less and less very detailed research sharing, even the videos or recordings by them. I guess they kind of learned the lesson.

But like, yeah, yeah, that could be part of the rhythm. On the open source world, like, it’s kind of different. Like, so before DeepSecR1 was released, lot of people were speculating how OpenAI was doing o1, and they’re trying things in different way. And after DeepSecR1 is released with all the recipes and all the data they shared, like, the open source world seems to be converged on the path. Although that path might

Grace Shao (29:46)

There can’t be compliance reasons.

Tiezhen Wang (30:12)

be different from o1 because we never know how o1 was made. But then because of DeepSeek’s contribution and sharing, everyone knows how to make thinking chain. And the whole ecosystem is evolving really, really fast. That’s one of the real value of open source because everyone can just collaborate. No one is holding secrets. Well, there are still a lot of secrets on how you can run as efficient as DeepSeek, but that’s

Like too technical, like that’s not too much on the research side. yeah, like I...

Grace Shao (30:44)

So on that note, yeah, I want to follow up on that. I think I recently wrote about something which is, speaking to our researchers, it got me a sense that DeepSeek in a way is now becoming essentially like a foundation for everyone because, you know, a lot of the labs in China are looking to DeepSeek to see if there’s any like, you know, engineering breakthrough, like your point, and they build on top of each other. Help us understand like each of the labs, because you said, they’re cost constraint, they’re compute constraint.

Tiezhen Wang (31:06)

Thank you.

Grace Shao (31:12)

They’re teleconstrained, right? Their resources are constrained and every single asset you can think of compared to the American peers. Now, why does it make sense that they all open source and how are they all optimizing for their own goals at this

Tiezhen Wang (31:24)

Yeah. So open source by itself, as we just talked about, is an accelerator of the whole ecosystem. So DeepSeek shared all their like, knowings and discoveries and what things work, what things doesn’t work. This by itself is accelerating the whole industry, not just Chinese open source, but also like US open source and US like closed source. Like they just don’t say how much they learn from DeepSeek, but I believe everyone is learning from DeepSeek.

Not just that, DeepSeek also contributed to GRPO, which has become the most used algorithm, reinforcement learning algorithm in the industry. So they did a lot of contributions. if you check recent model architecture evolution, what’s proposed by DeepSeek is becoming the standard and getting adopted by many people. For example, Kimi 2.5 was using a model architecture very similar to DeepSeqs. And GLM 5.1 was adopting a lot of components from DeepSeek architecture as well. So it’s kind of sharing and learning and co-evolvement is one of the, I would say, secret of how China is able to. catch up with the US in certain area, although having restricted compute and restricted capital, I would say. If US open source is working again, like the whole ecosystem, like everyone was trying to open source, I would say the human race would be evolving much faster than what we are doing now.

Grace Shao (33:01)

So on that, how do we understand the accusations of what is being distilled? What is technically shared? What is, how do I understand the gray area of that? Like the accusations from a lot of American labs, Chinese labs right now, like you just said, a lot of American labs are learning from Chinese labs. Frankly, within the researcher community, it’s not even Chinese versus US, it’s really just labs with each other and against each other if they have to, right? Intellectually competing. So then how do we understand what the, industry agreement is on a distillation, why is it so contentious right

Tiezhen Wang (33:32)

On distillation, yeah, that’s a great question. I can only give you my perspective. first, distillation is a very broad word. We are distilling from each other as well. I learn from you, you’re learning from me, and we are all learning from books and papers and all this public information. So I would say,

distillation is a very common practice, like basically how you learn from others. Like you might have a model which summarizes the books and like doing bunch of explorations. And the way for the model itself to move forward and evolve is to distill from its historical data and historical experiments. And like that works for like another model trying to learn like your model as well.

And on the research field, distillation is very common. DeepSeek R1 was released with MIT license. Specifically, so I actually asked the team about it. They choose MIT license because they want their model to be distilled by others. Because that was the only model that works really well with the thinking chain. And they want all the open source model to be able to have that.

Like they have shared all the recipes, but others do not have data. So DeepSeek design like they’re like small models so that and also the recipes so that other people can easily distill DeepSeek, getting the thinking chain and use that on their own models. like this distillation is happening like everywhere. And I think like US companies are distilling from each other as well. Like I’ve seen like the recent discussion on Twitter in public.

where Elon Musk and Sam Altman were kind of battle on that. yeah. And if you think about it the other way, so if you do not allow a model to distill, I mean, the output of a model to be able to train a model which is from a competitor, it’s kind of a very interesting point. Like if we say, I’m reading a book.

I’m telling you the story. So you, after reading the output from me, which I think of me as a model, you’re reading my summary and you are not allowed to share the summary to others. You have to read the book, the initial book, not using my summary because of the license, et cetera. That’s kind of ridiculous. That’s not how human transfer knowledge in the past a few thousand years.

Like I have a very bold argument. I think that anything like generated by AI should not be copyrightable. So like it should be in public domain, like anything generated by AI, because like anything generated by AI is a distillation of like human entire history and everything that human has created. And if you just take that for free and asking other people do not use that.

Like it’s kind of a waste and it’s kind of like blocking people from evolving forward. Because like human content do not have this restriction and why you are putting this restriction on something not copyrightable and generated by machine. So that’s something I do not really understand. So I do see there are terms and conditions saying that my model output cannot be used to improve other models. But I don’t think that’s kind of valid.

I’m not sure if someone eventually will file something on the court and we can have a case on that. currently, think there are a lot of things to discuss, but it’s not about if we can distill a model or not, but about something bigger. Should the model creator even have this right to restrict others from distilling from their models?

Grace Shao (37:18)

That’s really interesting. I think that a lot of the discussions in the public space is really about whether you can use copyright work of human output. And then the argument is always like, just you cannot distill because the company said there’s no distillation allowed. But like to your point, there is no actual clear black and white rule of regulation around this right now. And in fact, it’s it’s bit murky. Yeah. Yeah, yeah, that’s interesting.

Tiezhen Wang (37:37)

I’m not a lawyer, but I can find a clear answer on that.

Grace Shao (37:43)

Okay, I want to kind of go to China. Like we’ve kind of talked a bit about the big picture. Well, a lot about the big picture. But let’s look at just the China labs. mean, I know that you represent APAC back then with Hugging Face and you worked around APAC, you lived in Australia. But for the sake of this, know, Chinese labs right now probably are the most relevant out of APAC. Do you think I’m missing anything actually on the APAC conversation? Like, do you think anyone else in the region is relevant in this space that we can talk about?

Tiezhen Wang (38:08)

Korea is doing really, well. Yeah, Korea is really well. Well, the most, one of the best model is probably Upstage. they, initially they create a Korean model leaderboard, like open source version of like model leaderboard. Well, no, no, the leaderboard was not funded by government. The, the, the,

Grace Shao (38:10)

Yeah, yeah, give us some picture on that. That’s funded by their government, right? That’s their government funded.

Tiezhen Wang (38:26)

So Korean, it’s actually a very impactful country, but as the other days, there aren’t enough Korean data. Even for ChatGPT, I think until ChatGPT 4, the model doesn’t speak good Korean. So the model was able to speak very good Chinese from day one, like from ChatGPT 3.5, but because of the data volume, et cetera, speaking Korean was always a challenge until ChatGPT 4.

At the time, like now, the open source model is able to speak like Korean. So Upstage create a leaderboard. So the way they solve problem is very interesting. They’re not solving problem by solving problem. They’re solving problem by helping others to solve the problem. So instead of creating a model right away, they create.

Grace Shao (39:09)

I heard about Upstage from VC in Korea as well, but I don’t know the detail about it. Tell us more about who they are, what they’re doing.

Tiezhen Wang (39:15)

Well, I don’t know too much about who they are, but I only see their open source contribution. I think the founder is a professor in Guangzhou, but he’s Korean and moved to US. Correct me if I’m wrong. I’m sorry. I’m not really up to date with that information. But I just want to call out because I think that’s a very interesting paradigm. For example, if you are a company, you have your own problem you want to solve.

Tiezhen Wang (39:42)

Like, how do you want to solve it? Like, you are going to hire some people and define a problem and try to use your own people to solve it, right? So that’s the old way. What’s the open source way? Is you publicly define the problem. You have a leaderboard. Like, you might do a private eval or public eval. It all depends on you. the problem is you have to list your problem.

publicly and you have to tell everyone that you can contribute to this problem by submitting a model to a URL and we will do evaluation and see how each model is evolving on this area. So they basically have a leaderboard. And you will like a lot of researchers would be very interested because now they have a problem to solve before they do not even know Korean what’s the problem. So now they have a problem to solve and you will see that the curve goes like.

it goes up because there are more and more researchers coming in and all their work are open sourced. So a new researcher wants to jump in the field. They will first have a look on the leaderboard to see how far away from a really usable benchmark. And then he can investigate all the previous attempts and find his own way of kind of just changing something a tiny bit.

and apply that to the past people’s work and submit to the leaderboard. And now we are seeing people making progress on the leaderboard. So that’s a very, very clever way because it’s not one company solving the problem. It’s like we are opening the door for everyone to come into this playground and try to solve the problem together. I think within a few months, they were able to get thousands of submissions.

which is really massive because just imagine you hire 10 % people, you won’t get that. And now it’s by this new way of doing things like building public, evolving public, you’re having a lot more submissions and you are educating people, et cetera. So they have this very impactful and inspiring leaderboard and then they release a model called Upstage for something. I can’t remember it has been a while.

And the Korean dataset and the Korean models are accelerating very fast on high-netics. I think it is now the fourth largest models, speaking Korean. Yeah.

Grace Shao (42:04)

Very interesting. Yeah, I’m going to shamelessly self-plug in. People can listen to the episode I recorded with one of the leading Korean VCs as well that was published last week. He gave a good AI ecosystem breakdown of stuff.

Tiezhen Wang (42:12)

okay. Yeah, could you help me like do some like DD first and like just make sure that are correct. Yeah, you can.

Grace Shao (42:22)

Yeah. No, no, no, he did talk about Upstage as well. It’s very interesting. Yeah, I want to... sorry, go on.

Tiezhen Wang (42:29)

Yeah. And also, so you asked for APAC. So in Singapore, there are a lot of great researchers, like lot of Chinese researchers will go to Singapore as well, like Cancun too. Yeah.

Grace Shao (42:43)

Yeah, I think the ecosystem is a bit overlooked by I think Western markets, but definitely there’s a lot happening in around Asia. Like APAC has been including Australia as well as Southeast Asia, East Asia, and Northeast Asia. Okay, I want to bring it back to China. We’ve been kind of talking about China kind of more on the high level sense. Now looking at the companies themselves or the labs, we want to break it down. Just give us a sense like, how do we understand moonshot?

Mini, Max, Deep Seek, Zhipu, if you have to put it in one bracket, versus the hyperscalers, Tencent, Alibaba, and ByteDance, in terms of their strategy, in terms of the capabilities. Like how should we understand this ecosystem right now? Are there other relevant players that you think I’ve missed, maybe like Xiaomi or anyone else?

Tiezhen Wang (43:25)

You mean like how the model creator, model lab, are collaborating with hyperscaler? Is that your question?

Grace Shao (43:32)

No, no, I just think it’s like the people, the people who are creating LLMs, like are researching on how to deploy LLMs. These are the main players, right? Now, how do they defer? How are they similar? What are we seeing like on the ground? Are some of them becoming more irrelevant? Are some of them becoming maybe say, we just talked about DeepSeek becoming almost infrastructure provider for the whole ecosystem.

Tiezhen Wang (43:39)

Yep.

Grace Shao (43:58)

You know, are that mini-max is very, focused on multimodality. Zhipu is very focused on coding capabilities. know, Alibaba really trying to push out commercialization by their existing applications. How successful that is, that’s a different question. Just like an overview of these players.

Tiezhen Wang (44:14)

Yeah, I do think they’re kind of converging. Yeah, because everyone knows that coding is, the whole market for coding is booming. And if you have a good coding model, you can sell it for profit, for large profit. And I do feel that everyone is rushing for coding. There are people exploring different things, like,

For example, Tencent is putting a lot of efforts on Hunyuan and doing OCR stuff. And lot of other companies are doing video generation. But at the end of the day, think from a strategy level, I don’t feel that there are a lot of difference. It’s more likely a case where, you have data? For example, it makes a lot of sense for ByteDance and Kuaishou to work on video generation models because they have a ton of data. And also, do you have?

like a large enough scale. Like for example, Kimi is not very active in making all the apps. Like Tencent is making models. They are making like great apps. Like for example, Yuanbao, like a QA app, like based on all the Tencent data, it’s very popular. Like they make QClaw. Like Tencent is able to do that because Tencent has a huge talent pool. Like Tencent is a huge company, Whereas like if you look at the Kimi, Kimi is very conservative in...

doing all that because Kimi is still a very small company. So I think from a very high level, everyone was on the same page about the strategy. It’s just more, how much resource do you have? What are the advantage of you? Do you have data? Do you have distribution channel? Do you have product design, success story, et cetera? So yeah, I’m not sure if I answer your questions.

Grace Shao (45:57)

No, no, that’s good. So we kind of talked about why researchers want to open source. We talked about these companies are somewhat doing the same thing. So then this leads me to the question. We know that open source, open weight does not actually mean they don’t make money. However, obviously means that it’s harder to commercialize as we like alluded to with the US labs, why they make those decisions. Then how do these companies find ways to monetize and sustain their businesses then?

Tiezhen Wang (46:21)

Well, in the US, are also labs dedicated in making open source models and still making money from other donations or from other parts, like selling apps, et cetera. It’s basically the same way in China, too. For example, DeepSeek is run by, I would say, donations from the people who play the stock market.

like there are labs run by VCs and lot of labs are already profitable by like selling tokens like GLM has recently raised the token price because like they see a huge number of demand and they’re like running short on compute. yeah, like open source can make money. Like there are a ton of ways for open source model provider to make money. I have a lot of ideas. if in case you are interested in like making your

not profitable, can contact me. But honestly, there are lot of ways. The simplest way is to sell token. If you have the best model, you can sell a token for profit and people will actually buy your token. so it’s very interesting because when we combine science and technology, always consider it’s the same thing.

Grace Shao (47:15)

Yes, everybody find Tiezhen Wang.

Tiezhen Wang (47:37)

For model, it’s the same. When we think about models, we just think of a model that generates tokens, et cetera. But actually, there are two different parts. The first one is training, where you have the model. And after you get training, open source the weight you trained. Another part is the inference. So you need to run a lot of optimized CUDA kernels in order to make your token cheap and fast.

Either bracket can make a lot of money. For example, you can open source the fine-tuned model, not the base model. So if a company want to use open source model for fine-tuning on their own data, they cannot be building on a fine-tuned model. cannot build. They have to find the base model. And if the base model is not open sourced, you can sell that for profit.

And also different clients might have different requirements on the model. The NeoLab can collaborate with the client directly and provide some kind of training and post-training support. So that’s a way of making a lot of money, actually, because training is very expensive. It involves very expensive researchers and data and compute. On the inference side, too.

Grace Shao (48:45)

Yeah.

Tiezhen Wang (48:52)

Because the inference is tightly coupled with the data center you own. So your optimization strategy does not, there’s no guarantee that your optimization will work on a different cluster. So a lot of people just do not open source the inference recipe because it’s not that useful. And also it’s kind of a moat. So the model provider who creates the model, they know how to optimize the model best.

when the model is released because they have seen the model for four months and they have done a lot of optimization on the model inference. And when the model is out, like everyone else, it’s just starting to know the model and doing some optimizations. So of course, the model provider will sell token in a much efficient way compared to all other competitors. Three months later, when the outside inference provider gets to

know all the secrets and do very optimized kernels, there’s a new model coming up. So the model maker, the people who know the model from day zero, always have an advantage on selling the tokens. So that’s one of the very important ways how they can make money.

Grace Shao (49:59)

I see what you mean. Mm-hmm. Yeah. And does DeepSeek v4 coming out have an impact on how the GLMs of the world or Kimi make money? Like essentially their strategy with the fact that you just said they just raise prices on their tokens.

Tiezhen Wang (50:16)

Yeah, so GLM and Kimi doesn’t sell DeepSeek or Qwen. So they are not competing with each other directly. I would say the capabilities are on par with each other. So it’s more like a user test. Which one is better? There is no clearly winning between all three models. So we’ll see like Zhipu’s stock price was getting down because people were so worried about DeepSeek. But then they realized that like the

Zhipu token selling is not quite impacted, so the stock price bounced back. But at the end of the day, I would say it’s actually a good thing for them. So GLM 5.1 is adopting a lot of core design in DeepSeek with 3.2, I think, model architecture. And they were able to cut down the cost.

by adopting all these exploration from DeepSeek. And now V4 Pro is out. I don’t know the details, but a very simple guess is that Zhipu is able to cut down the cost because they can adopt new things from DeepSeek architecture. So Zhipu on one side, because of the demand is so high, so they can increase the token price.

and they can learn from DeepSeek and cut down the cost. So Zhipu is going, yeah, exactly, exactly.

Grace Shao (51:34)

you have a higher immersion. Yeah, this is something I think they’ve talked about as well, like really being able to learn from the engineering breakthroughs that DeepSeek puts out every time. Okay, I have mindful time. I just kind of want to have a few questions on the future outlook. You posted on X recently saying that you’ve been thinking a lot about how do we make AI bootstrap itself? And you you’re going through this transition yourself, you’re thinking about the future of AI. What does it mean for the open source future as well?

Tell us a bit about where you stand right now and how you think of this bigger picture.

Tiezhen Wang (52:06)

Yeah, I’m still doing some exploration on my side. I think this whole AI bootstrapping logic has already been implemented by a lot of big lab internally. The idea is very simple. In compiler world, you can design a programming language and write a compiler probably in well-known languages like C. And then you will first implement this language using the C code.

In the next iteration or after a few iterations, you are able to implement this language using your own language. So it’s called bootstrapping. You are basically evolving on your own. You are not relying on something which is not from your language. So it’s like putting it another way. If you see how normal living creature, how they replicate itself and how they evolve.

I don’t need to have a screwdriver somewhere to engineer my kid, right? My kid’s just born. All by itself. But how far are we from AI to do similar things? Now we have a coding agent very powerful. We have our AI training pipeline recipe kind of stabilized, at least for small sized models. So are we really far from

like AI able to get one of my idea, like I give him the direction, and he’s able to like first bootstrap a very simple version and gradually evolve towards that goal. Like I think it’s like highly possible. So at the end of the day, we might be able to like just tell him what I’m going to do without like giving him all the harness and all the like detailed guidance and.

I’m not talking to him 100 times, and he’s able to first lay out what he needs to do and have a plan, and then probably design a DSL or agent all by himself. And probably he will create a model ways to help him to get adapted to this goal. And then he can just keep evolving. All I need to do is to give him more fuel.

which is compute, and he’s able to do some evolution and all by himself. It’s kind of like if you have recently read Andrej Karpathy’s Twitter, there’s a concept called auto-research. But auto-research is just evolving on the model weight. It’s not evolving on the agent and harness. I think on the agent level and harness level, there are also a lot of things to do too.

So I’m quite new on this journey. What I was able to do is to bootstrap a very simple agent and I can use that agent to optimize the agent. But I think eventually we will get the weights involved too. When the model realized, okay, I’m not just needing an agent, I can create a bunch of data and improve my weights. He’s able to evolve from that too.

Grace Shao (55:05)

So in the future, how important is the capability of the models versus the harness and then the industry expertise then? Because right now, so much of conversation is still about, you know, the models are very strong. We are seeing what you’re saying already, the agent’s starting to build out things auTiezhen Wangatically. But we still need the taste. We still need the industry expertise to guide them. I find it hard to imagine that, you know, you can plug in something just say, want this to be done. And the agent just starts doing it exactly to your taste and your...

imagination? Do you really think that’s happening?

Tiezhen Wang (55:34)

Yeah, I do feel that it’s happening. Like we are using agent to like especially coding agent to code something that we are completely unfamiliar with. And I’m quite confident that it will actually work. The reason is that like I have defined a set of goals and as long as I see that is moving towards that direction, like I’m good. I do not need to understand the code line by line. Like it’s just the box. But like the difference is I’m using the coding agent.

Grace Shao (55:58)

Mm-hmm.

Tiezhen Wang (56:02)

to do something else. What I can do is to use the coding agent to improve coding agent itself. And using the coding agent to generate the data and train the model that coding agent is using. And I would call that a bootstrap, not like just using, like, I think I’m already quite happy with coding agent to do something else. But just like, yeah, yeah.

Grace Shao (56:23)

Interesting. I want to end on a more philosophical note. So do you view the argument that AI is going to replace humans then? Or do you think AI is going to be in the role to support humans if we could keep on going down this path?

Tiezhen Wang (56:35)

Well, it’s actually a very, interesting question. And I feel that people do have different feelings. But from a pure technology point of view, I do feel that it’s one condition of like, so technology is not the only thing that will decide everything. You mentioned that if it’s going to help humans, well, it’s not really technology by itself to decide.

Like it can be used in different ways, in different social structure, in different like tradition and such. It’s like giving you a gun and you can do things in good. Yeah, it’s not. Yeah, yeah. But like just imagine that you’re going back to history with all your knowledge of modern society. Are you going to help the

Grace Shao (57:12)

It’s not a good analogy. But yeah.

Tiezhen Wang (57:26)

like the society, like the history you go back to? Or are you able to help? Like I think it’s basically the same. If you have AI that knows everything, you can just think of it as human being in like 2000 years in the future. And now you have it. And what it is going to help on the society. Like it really...

Grace Shao (57:31)

Yeah. It’s like that saying, your own capability of using it is the cap of itself. Also, I think there’s a lot of argument and discussion around the fact that even the society as we know it today, the knowledge work that we all have, that we normalize, are not even created until the recent 100 years. And if AI is to disrupt that and replace human in that sense.

Why is it so bad? Because it alleviates us to do other things that human multifaceted beings that we are can do. Is that kind of part of the argument as well where like, even if it replace us or helps us, it’s only helping us actually alleviate some of the things, if you take a step back, the things that we don’t want to do, right? Where we can maybe go touch grass. I don’t know, maybe this is very optimistic view of it, but there has been people saying like,

The cap on AI capability is a cap of your own intellectual, your own cap of your own ability to navigate or use AI. So the more you can use AI, the more it can help you. The less you can use it actually, the more it will replace you.

Tiezhen Wang (58:45)

Well, I think it’s very interesting to define what is you. Are you defining you as everyone, or are you defining you as people who have compute? Well, no, it’s not. It’s actually a very, very, very interesting question happening right now. You know, Anthropic coding agent is able to do lot of things. But people are of imagining that we are

Grace Shao (58:53)

We’re getting really philosophical now.

Tiezhen Wang (59:09)

Everyone is getting a lot more powerful with models. But what if one day Anthropic just say, you cannot use your coding agent to do certain things? It already happened. Anthropic said, you cannot use your agent to do auTiezhen Wangated tasks. The other thing, there could be other limitations. I have a very bold argument is that the reason why we are able to use AI

so cheap that even us, like we do not own a data center, right? Even us can use that. It’s because our data is still valuable. You know, if you use subscriptions, your data is going to be distilled by Anthropic to further improve the model. And like they are able to give us a discount because they still need our data.

Grace Shao (59:52)

So your point is that one day when they capture enough data, they will not even give us this kind of access for free or for cheap price.

Tiezhen Wang (59:59)

It depends on how they define you. You ask them, like, you or something. How they define their user. How they define who could be part of the game. Like, if one day, like...

Grace Shao (1:00:09)

So then the question is, no, but then my question is, then there’s another argument where they’re saying too much power is in the hands of a few companies right now, right? Or a few founders, what not. We need open source, that’s your point, right? No, that’s really interesting. And that was actually gonna be the last question I was gonna ask you. What is one differentiated we hold? And I think you’ve already answered that in that sense, right? Yeah, I think it’s for us to really think about it. But then as the average user, my question is, how do you actually boycott?

Tiezhen Wang (1:00:18)

That’s why we need open source.

Grace Shao (1:00:36)

these companies or if not boycotting, how do you actually make an impact? Because if I’m not the developer creating an open source model for the average person to use, me as an average user, what do I do?

Tiezhen Wang (1:00:47)

Well, just use the model to do the thing you want to do. Try to embrace the model and be more patient for open source models because obviously the open source model is not as good as top tier closed source models. you kind of like, well, I mean, with open source models, you keep all your secret to yourself. So you can have.

like better security and you have better control. Open source model will never betray you if you just write on your local laptop. So although the model is not performing as well because he’s not distilling you, right? So still you can trust on your open source models and give it a more task to do.

Grace Shao (1:01:20)

You host yourself.

Tiezhen Wang (1:01:34)

I do feel that a lot of open source model is actually capable of doing things. But the expectation might be, think of it as six months, like cloud version of, sorry. Let me put it another way. So think of it as old closed source models and be patient with that. And you can grow up with the open source model together.

Grace Shao (1:01:54)

That’s very interesting. Thank you so much for your time, Tiezhen Wang.

Tiezhen Wang (1:01:56)

And thank you, Grace.

In this conversation, we discuss what Nathan saw on the ground: how Chinese AI labs differ from their U.S. counterparts, why open models have become such an important part of China’s AI strategy, and how labs like DeepSeek, Alibaba, ByteDance, Kimi, Z.ai, MiniMax, and others are navigating compute constraints, data access, and commercialization.

We also dig into some of the most debated questions in AI today: Are Chinese labs really 6-9 months behind U.S. frontier labs? How meaningful are distillation accusations? Can domestic chips like Huawei’s make up for restricted access to Nvidia GPUs? And is China’s AI ecosystem actually government-directed, or is the reality more fragmented and commercially driven?

Ultimately, this episode is a more nuanced look at China’s AI ecosystem that looks beyond simplistic narratives about subsidies, copying, or geopolitics, and instead examines the technical, cultural, and economic forces shaping the future of open models.

Check out his two recent articles here:

• Notes from inside China’s AI labs

• How open model ecosystems compound

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, and product managers. For more information on the podcast series, see here.

Share

Chapters

00:00 Insights from the China Trip

11:51 Cultural Differences in AI Research

18:15 The Role of DeepSeek in China’s AI Ecosystem

25:26 Overview of Major Chinese AI Labs

30:56 The Future of Open Source in AI

37:50 Market Dynamics and Consolidation in AI

42:28 Distillation and Model Convergence Controversies

51:58 The Gap in AI Performance: US vs China

61:09 Monetization Strategies in AI: A Comparative Analysis

62:32 Government Influence and Misconceptions in AI

Transcript (AI-generated for reference only)

Grace Shao (00:00)

Nathan, thank you so much for joining us today. Yeah, really, really excited to finally hear your thoughts on your big China trip, on what’s happening between the Chinese AI labs and the U.S. AI labs, what you think the potential compute constraints might mean for these labs and their performance in the future, and obviously the open-source ecosystem. So before we get into all of that, could you...

Nathan Lambert (00:02)

Yeah, thanks for having me.

Grace Shao (00:23)

Briefly tell us about how you ended up actually working on post-training and open language models. Just a bit about yourself.

Nathan Lambert (00:29)

Yeah. So I actually started my PhD at Berkeley in 2017, not working on AI things. I was an electrical engineer by training in undergrad, which is funny looking back, because that’s the same year that the Transformer paper came out. And I was like, I think I should do this AI thing, and tried to get the famous advisors to mentor me. And they’re like, we can’t take you. So I had my PhD as this wandering path to become an AI researcher. And then I ended up at Hugging Face after that, which was, realistically, the only industry research job that I had, but also a very hot startup and very fun to learn kind of at the intersection of these tools that people use a lot for AI and research, which is what I was doing.

And then when ChatGPT hit, the kind of RLHF thing blew up as the hot word on the technical side of things. My PhD had ended up being in reinforcement learning, which is just the first half of reinforcement learning from human feedback. So it was kind of a natural pivot to be like, well, I might just do that. And Hugging Face was a good place for doing that, because the whole company is kind of all for that, which is like: figure out how to support the community on the hot thing and build platforms there. So they were very happy about that. And I helped build a team at Hugging Face.

And then I was kind of burnt out on the remote-work time-zone thing and found out that the Allen Institute was doing such similar stuff. And I was like, wow, I have people that could be in-person friends and do similar things. I was like, quality of life — I need to do this. And a few years later, I ended up building a bunch of models. And I think being at a nonprofit opened me to this ecosystem vacuum of information, where there aren’t many people who can talk about what they’re doing. So then, with some luck and committing to write every week, I just feel like my influence filled the vacuum of nobody saying reasonable things.

And it is this nice synergy between what I write about and what I work on in my day job, and it just kind of got bigger and bigger in a very fun way. I think that, generally, at the highest level, I’m motivated by wanting AI to go well on this trajectory. And I worry about a lot of near-term things, whether it’s social unrest in the U.S. and just kind of the massive hatred for AI — I think is a very big near-term problem — and then, medium term, concentration of power, because I think AI will be super powerful in ways that people don’t expect. So generally, open models are a nice way to curb both of them by being a bit more transparent to people, and it naturally is a hedge against concentration of power. There have been different reasons throughout that, but that’s kind of a recurring theme in my life in the last few years.

Grace Shao (02:50)

Definitely. I love your work because I think you help non-technical people like myself really understand what’s behind what’s happening in these labs a lot better. And then I actually just spoke to your former colleague, Tiejin Wang, and he was with APAC Hugging Face just last week. He was saying the same thing. Open source, in many ways, is kind of the best way to go forward as we know that this technology will not stop evolving, but it’s the best way to kind of put up guardrails and checks and balances for the monopolies.

Okay, I don’t want to take up too much time on that side of things today because our focus really is about your China trip. Before we get into the weeds of all that, I want to hear about the trip itself. Most people who are writing about Chinese AI are getting their information secondhand. You really went there, you spent time with the researchers, you met with people who are building the models. Tell us about what you meant when you said you came back with great humility, right? Your eyes are a bit more open, whether it’s the good or the bad. Tell us about your trip.

Nathan Lambert (03:50)

I feel like I kind of went in — I mean, I had this horrible English phrase in my writing, which was like, “I knew I knew nothing about China,” which kind of tried to indicate that I knew going into the trip that I knew nothing. And it was still the fact in my current writing. This is a horribly written sentence that I had in there. And I only talk about it because somebody called me out on it. It’s like, what is this? And it’s like, leaving, which is knowing that it’s such a big country, there are just such vast amounts of talent working on these problems, and how unpredictable it is as a human to model people with very different worldviews and upbringings and training systems. Realistically, the way that people are trained in China is very different.

And I just think that even being there, you can’t fully grasp: what are the pockets of three to six researchers doing that is actually a bit different than in the West, even if they’re working on the same goal? I think you could get down to that level of granularity and a sociological study and actually see differences in what they’re working on, and that’ll always change the output. I didn’t get to that level of granularity, but it’s just to start having real experiences and understanding how people explain how they work on these problems.

And for me, realistically, a lot of it is coalition building, which is just like: I want there to not be vitriol at the level of the technical companies doing things in international bodies. So just meeting all the labs on both sides is really nice, because you need to do that for them to talk to you about more sensitive issues in the future. I got some criticism on the piece, which is like, this is how you shouldn’t visit China. And it’s like, well, what are you going to do if you’re going on an official visit to a bunch of companies? How do you expect to get in the door without being nice? You have to start somewhere, and I think it’s important to be respectful.

Grace Shao (05:31)

I think the piece was, frankly — I don’t think the criticism was fair, to be honest, because I think you were really transparent with the fact that you’re not a China person, right? It’s not like you’re going there and exoticizing everything. And if anything, a lot of people, even with China backgrounds, like to use certain dragons and tigers to describe things. I feel like you actually were really humble going and being like, I’m just a technical dude meeting with these labs, talking about their technical research, right? And then because you were physically there, you had observations of the culture and the people. So yeah, I actually thought your piece was quite good. And yeah, sorry.

Nathan Lambert (06:05)

I agree. I was willing to let that sail past, but I think it’s important for people who listen to realize how actively these companies are trying to court Western audiences, which is why we could get in the door. I mean, we had some prominent people on this trip, but that’s why we got all of them in the days that we wanted them, except for DeepSeek. So essentially some, like Catherine Rintel, who works with me at Interconnects, and some other creators...

Grace Shao (06:23)

How did you get everyone? Yeah, how did you get everyone?

Nathan Lambert (06:29)

He used to live in China and has connections in China. So he kind of orchestrated the mix of his connections and leveraging my connections to labs. We had some bigger names on the trip as well. Just stringing all of these together to get all the various labs in place is a few months of networking to make sure the trip lines up with people with established networks and contacts with the various labs. But these people want to look good to Western audiences, so they’re only going to say yes to the right researchers.

And the researchers know that there are two to four comms/ops people in the room, hanging out, making sure that it goes well. Especially the bigger the company, the more comms people. You go to Alibaba and there are three to five various people, from the head of comms to some special offices. You’re not going to get these people in the office, or at all, without accepting the cost of these types of handlers. It’s the same thing in the U.S. You’re not going to just plop a senior executive into a chair.

So it’s also good because now I have the WeChats of a bunch of researchers from China that I could just text about things. It’s like, hey, congrats on the new model release. It’s like Lay Lee works at Xiaomi, Xiaomi MiMo. It’s like, talk to this guy for an hour at a mall — I don’t remember the name of the tea store — but it’s like...

Grace Shao (07:27)

No, of course. No, of course.

Nathan Lambert (07:49)

Now we have these relationships, which is very useful, and that helps information spread across the ecosystem to these trusted parties, which doesn’t really exist. There are not that many, I think. And the opposite direction of the trip is very hard because Chinese researchers can’t really enter the U.S.; the visa purgatory is too complicated. A lot of us on the trip were either Canadian or entered on a transit-without-visa entry, which makes it very easy for American technical talent to go to China right now, which is why I think there are so many trips. I think there’ll be more of them.

We’ve got a lot of inbound from VCs and open-source labs in the U.S. that want to establish collaborations with these various labs because they’re the best open-weight models, and they want to build a stack for companies in the U.S. building open-weight models. So I think there are going to be more prominent, but not gigantic, U.S. startups going to try to build these relationships, which I think is a really interesting technological development because we’ve never seen this type of professional work trip in China from U.S. tech companies. Most tech companies have a “bring a device to China, it auto-bricks itself, and you have to hand it into IT.” So to actually proactively send people in a professional capacity is a really big change. There are a lot of angles you could take this, and I think it’s cool to see how it unfolds. This isn’t even really about the trip. This is the follow-on that we’re hearing from people that are like, hey, how’d you do this? We want to do this trip.

Grace Shao (09:06)

Yeah, definitely. Actually, from my end, I hear about VCs or investors always being quite active going to China because previously American funds were very, very active during the internet era. People were kind of always trying to find a way to either get into these good deals or potentially keep their pulse on it. But I think it’s really, really positive for the whole AI ecosystem to have this kind of fair, transparent exchange in some capacity. But to your point, there’s no way that star researchers can come out and talk to you off the record without any compliance, because that doesn’t happen in the U.S. either. That’s just companies protecting themselves.

I just think your trip was quite meaningful, and I want to bring it back to your observations. You talked a lot about the cultural aspects of it. You talked about how you felt like in China there was less of this star-researcher celebrity status around people. People were more humble, or there was more humility. It was very focused on execution. You argue that Chinese labs are particularly well suited to the current LM-building game because they’re very focused on meticulous stack-level work. And there’s less ego sometimes to work on the dirty work, or the non-sexy work. So kind of unpack that for us. Why do you think that is? You kind of touched on it — you said they were brought up differently, they were taught differently — but what’s so different?

Nathan Lambert (10:27)

So essentially, an interesting part that synergizes on this trip is that we stopped by some academic institutions. I think it was like AIR and Tsinghua and stuff. And you hear all of these academic leaders talk about how they’re pushing hard to try to change it. So yes, they know China is producing more papers than anyone else, but they still think that it’s not as transformative of research. And they think that they’re trying to cultivate the academic domestic ecosystem to change just the type of work it works on, and the distribution, and take more risk.

And then you would talk to some industry leaders off the record behind closed doors, and you would hear things like, it’s never going to change because the education system is so structured. There are so many layers of the funnel that reward things like memorization and stuff that they’re just like, this research culture is not going to emerge. And then the follow-on with the AI labs is that these labs are doing fast-following. They kind of have a proof of concept, and they know what it needs to look like. Therefore, in that domain, you’re not trying to invent the new paradigm. You’re not trying to make the model that is o1 or o3, or the first model to work in Claude Code. You’re like, I see it, and I’m going to try to do that and make it the best thing. And I’m going to try to make it cheaper and just maximize that goal.

A lot of companies don’t need to invent the new paradigm. OpenAI has done this so many times. That’s their bread and butter: never doubt OpenAI’s ability to release a blog post and a plot that changes how people think about AI. I still think it’s going to happen a few times in this massive boom over the next four years. OpenAI just kind of has that sense of what is the thing that you can push on a bit earlier and just transform things. But I don’t expect — and other people wouldn’t expect — the Chinese companies to do that as much, because it’s just such a culture of, I guess, building. I don’t know how to describe the positive version of this. Maybe it’s slightly more practical-minded, in terms of: it’s your job to build this thing.

A lot of the researchers, maybe because they knew their managers — some of them had managers in the room — see their role in the company as being to make the models excellent. And especially for students, I work with students and that’s what they say. I work at the Allen Institute and we have students that will co-lead our language models. It’s not that surprising, because if you do an industry research job in the U.S., a lot of mentors will tell you that you’re kind of free of the burden of bureaucracy and politics. So the naivety of students, and the simplifying, is actually so good at just getting a lot of technical work done.

There’s also the life-stage side. If you’re younger, you don’t have as much family, and you normally haven’t built up as many habits and other things you do with your life. Language models are so complex, and the amount of context that you need to absorb to understand what the bottleneck is — there’s so much information, and you have to be able to pick what the bottleneck is and break it. If you just don’t have the mental space to absorb all the context, you kind of end up doing things that are cute but don’t make breakthroughs on the model.

So that’s kind of a difference that I’ve seen in people who were both very successful academically before language models. Some of them are able to pivot to this practical mind, which is: what is the state of the system? How do I improve it? And then some try to make kind of these abstract frames of what’s happening and approach it like an academic, and it normally doesn’t improve the model as much. So I just kind of see, if the academic system is a bit more practical-minded, a bit more structured, and the work you’re doing is structured in the language model — make this kernel implementation faster, make this idea work — then maybe it can be...

I think it’s an oversimplification. I push on that a bit in the piece just to really contrast what you could think a U.S. lab would look like. And I have a few anecdotes. I’ve heard a U.S. lab paying off a researcher to be quiet about their thing not being in the model. All of these one-off things are more storytelling devices than anything, because most one-off things don’t matter at all. But also Llama 4 imploded, and that was because it was described as a Game-of-Thrones political-style environment, with all the VPs vying for influence and showing that their thing made the benchmarks go up. It kind of fell. Many, many people will tell you that. And we’ve had the Qwen turnover, but it doesn’t seem like it was quite the same type of thing as Llama 4 or xAI. xAI barely exists now. There have been some dramatic things in the U.S. with how these companies have kind of come and gone out of the fold.

Grace Shao (14:55)

Yeah, I kind of agree with you, but also I would push back on that. I think there’s obviously a more rigid and competitive academic system, which by default in East Asia results in a culture of students following the bureaucracy and authority a bit more. So I agree with you in the sense that they’re very pragmatic. They focus on the task that is given to them. However, I wonder if things will change with how AI will disrupt education. That’s number one. But also, a lot of the young researchers that you’re working with today seem quite different. At least a lot of the entrepreneurs I meet today are born in the ‘80s and ‘90s, some even younger and born in the 2000s. And I think there’s a kind of aura or confidence coming from them. If anything, you want to say they’re a bit more individualistic-minded. You went to Shanghai, right? They are dressed very, very uniquely. They have these outrageous outfits on the streets. People are seeking individual ways to showcase their personality. So I wonder if that will shift.

But for sure, for the academic institutions like the Tsinghua and the Beida of the world, they are still very old-school. But I would say that is the same maybe in some academic institutions in the West still. Okay, I think on this topic we can go off on a tangent on academics, but let’s go back to China’s ecosystem.

When DeepSeek V4 came out, we talked about it offline, the two of us, quickly about a piece I wrote saying how DeepSeek is starting to look a bit more like a base layer for China. And if anything, some of the labs kind of admitted to that. They’re like, we have very limited resources. And to your point earlier...

Nathan Lambert (16:11)

Yeah, you could take that in so many tangents.

Grace Shao (16:34)

Limited people — these labs are tiny. They’re run by 100 to 200 people max. Limited capital, obviously limited compute. They have constraints all around. And in that sense, in a way, the ecosystem’s looking less like a zero-sum game and more like different players optimizing their own strengths. So correct me if I’m wrong, but DeepSeek is providing a base layer where a lot of labs will quickly follow and basically adopt a lot of their engineering breakthroughs. And then Zhipu, Z.ai, will focus on the coding; MiniMax focusing on the multimodality, et cetera. There are a lot of these different players. ByteDance, obviously, very, very focused on their video models. And Qwen, like you mentioned, had the whole open-source saga break apart with Lin Junyang leaving. But in general, they’re still kind of the leader in hyperscalers on that front. So everyone’s doing their own thing almost, instead of really...

Nathan Lambert (17:27)

I agree with the people specializing, which I think is normal business evolution. You figure out a bit where you’re good at. And there’s so much opportunity that they are like, okay, I’ll follow this because they see that they’re good at it. I just am more skeptical of DeepSeek as a base because I have no idea what DeepSeek is doing. And some of the labs when we were there, because DeepSeek V4 had just come out, were like, yeah, we look at the things they’re doing, but they seem more intricate than needed. And if you read the paper, there’s just so much going on in this model. As a researcher, I’m like, some of it seems a little fake or a little dependent on their setup and not necessarily going to work in every model.

Grace Shao (18:04)

What does that mean? Break it down for me.

Nathan Lambert (18:18)

Essentially, I will say that building an LLM is dependent on where you have your GPUs, your pre-training dataset, your intended deployment setup, and stuff like this. So you make decisions based on your constraints, and you build the model. DeepSeek has these constraints and they end up with their model, but Moonshot and Zhipu have different constraints, maybe more flexibility, and they ended up building a different model. They will test the DeepSeek innovations. So they’ll say things like, X innovation doesn’t improve our model. These two organizations are on different development paths that have core similarities, like these large mixture-of-experts models and the general methods are similar, but a lot of the parts end up being a bit different.

That’s why I’m like, I don’t know exactly. If DeepSeek was a base, you would see the Chinese labs just do post-training. We just take the base model that’s out there and we adapt it to our domain of specialty. And we have users that do that, which is something that I think about a lot. I’m thinking about starting a post-training lab and how to format post-training research better. So I think about this a lot. I think about what a shared base actually would be. They go through — some of these labs put an extreme cost on creating their base model. And if they didn’t need to do that, they wouldn’t.

One of the labs told us how long their pre-training run was, and my jaw dropped. I was like, that’s way too long. Any U.S. advisor would be like, you’re taking way too much risk on this pre-training run. If they didn’t land that pre-training run from one of these past big MoEs at a Chinese lab, I don’t know if the company’s dead, but that’s a huge amount of time. Most U.S. companies now know that you don’t want your big pre-training run to be more than a few months because it’s just so much risk and time to put all your eggs in that basket.

That’s a sign that, in that case, they don’t have as big of a peak-size cluster. Essentially, pre-training time can come down a lot when you have a bigger overall cluster; you can just get more throughput on it. But if your biggest cluster is smaller, it’s harder to get a certain amount of throughput, so you use that one for longer. That’s a compute constraint. To loop it back, I think the specialization is real, but I’m more like, I have no idea what DeepSeek is doing. I know they’re raising money now. I don’t know what the plan is there. They seem the most without a specialty in the Chinese ecosystem.

Grace Shao (19:59)

Dependency on. Yeah.

Mm-hmm.

No one knows, though. No one knows. They’re secretive.

But that’s my point, right? I feel like they’ve been kind of nationalized, whether willingly or not, because they’re taking the Chinese government’s money. They’ve kind of gone secretive. And it’s not like there’s a secret that they prefer Chinese-educated researchers. They’re keeping a very domestic stack, from talent to capital to the whole stack. So to me, it seems like they’re being Huawei’d, in some ways, because they did well and they got their name globally, and then by default they’re becoming the next Huawei, willingly or not.

Nathan Lambert (21:01)

I don’t think nationalization makes you a base for the other companies, at least not at this stage. There could be something, but it’s hard to force.

Grace Shao (21:06)

But then you have some incentive, right? But then it is some incentive. You’re like, well, if you can propel one of the teams and propel the whole industry as a whole, it could be in your KPI or some kind of unspoken expectation.

Nathan Lambert (21:17)

The coordination problem is so hard. Essentially, both in the U.S. and China, even the open labs, what they do is they fork open-source code and match it to their internals, and every company does this. Therefore, all the improvements that could potentially be going to the open code and forming this base that is far more efficient — they’re not completing the feedback loop. I think China could be closer to it. If people really lean into DeepSeek as a standard architecture and DeepSeek shared their training code and all the specifics and how to do this, from a Chinese economic perspective, that would be a huge win because you’re just saving compute. But I think it’s too decentralized and too competitive to have that happen. It wouldn’t happen in the U.S. either.

Grace Shao (22:04)

It’s so cutthroat. Yeah.

Nathan Lambert (22:08)

Even though I think for open models to be closer to the frontier, it would be better. I talk about open models in the U.S. needing a consortium. But there’s definitely enough money to make a consortium in the U.S.; then you fail because the model won’t be good because you’re feeding too many asks into the model. That’s the only way to create a shared base.

Grace Shao (22:25)

Interesting. So it’s not really just commercial. Yeah. It’s not the commercial reason.

Okay. So if you had to give a high-level commentary on each of the major labs, what would it be? If you look at ByteDance, Alibaba, Tencent Hunyuan, if they’re relevant, DeepSeek, Moonshot, Zhipu, MiniMax, Meituan, Xiaomi now being part of the ecosystem too.

Nathan Lambert (22:46)

You might have to prompt it or say more, but I could just kind of ramble through them, which is kind of fun. Alibaba: cloud-focused, understands that open models can enable more usage of platform. So I would say Alibaba is very, very cloud-focused. ByteDance: mostly characterized by everybody else being intimidated by them, and very user-focused, including multimodal. Kimi: vibes of the office were great. It would be one of the best startup vibes that you would visit among U.S. or China. Zhipu: very AGI-pilled, surprisingly cautiously excited about being entity-listed, even though they have no idea why they are, because they’re like, it stamps them as a big deal. And then there’s some...

Grace Shao (23:27)

I think they previously worked with SOEs. That’s the main reason. Or they still do, but that was one of their main sources of income. And unfortunately, because a lot of these labs spun out of Tsinghua, and Tsinghua is, for people’s context, in Beijing. It’s really close to the government, obviously. But the thing is, when it’s close to the government, it could mean there are three layers of agency underneath the actual government apparatus. But then people like to link it to the fact that it’s taking government money, so therefore they are suspicious. It’s very unfortunate, I think. A lot of companies get thrown into that category. Even companies like Lenovo and a few other Chinese companies have previously been called out by U.S. senators saying, they’re taking Chinese government money, but really it’s that their scientists or their research labs spun out of a certain government-affiliated or government-funded academic institution. That’s what it is. Anyway, yes, go on.

Nathan Lambert (24:23)

Yeah. Some more would be: Xiaomi — surprisingly great research vibes for a new team at a random company. They seem to be crushing it.

Grace Shao (24:31)

What do you think of Luo Fuli? The star researcher.

Nathan Lambert (24:31)

I didn’t get to meet her. I think she’s as close as they have to a star researcher right now. There’s the tier of star CEO, which there are obviously others — Dario and Sam, the analogies are there — but the star researchers, like the Sholtos of the world in the U.S., obviously you can come up with many more. She’s the closest you have to this. I need to watch more interviews. We’ll see. But she wasn’t in our meeting.

But they just seem to be doing the right thing. They’re making general models. They don’t really have specialization yet. Florian, the person who helps me write about open models on Interconnects, and I took a detour to go see Meituan because we’re like, why is Meituan building these models? And they’re very practical about it. It was a less glamorous visit at a normal tech office. It wasn’t an official visit for them. They were like, yeah, we’re a major online platform. We obviously are going to use LLMs everywhere once we need to build our own LLM and specialize it to our products, which, surprise, is very practical-minded. I’m guessing there are many more companies in China like this.

Grace Shao (25:39)

That’s what Tencent’s saying too. It’s because they want to serve their existing consumers and optimize their LLMs for their own distribution and their own basic interface or activity loop.

Nathan Lambert (25:52)

Yeah. After I left, some people in the group went to Xiaohongshu, like RedNote, and they’re there. They’ve released some language models that are multimodal. They’re like multimodal data-processing things. So a lot of them are not that surprising. The startups just have different cultures. I have met some MiniMax people before, so I left the trip early before MiniMax on this one. But MiniMax was quirky. They have a ton of women in their company, which was very fun. And they have products. They’re maybe slightly more product-focused, but I feel like the quirkiness of the company kind of matches maybe Western confusion over what their products are doing and what they’re trying to do. But it kind of matches their language models that are a bit more efficient.

Grace Shao (26:35)

Well, they came out with a lot of very consumer-focused applications, right? They had Hailuo and Talkie, all these character companion-bot products before.

Nathan Lambert (26:45)

Yeah. And then the last one I went to was Ant Ling, which is also very corporate, but in a less intense way, because I think they see it as serving their own products, whereas Alibaba Cloud is like, this is the gold mine we have to win. It’s a much bigger deal for them than Ant Group. But a lot of these things, when you list them — I don’t know, eight to 10 companies — they’re all pretty reasonable with respect to the age of the company and what the company does best. There’s not as much confusion.

Grace Shao (27:14)

Yeah. And Ant is low-key best at medical chatbots right now, which I guess makes sense because everyone has access to Alipay. And then for seniors, apart from WeChat, it might be the only application they’re using on a regular basis. So it became the default medical consultation app, which is really random, but it’s their niche now. Yeah, I think you’re pretty spot-on. It’s pretty cool that you got those takeaways, even just meeting with them for a couple hours.

Nathan Lambert (27:41)

I have been reading about them for so long, so a lot of these priors are easy to confirm when they kind of fit with things you have seen. The Chinese showroom culture is so interesting, and also one of the most surprising things to have at software companies. It’s so funny. They’re definitely appealing to Western audiences. Z.ai had poorly translated merch. What was it? Something so — it would be borderline inappropriate translation in the U.S. It was like “ship big, go hard,” or something. Just some really weird translations. And they have live API statistics in their showroom. So Z.ai was like, we’re serving 5.5 trillion tokens a day. All the U.S. companies are so closely watched for when they announce token statistics.

I know at least one of these numbers is wrong. It’s something like Fireworks does either 30 or 300 trillion tokens a day — or I meant Together for that one — and then one of Fireworks or Together, and the other one, are like 100 trillion tokens a day. Don’t take these as sourced; go look them up. There were some public announcements recently, but those were the first updates that anyone has on major infra companies in the U.S. Inference is a huge market. You don’t hear anything from Fireworks because they’re just struggling to demand and they’re making bank, because inference is a much better thing to sell than bare metal.

Essentially, inference is selling the software implementation to serve tokens more efficiently, and you can just get more margin when you improve the stack for a fixed model. So a model comes out and you host it, and then you can make your stack more and more efficient on that model. You just get more margin and hopefully growing usage. That’s way different than GPUs, where the best case is that you lock in a huge commitment for a long term.

Just being able to walk into an office and learn about their API is interesting because they also had geographic distribution, which was like: China was, I don’t know, two-thirds; U.S.A., 20%; and then the last percent was Singapore, Korea, Japan on the Z.ai API. So that’s cool. This is shit that I always want to know about the companies, and I have no idea. One of the things I always want to know is: how are open models being used outside of the U.S. and China, and has this decades-long process of technological diffusion started to kick in in a way that any company can measure? I don’t think anyone has good data on it yet, but I think it’s obvious that at some point, open models that are cheap to run are going to have some interesting playbook across the globe for the long tail of countries. Maybe I’ll just walk into the front door of a Chinese open-weight company and get my answer.

Grace Shao (30:30)

But actually, I think the culture of these labs — a lot of them, because they’re run by really young, passionate people — you would feel like they’re a lot less commercialized or less corporate, or at least less sleek. They’re not sophisticated with, you can say, the capital-market side of things, but you can also say that they’re just really naive and open-minded and passionate about the product they’re working on, with less of a corporate guardrail built around them.

Nathan Lambert (30:56)

Yeah.

Grace Shao (30:57)

Okay, I want to talk about...

Nathan Lambert (30:57)

Yeah, go ahead. It’s like one of the people at Z.ai who’s known on X — I don’t know, 9,000 followers — it’s like Lu. She came up and was like, hi, I’m a student, I’m 20. I’m Lu from X. And I was like, that’s hilarious. There was a lot of shit like that. It was like, oh, okay. I don’t want to call her a kid, but it’s like...

Grace Shao (31:06)

Yeah, yeah, yeah. And I think the one that runs Moonshot’s developer ecosystem or something is literally a girl fresh out of school, right? And she just posts hilarious memes all day long. There’s no filter on her social media. It’s funny.

Okay, we go on these tangents, Nathan. We need to come back on track. Open source, open weight. Why? Why do you think Chinese labs are adopting it or embracing it, however you want to put it, especially after visiting them? Is it because they simply have to, because of what we talked about — they are leaning on each other because of all the constraints they have? Or do you think the philosophical drive is actually bigger in that ecosystem? Or is this a bigger strategic thinking for diffusion in the long run?

Nathan Lambert (31:54)

I actually don’t feel like it’s that special ideologically. I think it’s easy to say the ideological line when you are doing it. Now you can look at Zuckerberg: he said the ideological line when he was doing it, and then he stopped. I think it’s mostly just that, for one, distributing within the U.S. ecosystem, especially to enterprises, is the highest-value market, and they can’t sign many enterprise deals. And the closest best thing is things like Cursor adopting Kimi’s model. Even if Kimi doesn’t get paid for that, they’re happy. That’s the biggest sign of credibility for them, and they can figure it out in selling tokens or whatever in the future.

Practically speaking, one, the only way to influence the U.S. market is by releasing these models. And two, it seems like they don’t feel like they’re losing as much if they release and share things. If the model was closed, they just think they would get less influence, they would be seen less, fewer people would use the model, their actual paid offerings would be adopted less. It just seems almost overwhelmingly obvious, because there are all these benefits and not as obvious of a drawback. There will always be better models, and just keep going. But I think every scientist loves...

Grace Shao (33:08)

Then why are so many U.S. labs against it, or not willing to?

Nathan Lambert (33:12)

Because they can make as much money without it. Anthropic and OpenAI make more money by not releasing them. They can just make so much money, so why bother thinking about an open model that doesn’t make money? There are different scales of influence. Same with Google. Google’s making so much money. I think Meta will make a lot of money by having good AI models in their products, if they get their act together. Even Google could release more models. They have so many surfaces other than Gemini that need AI to be commoditized and used, like the cloud and all of this. Meta could release the models. It’s just not worth the effort for some of them. They’re like, we need to do this high revenue target; it’s too much of a pain to go through legal and make it ready to release. Why bother?

I don’t know, maybe it’s a little bit of a cynical take, but I think Microsoft and Meta could release their best models openly because they benefit if it’s a commodity layer. But I don’t expect them to, because it’s just kind of like the benefits of focus are so high, and they just kind of see it as something they don’t have to do.

Grace Shao (33:56)

And it’ll be good for them. But then eventually, we will see some consolidation in the market as well, assuming — because you can’t really have 10 labs in each dominant country right now all exist.

Nathan Lambert (34:26)

I do expect consolidation. I think this is potentially a subtle cultural point, which is that the U.S. labs are more likely to buy into “we’re special, we need to go fast, keep it closed,” and the Chinese labs are not. There could be something there. That’s also who the decisions funnel up to. I don’t know. I talked to the Alibaba people that make these decisions. I can’t say all the things that they say about them. Some of these were two-on-one and off the record, so I can’t say all these things. But at all the other labs, there is a person that makes the call, I’m guessing. I think those are senior leadership that we’re not talking to. So it’s kind of hard to know exactly what they really think.

I definitely expect consolidation. My thing is that I expected it in China faster because the capital markets aren’t as strong as in the U.S., but I don’t have a model for that. I think you can model it, which is: what do you think the revenue growth would be? What do they need to do to raise to keep training bigger models? What is the compute cost? Then you look at the potential raises and think about which country would not be able to do that race first. But also, it’s this wild thing with OpenAI raising $120 billion. Are you kidding me? What is that?

Grace Shao (35:47)

Yeah, the valuations in the U.S. are not really understandable by anyone else right now. I think in China — so on your point on that, I’ve been writing about this and I think it would make sense for Tencent just to buy out one of the labs. They have the money, they need the capabilities, and frankly, they’ve really been struggling to compete with their LLMs, with all the labs talked about just now. So my...

Nathan Lambert (36:05)

Their licenses are so bad. They release all these models that have horrible licenses. They’re not that good, and the licenses are just horrible.

Grace Shao (36:13)

So I feel like it financially makes sense for a company like that to optimize and just buy out a lab. Then the labs can also lean on their distribution, because at the end of the day, how are they going to win consumer mindshare or distribution in China right now when it’s really just dominated by Alibaba, ByteDance, and Tencent? That’s my spiel. But when I spoke to some of the researchers...

Nathan Lambert (36:33)

I think big companies have a lot of inertia, and the senior leadership has the call, and they can have inertia. I still think Apple just ends up buying some lab for $25 to $50 billion. It’s not the worst thing. Just golden-handcuff the researchers. Some will still quit.

Grace Shao (36:43)

Yeah. But I think right now they don’t want to. The labs still have a dream. Some of the researchers still have a dream. So when I spoke to a lot of them, they’re like, no, we don’t want to do that. We want to commit to our own frontier research. If I wanted to join one of the big tech companies, I could have. So why would I want to sell? That’s what the researchers think. But to your point, we don’t know what actually the one person or two people at the very top think, especially if they continue to have hurdles with compute access and capital access, which brings me to the question.

Nathan Lambert (37:14)

It also depends on your view of inference. You can ask your next question. I don’t need to cut you off. It depends on your view of inference. If these agents are just so much inference, I do think it’s going to be an oligopoly-style market, not a monopoly-style market. And what’s the difference financially between two and four or five big companies with great models? Is that actually not sustainable if there’s so much demand? There are a lot of cases where we have two or three, like the cloud, but what’s stopping that from being four?

Grace Shao (37:40)

I think they will be the infrastructure providers. Yeah, yeah. And they would kind of lean into each of their existing ecosystems or distribution, whatever you want to call it, and serve certain specific models for specific uses. So enterprises can choose what matches their needs the best as well.

I do want to bring the conversation to a more contentious topic, which is on distillation and model convergence. You raise the question of whether Chinese models are structurally different. Often we are hearing claims saying a lot of these labs are about three to six months or six to nine months behind U.S. labs. There’s obviously a lot of noise or allegations and accusations from certain U.S. labs saying Chinese labs are distilling them. How do you actually see that accusation or that kind of dynamic?

Nathan Lambert (38:36)

The biggest unknown that I don’t have an answer to, which actually has a lot of sway, is how much of the Chinese companies are actively trying to hack APIs versus just showing up as a customer and paying. If you’re trying to hack the APIs, normally you get reasoning traces out so that you can create a reasoning foundation that would be similar to the model that you’re trying to do this from. That’s very different than the API standard form, which is just the output of the model, which is a less direct process for learning from.

I don’t know the magnitudes. If it’s more just like, I walk up to an Anthropic API and I use it as intended, but I’m making a competitive model, I’m not very sympathetic to Anthropic. They could ban it if they want to. And I think the impacts are kind of a standard practice. You can do it with many different models and so on. The evidence Anthropic provided is not large enough scale where I’m like, this is industry IP theft at mass scale going on 24/7/365. So there’s definitely some gray area to what is actually happening in distillation.

That’s why, on the policy side, I try to push people to not call all of it the same thing. Essentially, using any API endpoint to make synthetic data to train your model is some form of distillation, but it’s very different if you’re trying to break this model so that it gives us a different behavior that is hyper-useful for training and not get caught. Those are pretty different actions, and they’re all looped into this common phrase of “distillation” right now. That’s my biggest problem, which is that academic researchers and small companies use distillation extensively as the core of their business and the core of research methods. So if the U.S. government nukes that as a thing that could be done in the AI ecosystem, it’s mostly bad for small players, bad for U.S.-China tensions, and bad for academics. That’s my primary concern.

And then trying to get the labs to actually say more. There’s a distillation side and then performance is the other side, which on benchmarks, it does seem like the Chinese labs tend to be six to nine months behind. When it comes to general use, I’ve always found the closed models to be better in ways that are hard to measure. So I go very back and forth on whether the closed models are better. I think we will especially see Anthropic and OpenAI pull ahead on knowledge-work tasks like legal, healthcare, financial services, because I just don’t see the Chinese labs paying for that data. All that data is going to be people that charge hundreds of dollars an hour to annotate and create these environments. So it’s a whole new capital build-out that goes on there right now. It’s going to be billions of dollars if you’re going to buy a billion dollars of data and a billion dollars of compute and a billion dollars of talent to train your model.

Grace Shao (41:30)

They don’t have the money.

Nathan Lambert (41:30)

I don’t think they have that. Mercor has some of these evals, and I think there is a bigger gap there. So it’s very interesting. Florian, the guy that helps me, and I disagree on it. It’s this fine line between, yes, the evals — coding and lots of these things, and even random evals that surely the Chinese labs aren’t training on — the open models really are genuinely crazy impressive scores. So I think there’s also a tester’s bias, where I don’t use the open models as much. Maybe it’s hard to ground in my head what I was doing with AI six to nine months ago. I wasn’t even using Claude Code as extensively.

I guess the question is, at the end of this year, can I use an open model in something like Claude Code and feel like it works at all? That’s the test on the performance gap, starting in June, June to August, and whether or not that hits. I don’t think the open models have hit that yet. I think it would be way more of a narrative if all the companies spending billions of dollars on Claude are like, oh, we can spend 1% and just use DeepSeek. These CIOs and all the big companies — some companies spend more on tokens for their employees than on headcount. These are normally startups. But they would happily reduce that token cost to 1% expenditure if it really was that similar, because then you could just use 10x the tokens. I don’t expect that to happen. And I expect things like the latest Claude and GPT-5.5. I expect more of these things through the year, and we’ll see if I end up being right. Both are right at the middle of us, as a world, getting more clarity on them. They’re like 18-month-long stories unfolding, and I feel like we’re just in the middle of performance gap and distillation and learning more.

Grace Shao (43:25)

Yeah, it’s interesting. You mentioned — it helped me recall a conversation I had with other people as well. The point on distillation is that I just had a conversation with your former colleague at Hugging Face, who leads APAC, called Tiejin Wang. He was just saying, look, the distillation accusations don’t really make sense because we’re all distilling off of each other as we speak. I’m learning from you; you learn from me. We’re distilling. It’s so vague of a terminology to just use that to accuse all these various behaviors.

So to your point, I think people in the technical world who understand what’s happening actually want more clarity on what is the gray area, what is actually black and white, and what is not appropriate or unethical. That needs, I think, the industry to come together to really put guardrails and rules around.

Now, number two on the compute side and the data side. Something anecdotally will be interesting to you is that when I spoke to one of the lab researchers in Beijing, I think in February around Chinese New Year, they were saying, look, they want to get better data, but they can’t because usually a lot of American labs would pay tens of millions, if not even more, like a hundred million dollars, for a set of very obscure or niche datasets, but they would have an exclusivity contract. What the Chinese labs will do is that they will literally wait out the exclusivity contract and then, say two or three months later, pay for it at one-tenth or one-twentieth of the price for that same dataset. So then once they start post-training on that dataset, that’s where the three to six months or six to nine months come in as well. Yeah. On that note, I want to...

Nathan Lambert (45:00)

Yeah. I think the data industry in the U.S. has two things. One, the lab asks the data vendor, we need this specific type of data. And the data vendor is a network that connects the people to the lab. The other thing is the data vendors know evals that are important, so they try to create good data for hill-climbing on specific evals. That data could be sold to multiple people, but is less expensive because they make it once and expect to eat margin or take margin on it. There could be a pipeline where once OpenAI is at the cutting edge, creates this new thing, they create deep research, then the data industry is like, let’s make things that are a little bit cheaper to sell. So there is time lag in these various things.

But I heard the same thing on the ground, where they have a negative view of the data industry. It’s like, quality is bad, we don’t really have access, we do some in-house. That’s a very big difference from today, which is that you have the data companies in the U.S., which is insane.

Grace Shao (45:53)

Yeah, the American data companies are so mature. It’s its own sophisticated ecosystem.

Before we get into data, I actually want to ask you this question. I think recently a lot of the narrative is now saying, look, Anthropic and OpenAI have kind of proven that pre-training scaling laws continue to hold, especially with the recent models. There’s an obvious compute constraint on the China side that we talked about. And then it will likely be even more amplified with the absence of Blackwells in the coming months.

So as we move forward in this race, per se, if you have to put it in China versus U.S. in that sense, will we see a wider gap between the performance and benchmarks between the Chinese labs and U.S. labs? As in, will we see the gap going to 12 months, 24 months, as Chinese labs are very, very constrained on compute for pre-training breakthroughs?

Nathan Lambert (46:40)

I think it’s more of pre-training as a thing that you could actually finish. How big can you pre-train a model that you can finish and serve? The Chinese labs could train models that look like GPT-4.5, which is this giant model, but you can’t serve it. They end up training a model that is 2.5 trillion parameters and they release it, and no one can use it. They could barely serve it on their API because they don’t have Blackwell NVL72 racks or something — these racks that are definitely what are serving these large MoE models. They just don’t have the quantity of these.

So there’s a difference between models that you can build and models that are actually useful. I think some of the Chinese labs are definitely like, we don’t need to release the gigantic models because nobody is going to use them in open weight. The biggest models end up getting served via API. So there might be some segmentation in that market. But I do think the inference and amount of economic resources that you have to serve your customers is becoming a thing that dictates what models are built. That’s why I think the gap will continue to rise. All signs point to GPT-5.5 being a bigger model, and I don’t expect that to stop.

And then the economics of it is just the basics of: you need a certain volume to have the margin to support the research, because you can’t keep raising these ridiculous rounds forever. I think OpenAI, Anthropic, and Google are the only people with that AI usage volume to keep marching down the scaling laws to another 10x of training compute, which is mind-boggling amounts of investment in a model. That’s why, when the economic markets slow for fundraising, the model gap between these big three will just show a lot more. That’s the distilled way to say my prediction of when things will look different. It’s like these labs can’t fundraise, they go public, they can’t generate revenue more on their paid services, and then it’s just: look at how much training compute can be allocated or can’t be allocated.

Grace Shao (48:41)

Yeah. Basically, we’ll see a bigger gap, I think, in the coming months. Then what can make up for that? Domestic chips, or, like you said, better data. And why is it that sometimes people assume China has a very strong data ecosystem or data products, but actually the data vendor ecosystem is very weak in China?

Nathan Lambert (48:41)

So generally, I think I agree with what you said. I don’t know on the data side, but the way domestic chips could help is that if Huawei chips are fine for inference, and if they have sufficient volume to support the inference economics, which then trickles back into revenue, my read is that they just don’t have the volume of the chips, especially spread out across the amount of companies that they have. Essentially, the total FLOPs of Huawei, all the things produced, and it’s going to all these different places — it’s just not big enough.

It could be something like ByteDance and Alibaba, with offshore data centers, can keep up a lot longer because they have access to Nvidia compute and have for a long time through this kind of offshoring. Maybe that stabilizes the ecosystem, and we’ll see what the AI startup, the younger startups like Kimi and Z.ai, end up doing. No one wants to do this, but if they pool resources, they last an extra year. You get another order of magnitude if they all pool together, but I don’t see them doing it.

Grace Shao (50:00)

But that’s the thing we were just talking about, right? MiniMax and Zhipu, how can they possibly compete with the hyperscalers at this point if you need offshore data centers? And the fact that Zhipu is on the Entity List doesn’t help, right? It’s not going to be easy for them to access these data centers either.

Nathan Lambert (50:12)

Yeah, I think they can’t. I think they won’t. Human nature will make it so they won’t collaborate. They’ll just do something smaller. They’ll just have successful businesses that are different.

Grace Shao (50:22)

They just have smaller ambitions, want a smaller piece of the pie. Yeah.

Okay, so you wrote something like, nothing’s a secret, but everyone wants Nvidia chips. They want it, they don’t know how to get it, they’re fighting over it.

Nathan Lambert (50:34)

Yeah. They’re the only thing that works for training. All the models are trained on Nvidia. I don’t believe the DeepSeek propaganda that it’s trained on Huawei. The only models that are trained on Huawei are tiny. Inference on Huawei works. Every lab is like, inference on Huawei works. The labs that don’t have meaningful inference are like, we are told to get Huawei, so we buy them, but we don’t use them. Earlier research labs are like, we don’t have any inference and we don’t have a need for Huawei. Any company that has meaningful use of their models has figured out how to run them on Huawei for inference, which, to Jensen’s credit, is like — it’s happening when he said it was going to happen, but it’s not that surprising.

Grace Shao (51:11)

Yeah, the Dwarkesh interview. I don’t actually understand why he got so much hate for it because even without your political stance, what he said actually made sense logically by saying, if you don’t sell them the crappier versions of what we have, they will have an equally quite crappy version to serve themselves, or they would just want...

Nathan Lambert (51:28)

I think they would buy both. Buying both is actually true. The amount of Nvidia chips that you would have to sell to China for them to stop buying Huawei — because Huawei is almost surely way cheaper because Nvidia margins are insane — when would they actually stop buying both?

Grace Shao (51:43)

But then you have to go on CANN. You have to reroute everything back on CANN. The developer ecosystem is not there. That’s Jensen’s point, right? Or the habits are not there. So I think that’s what, when I talked to a research lab...

Nathan Lambert (51:50)

Yeah. But I’m saying they would also use Huawei. I think they are so supply-limited, they would use both. Anthropic uses everything. A lot of companies in the U.S. will use multi-platform. Meta is a huge buyer of AMD. Demand is so high that any chip that is potentially viable on the models within a few generations is very valuable. And the fact that you can run some reasonably large model on any Huawei chip is a big line crossed for Huawei.

I don’t know if they can produce the volume of chips and scale that quickly, especially as they try to move to lower nodes. That’s the standard semi debate. But the question is: can Huawei scale production? That’s the only question. And if Huawei can manage to scale production, Jensen will just look really right. If Huawei can’t scale production, Jensen will look a little bit like a lunatic, but it will be outside of his hands.

Grace Shao (52:43)

And we don’t really know what happened during this trip. It seemed like nothing really substantial happened after this big Trump delegation. It was more like a high-profile tourism trip versus an actual deal trip.

Okay, I want to ask you something you wrote about that’s a bit niche, not something you usually write about. It’s on the SaaS side of things. You said that there’s a common argument that China struggled to monetize AI because they’re unwilling to pay for enterprise software. We looked at how China tries to monetize on consumer AI, but clearly that’s not really been proven yet. In your piece, you push back on the claim and say that there’s a distinction between SaaS spend and cloud or inference spend.

Tell us about what you think about that ecosystem and how Chinese AI labs are trying to make money maybe a bit differently from American AI labs.

Nathan Lambert (53:32)

I don’t know if it’s necessarily different, but I ask a lot of researchers about this. They say that everybody is trying the new AI tools when they come out. If they don’t like them, they stop using them. If they like them, they keep using them on the consumer side. So something like Claude Code would be an example: tons of people tried it. I’m guessing lots of them churn in China, just like in the U.S., but consumers are very quick to adopt and try new things, but won’t stick if it’s not actually serving them.

And then the enterprise is like: there’s definitely cloud that exists. Digital services are gigantic. They essentially think that there’s more runway for making money on AI models that falls into that. And they all use coding agents; they all use Claude. It’s a hilarious thing. They’re all very Claude-pilled. There’s almost no mention of Codex, where in the Western media, Claude versus Codex is this whole thing. They all use Claude. And that is obviously a paid service. So I think there are cracks in the argument, and I expect AI models to be seen as a bit of cloud, but potentially it is the thing that changes some of the expectations, where it’s just so transformative because they’re so competitive, and it could be seen as a bit of a phase shift.

Grace Shao (54:41)

Yeah, and I think it’s a generational shift, a phase shift. Also, actually, recently Doubao raised their prices on Seedance usage and whatnot, and it’s a shift into trying to capture the prosumer market. You can say the average uncle and auntie on the streets still don’t want to pay for a consumer app, but I think there’s more prosumer market share that could be captured in China, maybe not fully enterprise either.

I want to ask you about government roles and geopolitics. I know there is a common narrative that usually people assume Chinese AI labs are heavily subsidized. Actually, when I was in San Fran in March, I was at a dinner with a couple of investors, mostly public investors, and one guy asked me, “Hey, are all labs just basically subsidized by the government?” I was like, definitely not. The majority of them are not. If not, they frankly don’t want to take money from the government.

It was really hard for him to understand that, because I think the misconception is all Chinese labs or Chinese tech are just funded by the government. Kind of to our point earlier, where any affiliation to any government agency, just by default, is assumed to be therefore backed. First of all, the government, I don’t even know if they have that much money to give out. Number two, I don’t think that’s how competition works, right? So what’s your thought on all of this?

Nathan Lambert (55:55)

It seemed more like a provincial government trying to help the companies do stuff, which is like get offices, get talent. I don’t know what the provincial government can do. In Beijing, there’s Beijing Academy for AI or whatever, which is a real research institute that’s just funded by a certain neighborhood in Beijing. It was like, okay, the U.S. could do that. But much less of the Ant Group-style thing, which is government takes major ownership stake in an investment round and goes on. Maybe Kimi’s latest round, there were mentions of government-backed VCs, and I don’t know how that kind of intermediary works. So I still think it’s very indirect. And because the government system is so competitive across the different layers, each of those layers are competing to help the companies, but they don’t have piles of cash sitting around to buy GPUs.

Grace Shao (56:44)

No, they don’t. And they frankly don’t know what they’re doing half the time. This is an argument like what you said: Haidian District or Chaoyang District of Beijing will be funding an academy, and the academy will be in the effort to help AI go toward AGI. But the reality is they’re trying to follow this high-level KPI of being like, let’s make AI happen. All they want to do is write in their report and say, we funded something about AI, so we’ve hit our quota. I don’t think it’s as hands-on as people assume.

Nathan Lambert (57:09)

Yeah. If you read Breakneck — most U.S. tech people haven’t read Apple in China and Breakneck — and all you need to do is read these books and learn a little bit about the interface between tech and China and understand that they are also hyped about AI, and then you’ll understand that it’s a messy trickle-down process in the Chinese government. It would be very obvious if they were nationalizing a lab. It would be as obvious as if it was in the U.S. It has not happened.

Grace Shao (57:18)

Both great books.

Yeah. So to close, what is the biggest disconnect between how the U.S. AI ecosystem right now thinks of China, Chinese AI, and what you saw on the ground? And what is something you think we didn’t touch on today that you want to share?

Nathan Lambert (57:52)

This is the thing that everybody asked me. They normally asked me the first thing when I got off the plane: what’s the big thing? And it’s like, I don’t think there’s anything that shocking. I think that many people just haven’t read basic books about how tech is interfaced with the government, and know these things, or hear narratives that are very geopolitical, which is targeting the top end of the government system and how that in the U.S. engages. And there’s a lot of shrapnel from that. Anthropic pushes very aggressive China narratives, and Anthropic is a very followed company in tech.

Most people don’t spend the time on this in the U.S. ecosystem and just don’t go deep on it. I don’t have anything shocking. It’s good to encourage people to do some of that because these dynamics impact things like: the Chinese open models are really influential, and now Silicon Valley is building AI. So it matters to a lot of people, but they don’t study the causes of why they might do this. They just are like, it’s here, I don’t need to think about China.

Grace Shao (59:00)

Yeah, and I don’t know what it is. You and Bill Gurley were saying this in a couple of his public appearances. It seems like Chinese researchers, tech people, CEOs, whoever, are a lot more aware of or following more closely U.S. leaders and thought leaders, tech leaders, business leaders, than vice versa. There’s something about that. I don’t know if it’s just easier to dismiss it or easier to not have to learn something new. But the goal of AI...

Nathan Lambert (1:00:27)

I think it’s American culture. American culture is very obsessed with its own weird world. Yeah, it’s hilarious. American culture is ridiculous. It’s so ridiculous.

Grace Shao (1:00:27)

As a Canadian, I can’t say things like that. You said it. I’ve lived in the States, but I can’t say this. But I think, look, shameless self-plug here: as a Chinese Canadian, my life goal here with AI Pro is really just to bridge that gap. I think to your point, there’s going to be geopolitical narratives and rhetoric at the very top, but for the average person, or even for builders, tech people, whatnot, it’s probably in everyone’s benefit to understand what’s happening on the other side and stop alienating it or stop making it as if it’s so different.

I think throughout this conversation, it’s really just to say, look, so much of it is so similar, but so much of it is slightly different. The difference is not really a government mandate versus maybe a cultural difference or resource-constraint difference, especially in building technology. But that’s kind of my view.

One last question for you, which is a question I ask everyone on the show. What is one differentiated view you hold? Throw me something crazy.

Nathan Lambert (1:01:28)

I know, I’ve always kind of been open-models doomer, even though I build on them. It’s just that it’s so unsustainable, and there’s so much money to be made with building closed software, that I’m constantly doomy about the prospects of open models. I’m always a skeptic.

Grace Shao (1:01:40)

It is a bit sad, isn’t it? How does a company like Hugging Face actually make money?

Nathan Lambert (1:01:45)

I don’t know. You can look into how much money they actually make. It’s not very much, unfortunately.

Grace Shao (1:01:48)

Yeah, I think that’s the unfortunate reality of the capitalist world we live in. As much as it incentivizes the competition and breakthroughs, that doesn’t help with what we just talked about earlier. Yeah.

All right, Nathan. Thank you so much for your time. Really appreciate your insights and your sharing.

Nathan Lambert (1:02:04)

Yeah, thanks for having me. Good to see you.

We get really into the nitty-gritty of what makes “good” tech versus “bad” tech for kids and why the category of ‘children-first tech’ is very overlooked. Reni explains why most children’s apps are built on an “attention economy” model that forces them to compete with addictive content, and why his team needed to build physical hardware to break that cycle.

We tackle the hard questions, including the pushback from parents who believe in “no tech” childhoods. And he shared his most non-consensus view: that the era of standardized, industrial education is over. He believes we are entering a golden age of “scaled homeschooling” where AI meets kids where they are. Whether you’re a tech investor or an anxious parent, this conversation about nature versus nurture, “nei juan” (involution), and raising resilient humans in an AI world is a must-listen.

Subscribe now

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, founders, and product managers.

For more information on the podcast series, see here.

To find the previous episodes of Differentiated Understanding, see here.

Chapters

00:00 Reni’s Journey to Dex Camera

03:48 Designing for Children: Principles and Insights

08:05 Technology’s Impact on Child Development

12:09 Bridging the Gap: Business and Product Design

15:36 The Role of Parents in Tech Development

25:20 Leveraging AI and Language Models

29:48 Value-Driven Pricing Strategy

32:05 Defining the Product Category

34:33 Subscription Models and Content Delivery

37:58 AI and Parenting: Balancing Technology and Safety

43:29 Unexpected Use Cases and Impact

47:29 Personalized Education and Parenting Philosophy

AI-generated Transcript

Grace Shao (00:00)

Reni let’s start with your personal story. Who are you and who are your team members? Because when I met you in SF, I was so enamored by the product and I thought your story was so interesting. So please share that.

Reni Cao (00:11)

Hi everyone, my name is Renny, CEO and co-founder of Dex. We’re a technology company in San Francisco, almost all parent company, which is pretty special in a startup setting. We’re a bunch of parents that having trouble with the same kind of like a reality where like our education system is a sort of like cookie cutter and our entertainment is also cookie cutter for children.

So we’re like, can we harness technology, especially the latest development of the AI, in different way for families that really gives children a chance to become the best version of themselves and ⁓ give the digital autonomy back to parents themselves rather than accepting the fact that they have to struggle between technology versus no technology. So yeah, we’re the parents, of like a bunch of missionaries in this journey together to explore how can we make the best use.

of the AI and our first product is called Dexta Language Learning Camera where kids can take pictures and turn the whole world into language immersions. And it’s a product targeting young children three to eight. And we’ve sold 10,000 pieces so far and ⁓ ratings has been high and we’re pretty excited about this. But yeah, this is pretty much about us.

Grace Shao (01:24)

But Reni, tell us a bit about what you did before Dex actually. What kind of led you to this path? I know becoming a parent really did inspire you. You have a young daughter, I think similar age to mine, around three years old. But before that, what really led you to this path? Were you always passionate about children’s tech or education?

Reni Cao (01:41)

I actually have been a product management guy for the last decade in Silicon Valley, some big companies like YouTube and LinkedIn, some smaller s***, ZFS, Wish. But I have been a builder since the beginning. I would actually say that my passion for decks actually originated much earlier than I started my career. It actually started right when I was at school, but happy to say more if you’re interested.

Grace Shao (02:07)

Yeah, no, do tell us a personal story there.

Reni Cao (02:09)

So I was always this random kid with tons of questions back in high school. And very unfortunately, I think the education system, especially in East Asian countries, is not designed for meet kids where they are. So every time when I come up with a random question, my teachers are usually a little bit impatient and will be like, can you just go back and finish your quiz, et cetera, et cetera.

So the moment I saw when GPT-4 comes out, I was thrilled and I posted a long like blurb on LinkedIn. Basically saying like, you know, if I had, have this as a kid, I would have grown into a more complete human. So this kind of like, I feel like this like generative AI’s capability to meet kids where they are, especially meets your needs for curiosity. It’s game changing. So.

I feel like I’m building this product first and foremost for a younger me that could have benefited so much from this. That’s pretty much the story about me. yeah, I know we see it and of course our parents right now we see there is a tectonic shift in terms of the skill landscape and what the future of workforce is going to be and even the existential challenge of what does human mean in a future society.

So we do want to build something that’s centered around children, centered around the family to help them find what they love and build agencies around it at the end of the day. So yeah, that’s the two main driving force of me coming to Dex. But I would be honest about it. It’s like very random. When I want to start a company, a lot of my colleagues are very surprised, being like, oh my god, Renny, you’re getting into this field. But yeah, I guess I finally find the work of my life.

Grace Shao (03:48)

I love it. think you need to understand the passion and the personal reason behind the businesses to really understand why the design was frankly so intuitive and why you’re so passionate about building this and leaving such a comfy, know, like cushy corporate role. I think that’s the one thing that stuck out to me. The product itself is actually so natural to how children behave to your point, like my three year old.

from morning to night, know, morning she wakes up, it’s like, mommy, what’s this? What’s this? What’s this? What’s this? How do say this? Why do you know that? Sometimes she gets angry at me. If I don’t know something, she’d be like, but you’re an adult, you should know everything. But the reality, especially with languages, it’s really difficult. So for example, yesterday she was coming back from her Mandarin class and she said, liu shu, she was pointing at random tree. And I was like, that’s not liu shu. All I know is not liu shu, but I actually don’t know what liu shu is in English because I think it’s only really common in mainland.

I’ve never seen that kind of tree. Well, I guess it’s a willow tree. You don’t see it very commonly elsewhere. And then she kept on pointing at trees, but in Hong Kong, you clearly don’t have liu shu because Hong Kong is like tropical. And then she got really, really mad at me. And that moment I was like, wow, if we had a Dex camera, that would have been perfect. But I was literally trying to take a picture of it while we’re moving car and try to upload it to GBTB, like what tree is this? What’s the name of it? So anyway, I think it’s really great product design. And I want to kind of get into that a little bit.

When you were designing it, what was the thinking? Like, what does it mean to be children first?

Reni Cao (05:10)

I think there are three layers of children first as a principle. The first layer we already touched upon that. So young children, their hand anxiety is very different from adults.

they tend to use one hand to operate a device and another hand they want to use for sensory explorations, like they want to touch. Sometimes they want to just move things around. So this requires a different form factor that one handed use, very tactile, very intuitive for young children such that they can explore a world while harnessing the power of AI in this case. So this is kind of like the user, the special things about the user.

And it’s a different design. think that’s layer number one. I think the layer number two is also that the device itself is a metaphor for the market as well. And in the market, we want to build something that’s drastically from the so-called adult-centric smart devices, namely the phones and tablets, to send the market a message that there could be a different option. There could be a good technology. There could be a family-centric technology. And we’ve picked this form factor

utilizing the metaphor of magnifying glass. It is something you use to see some hidden wonders, otherwise you cannot see. I do think that’s the ⁓ second layer of the things, which is like metaphor and category creation. And at the end of the day, I do think we intentionally make the device kind of worth finding this fine balance between engagement and

learning or kind of like a healthy aspect of the technology, meaning like we add a assistive screen, but we make it really kind of like limited and not the center of the whole kind of like a user journey. And we want to kind of like find a new way to put all the components in our consumer electronics world in a way that it strikes a more delicate balance and ⁓ let the device itself to be kind of like, you know, retentive.

for children without getting them to be addicted. So it’s kind of like we intentionally make it a little less stimulating, actually much less stimulating than a lot of a thought-centric ⁓ product. So that’s the main three kind of like principles around the product design. There’s a lot of conflicting constraints here, as you can see, but we do our best trying to find what is the answer. And here we go. Like what you see right now is our first, you know,

⁓ answer we have thought through and I think the market validated the answer quite well so far.

Grace Shao (07:39)

Yeah,

definitely. think exactly to your point, know, like a lot of times, I think when we as young parents looking at introducing technology to children, really worried about the big screens, addictive nature, or even the parental, even though a lot of them allow parental control, it’s the unlimited access to a wild, wild internet out there. Like all of these things are basically concerns and or reasons why we hold back technology from our kids.

So actually on that note, do think you kind of mentioned it, right? Like technology over the years, especially big tech frankly, has garnered a bit of a bad reputation. And I think that was really tied to the rise of social media and all of this mental illness that came with it. And obviously like you mentioned the addictive nature. So what do you think is actually harmful to the children’s development when we are looking at tech? What are areas actually we can really embrace technology?

I think you kind of touched on it lightly, maybe explain it to us in an even deeper, more technical way.

Reni Cao (08:37)

Yeah, our thesis is that why a lot of parents think technology is negative for a good reason. And the reason is that all the main status quo technology for children are built on top of the attention economy, as we call it. Everything revolves around time spent and how much attention, how much engagement in terms of like a minute, seconds, sessions you can get.

That, is the reality because, think about it, you build an app on an iPad, immediately you’re entering a competition with Roblox, with YouTube Kids, with all the videos, all sorts of things out there. You could do well. You can try to do good for the society, for the families, but you’re effectively competing against more like...

addictive kind of like a form factor of information and it’s a losing battle and as we call it is a rat race. So no matter what type of like educational apps or content you’re trying to deliver at the end of day you have to deliver them in more and more engaging way more and more gamified and more and more animation used etc etc. That’s I think that’s why it’s another reason why we need hardware at the end of day. I think the first step

how we can create alternative reality is that we need to create a new world, a new kingdom where the business is built upon outcome rather than attention. Meaning like it’s not the time spent logic anymore. It’s like, can you use this device? For example, for Dex, you can use the device and you see the child speaks better after two months or your kid starts to have a like a love to speak Mandarin and not preserve the rest of their childhood.

I do think there is a business model there like that, but I believe that business model warrant a totally kind of like a different design of the experience from ground up, from the device layer to the software, to the content, all the way to like user interaction. So I do think like that’s why the current technology is considered bad because it raised towards attention. And I think ultimately, inside Dex,

I believe the final answer to create that alternates like a reality is can we deliver something that’s purpose built for children before we build a general sort of like, you know, like time spent logic, like a product in, in, in the case of the decks, is something that, you know, purpose built around the languages. cannot do a lot of things. It cannot, it’s not a chatbot. It cannot, it cannot play videos.

But I do think even do one thing super well with the Frontier technology already delivers so much value to the families such that you can build a viable business model on top of that while creating values for families. I think being courageous enough to limit our scope to something to begin with, like really hold onto our principle, deliver a promise, create values there.

is another internally operating principle to get there in terms of how to harness the technology. And I want to say that it’s very interesting. What we noticed that is a lot of people are trying to use the AI in quite an all-in-one way. So you can see a little device with tons of features in there. can generate pictures. You can talk to celebrities at chatbot. You can talk to Elon Musk on that device. And we think, actually, that would be a very slippery slope.

⁓ in terms of harnessing the technology at the end of the day. yeah, purpose-built is another very critical principle we’re holding on to, to create a good technology.

Grace Shao (12:09)

No, I love that. But I mean, from a business perspective, sometimes people might not have purpose built businesses, right? Unfortunately, some are not. Then thus, how do we basically help the industry align the business incentive to the product design incentive? Because, know, like what you’re saying right now, it makes a lot of sense. And I think once I saw Dex Camera, I was like, wow, why is there not something like this on the market?

but it does feel like there’s a huge gap where, like you said, there is a big devices and the big tech. There’s this tiny niche little products, whether software product or hardware for children’s ⁓ use, but it doesn’t feel like people are taking it seriously, even though we all know parents are willing to spend on children if it’s for their good. It’s not like the economics doesn’t make sense. So why is there’s that gap right now?

Reni Cao (12:56)

I think you’re hitting on one of our most recent realization that the parenting needs and the children’s needs are quite long tail or as we call it, very like a versatile, right? Different parents have different parenting needs. Even when you look at the language as example, there are tons of different languages and even more dialects you wanna learn. Like let’s say you wanna learn Mandarin, you still got so many like a dialects there. There hasn’t been a real...

kind of like technology that can enable a venture scale business that attracts talent, that attracts a good backing in terms of like a capital to build something that’s like a generational. But I do think this is the moment AI is strong. We finally have to make sure we can build one system.

that can consolidate all those long tailed needs. Even for Dex, very specifically, you can learn a lot of languages and even more dialects with just like a nine person team building the hardware plus software. I think it’s the catalyst that’s much bigger than Dex itself. And I’m really excited about that. But I think another very interesting angle is like, despite the technologies there, you have another question. It’s like why there is not

more company like Dex. I have a personal opinion here. When new technology comes out, people will tend to use it in the most sloppiest way possible. They were trying to just like, OK, you can chat with the AI, so why don’t we just shovel AI into a little box and put it into a Talking Fluffy and call it an AI toy. And that’s it. That’s my business. I do think it is like a gravity that’s pulling people away.

from deeply think how to harness technology and pulling them towards something that’s so trivial and it’s just almost like a shortcut. I think that’s kind of like also, I would call that a trap on the entrepreneur side, that the technology is changing so fast and everyone’s a full mowing, everyone just wanna use it in some way. But I think in this sense, we as Dex, the company, we believe in that.

we need to think very deep about how should we use this technology to meet users where they are and deploy like AI in certain ways so Shell can deliver the value. So that’s why we start small, but we’re going to expand from there.

Grace Shao (15:09)

Yeah.

No, it makes a lot of sense, but I think I wonder if you guys all being parents like you just said have made a huge difference. I hate to overgeneralize, but like, you I’ve been in the tech space for 10 years, but usually either I meet men who are like 20 years older than me or they’re very young men who have not, you know, settled into a family yet. And I’m just saying when I tell people my mom, it scares people. They’re like, I don’t know what to say. I’m like, OK, like.

I’m not trying to scare you off by telling my mother, but the reality is most of us one day will all have families. And when we do, we start thinking about the things around us very differently, our perspectives shift. And I think to your point when you guys had a lot of purpose designing this product, I wonder if it has a lot of, you know, reason because you guys are parents. Whereas if someone is an entrepreneur for the sake of being a business person, they might not have the nuanced understanding of what a kid needs and what they even think is good for a kid.

So to your point, they create little stuffed animals with an L-I unplugged into it, which is horrendously scary. I would never introduce that to my kid, right? I’m getting very, very agitated about this. But you know, another one that we talked about kind of offline was like, I should be ambassador and be paid by Tony Box at this point, because I probably gifted at least like 20 of them out to friends with kids. I think they’re just like, on the surface, you think about it, they’re like, ⁓ a little box that plays music. You’re like, this is so easy. I can just use my iPhone.

Reni Cao (16:13)

Me neither.

Grace Shao (16:32)

to exactly to your point. It gives the kids agency, allows the kids to start navigating the world themselves and have preferences. For context for people who don’t have Tony boxes or kids at this point is you put these little miniature IPs, essentially they’re Disney or whatnot, and you can put them on the little box as a magnet. And then the box starts singing and has like seven or eight pre-programmed music or ⁓ stories. And then you can control with your little hands. And basically like you press the

Reni Cao (16:54)

stories.

Grace Shao (16:58)

big ear, the ear just like the volume goes up, small ear, the volume goes down. It’s like really, really great. So basically introduce technology to kids where they’re like, oh mom, I can control what I want to listen to today. But I don’t need to nag you about it to control the iPhone. I don’t get exposed to a screen. And I can sit there and be entertained for like half an hour myself. So I think Dext really falls into that category for me. Like, you know, we kind of skip the part where we explain how your technology work really and in a very day to day way.

It’s basically like you hold a camera, you point at things, you click the button, you can say, what is this? And you default choose languages, right? You actually explain better than me, please.

Reni Cao (17:35)

So there are actually four questions here. So I want to actually react to all of them one by one. I think this is a lot of good insights here. I think Tony Box and Dex share one thing in common, which is they are children-led, or they are child-led in this case. Think in the POV of a child. The world is kind of like a scary place that you’re told to do this or that.

you are brought to here or there, there’s not much quote unquote autonomy you could have. But now there’s a device that your parents actually are willing to let you operate and you can decide what type of content media or interactions you can get. That is just a huge reward to children’s like unlimited curiosity and their like a strong needs to be considered sort of like, you know, a big kid or a

even grown up in a way. I think that’s the intricate magic that if you were not a parent, you haven’t interacted with children a lot, you will miss. So instead of saying like a parent’s made us a better product builder, I think at the end of the day, it goes back to the product 101 that you really need to know your user. You really need to know who are using your product. We spent such a long time with our kids every day.

And early days, which is very funny, like ⁓ the first group of users using DAX is just our own children. And that gives us a huge edge there. Right. And I do think you mentioned that a lot of like startup founders in this category, sometimes they’re doing something with raised eyebrows of the parents. I do think they’re a little bit distant from the kids is one reason. And another reason is I do think there is a misconception that children are less.

at the end of the day, lot of founders think, you know, those are toys or some gimmicky stuff. Kids, you know, you just give them something that can flash, they can make some sound, and children would love to use them. But I reject that answer. I think that assumption is completely wrong. Children are actually smarter than adults in certain ways. They just cannot verbalize it. But as I said, they already got their little taste.

as the famous word, popular words, they got their taste and they sometimes can tell what’s a soulful piece of story versus it’s a very sloppy kind of story. So children actually knows that and they want quality experience, they want quality product, they can actually absorb something that’s really built well for them. I think that just gives us kind of like this endless.

sort of motivation to polish our product as if we’re building this for the most critical sets of adult users because we think actually children are more and they deserve more. Now, coming back to how Dex works at the end of the day, I think the core loop of Dex is quite simple. You just take the little camera. I’m happy to actually send a video to be the bureau here. You just take a picture.

⁓ And they would just literally just tell you, let me actually take a selfie here. Hi. Let’s see what I can learn about this. Look at that big smile. It’s like spreading happiness everywhere. Can you say a smile?

Just smile.

smile.

Yeah. This is like you get unlimited, like smile comes with some laughing too. It’s when you make happy sounds like, ha. Can you say laughing? Laughing.

like the ones we use to listen to music. Do you like music too? Can you say headphones?

This is actually English immersive mode. So you can, you can improve your vocabulary there.

Grace Shao (21:06)

how many languages you have now.

Reni Cao (21:08)

We have 16 languages and more than 30 dialects and it’s still expanding. And interesting observation here is like the smaller, the more niche the languages is, the stronger the demand is there, which we find is super interesting.

Grace Shao (21:21)

probably just harder to find offline solutions otherwise, right? Or like harder with the communities, assuming you’re an SF, finding a Mandarin community is not that difficult. You know, if you’re in England, finding a French community, probably not as difficult. if you go, you were saying like maybe like Arabic languages like that are not as mainstream, maybe in San Fran, you have people in San Fran wanting to do that, right? Or like people in Dallas last time you said, trying to learn Mandarin, which again, you don’t have a huge community. Very interesting.

I’m sorry, I got very passionate about the topic. So I want to of swerve back to our conversation here about raising children with technology. I’m sure you get pushback. think people right now, there’s the other side of argument where everything should be organic. Everything should be very simple.

Reni Cao (21:53)

Yeah, of course.

Grace Shao (22:08)

And I myself, I’m a big fan of a lot of the Montessori toys. You know, they’re not buttons or not even power charged. They’re just little wooden blocks, but they’re designed very well for them to, you know, develop motor skills. So how do you kind of explain to parents today who are saying technology should be rejected in the childhood. Kids should just be reading physical books. should learn the way that we learned or even like previous generation learned. We should go back to touching grass only. So

Like, yeah, what’s your argument there?

Reni Cao (22:37)

First of all, you are completely right. Every once in a while, we got a comment on our social media that, why don’t you talk to your own daughter to teach that language? Why do you need a device to do that? So your assumption is completely right. And my response to that is, first of all, actually, I respect that parent a lot. I believe in the most ideal world, organic human-to-human interaction and free play in the real world is great. There’s a lot of tech, like researchers actually

Prove that right, right? However, I do think the parent miss out constraints here. Number one, you may want to talk to your daughter, but you don’t know Cantonese, for example. So there’s no way for you to teach some subjects or some skills that you want them to learn or you want to immerse them with. And second, all of us know that the contemporary society is more and more fast paced. Not all the parents enjoy this privilege.

of saying, let’s slow down, set up a dedicated time for children to go out to places. All sorts of this ideal family style back in the 80s and 90s changed a lot, I would say. So we are, believe, rather than just blaming the parents, not spending enough organic time with their children, I do believe that technology should be introduced more as an option, as kind of like a gap stop.

as one of the extra tools on the table. That’s why when we design decks, we don’t introduce chatbots, but we spend so much time on sharing the insights that what your children are interested in. What did they take a picture of? What do they want to geek on? What did they learn today towards the parent app? And just give them this little window to see the world through their children’s eyes. Give them good downtime topic.

giving them a way to reconnect even as asynchronous. So I do think the concern is real and the overall kind of like, you know, judgment is very well reasoned. But I think what that’s the approach here is much more nuanced than saying like, let’s use technology to replace human. It’s not, it’s actually using technology to connect the humans, connect the parents and kids better. That’s the nuance I have to take a bit.

Grace Shao (24:42)

I see what you

Yeah. No, no, I love it because actually I’ve seen some parents even give kids like little Kodak cameras these days and these little toddlers go around the world, take pictures of how they see the world and they’re so cute. My own daughter sometimes takes my phone and takes pictures around the home and I come back with a lot of selfies and pictures of her sister’s foot or it’s just very cute because you see the world through their eyes, right? And it gives like, it’s like technology doesn’t take all connection away.

on technology. wanted to ask you about the technology. How do we understand that? Like how are you actually leveraging LLMs? How do you route through different LLMs or different languages? Is this something we talked about briefly? But I wanted to understand that bit more.

Reni Cao (25:20)

to share details. Where should we start?

Grace Shao (25:22)

Like how does it work? right now? So basically for the little Dex camera, can’t ask it, like he’s to your point, you didn’t build a chatbot. So I can’t ask a question. I can’t have a conversation. It’s not a companion, but I can ask it what’s this? How does all that work in terms of the back end technology and the guardrails you built up?

Reni Cao (25:38)

Yeah, I think

in a 30K feed view, Dex are utilizing basically all the multimodal LM capabilities to understand what the children are looking at. And on top of that, we build sort of like a profile, interest profile for the children and the parenting need profile for the parents to help contextualize, you what responses should we give in that case? To give an example, if you’re a three year old,

just starting to learn Cantonese and you are sort of like interested in a bunch of like a museum topics or you love like dinosaur skeletons and stuff like that, we will render you more challenges around kind of like hey let’s bring Dex to a museum and learn about different terms there and it will be English the primary languages teaching entry-level Cantonese things there. So basically like the visual understanding you certainly use like a multimodal LLM

The response definitely use kind of a conversation API of a lot of like an ALM. And I think building out this context layer or this memory layer of like a children’s interest and parenting needs, that actually is more complex. That takes kind of like a full agent system to try to understand what matters, like condensing or distill insights into a profile and gradually kind of injecting that into our responses. I think that’s on a very high level. That’s it.

We do use a wide range of LLM, mostly with Gemini and OpenAI. yeah, that’s kind of like the high levels.

Grace Shao (27:08)

I’m going ask a question you might not like, but I’m going to put you on the spot. When we talked last time, said specifically on Cantonese and Mandarin, you do use different LMS, but the accents can be quite funny. Like they’re a bit off. They’re not native sounding. Why is that? And how do you overcome something like that? Or other maybe non-English languages. Yeah.

Reni Cao (27:12)

No, ask me.

First of all, you need to try again because we have a solution already. But definitely, hit. We are already squeezing. I’m so hard that we’re hitting the boundary of a lot of like, in this case, it’s a TTS of the leading providers. Because I think about it, I’m pretty sure you’re using English plus Cantonese. It’s basically using English to learn Cantonese. Is that the case?

Grace Shao (27:50)

Yes.

Reni Cao (27:51)

is a mixture of languages cases. The challenge there is that without fine tuning, there is very limited sample of someone that speaks very good English and very good Cantonese, and they mix them in like one sentences. So the data, the training data to begin with is a little flawed. Either you have accent English or Cantonese as the more common cases. That’s the fundamental root causes of this. And we’re having kind of like heavy lifting tasks to kind of like solve that.

And with the foundational model getting better and better, think one day we’ll get there. And we can see that to be fully fleshed out in the next six months. You definitely hold us accountable. And I think this is right observation for mixed languages. It’s really hard. Yeah.

Grace Shao (28:32)

Yeah, I bet. how does it actually work right now? Like in terms of economics, like people pay you about $249, right? That’s the price of the product pre-tax. That’s not cheap. Like it’s much more expensive than a toy, but obviously bit cheaper than an iPad. How do I understand the pricing decision there and price? And then how does that relate to, I guess, how you pay for your token usage right now? Does that cover it?

Reni Cao (28:57)

Yeah.

Yeah. Oh, big time. We actually have a pretty healthy margin and the tokens are getting incredibly cheap. Much cheaper than where we started. I’m talking about like in 96, 97. It were a fraction of the token cost of where compared to when we just getting started, which is back in 2024 February. At that time we don’t even have GBD4, we have GBD3.5. that’s the kind of like, that’s the kind of like, actually that time we have GBD4 but

is we don’t have GPT-4.0. So it’s very expensive at that time. So now pricing. Actually, I have a let’s talk about the user-centric view and a business-centric view. On the user side, we’re actually adopting this value-based pricing model, which is like any enough day, language is a high value skill to acquire. I sent my daughter to a language immersion in the US. I’m very embarrassed to mention how much I spent on that school.

Grace Shao (29:33)

Okay.

Reni Cao (29:49)

And if DAX can offer 1 % lift or enhancement on top of that school, the price is fully adjusted and much more than that. So this is what I mean by like, and very funny that you mentioned toy, right? Toy is something that you get it, you play it for a couple of days, then you don’t see it, you don’t worry about it. And this is not what we’re trying to do. What we’re trying to do is we want to use a relatively high price to keep ourself honest about

the value we’re delivering to the parent. Do we really teach a language or do we really get the kids to fall in love speaking that language? If we do so, that price is well-justed. If not, we’re going to give you 90 days of free return period. No question asked, just return it to us. I do want to use this pricing model to push us to deliver more value for the user. So that’s one aspect of it. And on the business side, very funny, you mentioned, I hate when people box us.

into toy category. I don’t blame them. Natural reaction, but I want to send a signal to the market that if a team of talented people, hardworking parents, put their heart and soul in building a purpose-built device that harnesses AI and delivers concrete results, we could get out from the typical, stereotypical, like a toy average order value band and go much higher. Above that, it’s less about, I to keep my

is more kind of like, want to send a signal to prove that the market, we have enough parents waiting anxiously for something similar to this and want to pay a perceptually higher price for it, a premium for it. But yeah, that’s kind of like we landed on that price. And it’s so funny that so many people in the early days tell us, you’re going to do $1.99, because anything that started with a one

Grace Shao (31:22)

Premium, yes.

Reni Cao (31:36)

is night and day different than like two, that it started with two. But I actually, I’m like launching a suicidal mission. was like, let’s actually make it start with two, but let’s deliver more value there because it’s never like, it’s not a retail business at the end of the day. We’re trying to create a new paradigm of digital parenthood and childhood. We need to hold a high bar for ourselves. And the price is very telling, like in that case.

Grace Shao (31:59)

No, I actually agree

and I think would you categorize yourself in the same box as Tony box vertical? Would you?

Reni Cao (32:06)

Not really. ⁓ Tony Box is a, I would say they are a content business. they are, same thing with Yoto. Actually, their founders have deep backgrounds in labels, music labels specifically, and IPs. So they are effectively a distribution business that they are creating a new channel to distributing those IPs from Disney, from Spin Master, and et cetera, et cetera. And the other side, you can see that at Dex, we’re not

Grace Shao (32:19)

I see.

Reni Cao (32:31)

I think like IP partnership or putting characters on our device. And we actually optimize for value and outcomes, like I promised to you in one of our principle. So I would put ourselves in, I don’t know, the de facto smart device for families. Just very honestly, the family device, the family technology, maybe like this is where we’re trying to go to, but it’s a completely like non-existent category before we’re still exploring.

Grace Shao (32:33)

Yeah.

Yeah, yeah.

Okay, like family tech device.

Reni Cao (32:59)

and it may change how I call it.

Grace Shao (33:00)

think there’s some more

similar things maybe in East Asia because the audio learning like you know even when I was very young like I remember my grandma had a 步步高步伏机 I don’t know if you know what that is it’s like those like tiny little yeah yeah basically what it is it’s like people learn English with it and I think it’s very very like mainstream in China for a while but like you know these things been around I think in East Asia because everyone is using it to literally learn English

Reni Cao (33:11)

The steps are fine.

Grace Shao (33:24)

But it’s very one dimensional. It’s like one language to one language. They basically embed a dictionary, make the dictionary into a digital one. And you can ask search questions. You can ask what this word is. might, more advanced one might be even like with images, but I think, I don’t know, in the 90s, I didn’t see any images. But yeah, it does remind me of that technology and that vertical. haven’t seen something like that too mainstream in the West growing up, you know?

I think if I was when I was learning French and German growing up, that would have been so helpful to your point. But yeah, so I want to bring it back to sorry, I just want to bring it back to the the business. On the Tony box comment, I do believe their business actually could be really high margin because their product is only say like 199 or something like that, right? Like they’re the box. But each character is not a 20 bucks or 30 bucks.

⁓ My daughter is drying me up here because every two months she asks for a new figure. But my point is, it’s a great business, right? Like that thing just keeps selling. It’s like Spotify and a physical thing. So would you guys have add-on any services, software, hardware, anything?

Reni Cao (34:32)

We do.

That’s a lot of investor has been pushing us regarding this razor razor blade business model. I think for us though, what we are ultimately delivering is a business more like an app store.

It’s like where you can get personalized content and software for your parenting needs and for your children’s growth needs at the end of the day.

We’re launching, not we’re launching, we launched two tiers of subscription so far to validate that. One tier, $10 per month, you got unlimited LTE, plus you actually got a curriculum packed in like a content library. Every day we give you one topic and in the topic you can explore a lot of new vocabulary, expression, know, new languages and it’s good kind of like content to consume. And I think what’s most interesting is our future vision is actually a $20 per month tier.

In that tier, you can actually create activities for your children, personalize. Grace, can be like, I run this podcast. I’m a podcast host. How do I explain that to my kid and make it a little bit fun, exciting, and even adventurous as if the recording a podcast is a little journey? And by the way,

my kid likes this way of storytelling. You could give a lot of like a prompt there. They’re actually based on the profile, the context layer, we’re gonna build sort of like interactive, like a content that involves taking pictures, speaking, and just like looking at the device for explaining what does podcasting mean. And this tier actually got really good like attraction. And when we look at their subscription retention,

is above like 90 % in three months that shows early signs of product market fit. But this is what I mean by like our business setting of day. We are a channel to deliver like harnessed intelligence to parents such that they can build whatever content and software that adapt to their needs rather than just a purely search, then filter or control kind of like a timer. I really want the digital world to revolve around them, running out of way around. So in this case,

Put it in a simple way, we give them a tool to build whatever they want, and we charge on the usage of the tool, pretty much.

Grace Shao (36:41)

No, I actually really see that. I love it. Because I think my husband was trying to use chat GPT for a while to create stories with my daughter. Like, add a pig, add a dog, add a whatever in this. And obviously, it’s not made naturally for this. So the stories don’t come out as, I guess, natively understandable for children. So I see where this can go. And the funny thing, you use my profession as an example.

Reni Cao (36:49)

Exactly.

Grace Shao (37:05)

example, like my daughter just thinks I talk all day, that’s my job, and she thinks that her dad sits at a computer and press buttons all day. So between the two of us, none of us are doing it much, just talking and pressing buttons. So it’d be really great if, you know, I can, I guess, lean on technology to find a better way to explain to children modern day careers, you know, that may be not as easy to explain as, know, mommy’s a doctor, and doctors go help people and save lives, which is like what my family has.

you know, explained to us when we were growing up, you it was very clear. I want to kind of go on a little bit more about AI and parenting. I think there’s a huge discourse right now in the US, especially, I think from my point of view, where I sit in Hong Kong, in Asia, even yesterday, I was speaking to someone from South Korea, venture capitalist, they’re saying that parents and society seems to be a lot more open to bring technology into their day to day lives.

They’re much more open to the idea of leaning into technology for personal use and less worried about privacy and you know these kind of issues I guess. So at a high level, what do you think, should we be concerned when we introduce technology to children? will they, you know, for example, taking pictures themselves that automatically goes into one of the LLMs. Is that something that...

he should be mindful of or are there guardrails that can be built in?

Reni Cao (38:26)

We should be definitely mindful. That’s why we enforce ZDR, zero data retention across our stack for images. So even let’s say your kid take a picture of themselves, you cannot retrieve that picture even you want. You can ping me through my personal email. You cannot find that picture anymore. And OpenAI and Google signed a contract with us to burn a picture immediately, like zero data retention on all the usages. But overall, I do think

Grace Shao (38:48)

See.

Reni Cao (38:51)

It’s the company’s responsibility to introduce technologies to family and the family should hold a high bar there for sure. Because like the AI is so early and it’s way too powerful in certain way. And it’s like a kind of like a black box in certain way in a lot of different ways. that I definitely, I’m not a, I’m not that one of the technologies that wanted to like, you know, glorify AI and it is the future and stuff like that. comes with a lot of risk, especially like unproven.

aspect how it impacts the children’s cognitive development and something like that. That’s also a reason why we work with researchers and professors ⁓ closely like in Mount Eucalon from UCSF and Harvard professors doing education and doing research using text. I do think there is a substantial risk here such that the

And we as the entrepreneurs and we as the parents, we need to hold a high bar for ourselves and roll out things one by one. So I guess that’s why you will hear more about like, oh, that’s like, you could have done this. You could have made it more engaging. You will hear this much more often than you’d be like, oh, there is like an incident because, you know, we always prioritize, you know, safety first. We’d rather the device to be boring in certain way rather than introducing consequences that we don’t understand.

So I think there’s a very interesting dynamic between the Western and Eastern in terms of their views about technology. And I don’t think it’s a family parenting only. It’s also even a whole society, general perceptions. Happy to chat about that, but maybe it’s a little bit off topic here. Yeah.

Grace Shao (40:12)

comes from the mindful design as well.

⁓ No,

we can definitely talk about that a little bit, but I kind of just follow up on what you just said. So how should a parent evaluate an AI device or tech device when they are purchasing for children, right? ⁓ I’m sure there are different devices out there, maybe not exactly doing the same thing as what you’re doing, but other devices are tech native ⁓ or AI enabled for children. How should parents kind of go about this?

Reni Cao (40:52)

I’m not a parenting coach. I will share my views. Number one, do think we should bias, we should start from our needs first. Maybe let me put it this way. Don’t get carried away with all the possibilities of the AI. Ask yourself, what is the unresolved parenting needs you have and find solution there. Rather than, this AIX, then that’s just to buy that AI device and give it a try. That’s number one, I would adopt that.

Number two, I do think it’s important to see what a company’s method is. They definitely put their methodology somewhere, their belief somewhere, their principles somewhere, they’re kind of like, like how you ask me about how we use ALM. I believe that the parents should definitely hold the company accountable to explain those details and ask, verify, and that’s crucial step. That’s the due diligence on them, right? And I do think that

Finally, for any sort of AI product, I actually even think the parents doesn’t have to be getting into this searching and validating mental model. They could literally build their own in some sense. Given all the agent codings rising up and reducing the piece cost of software so low, I do think for lot of stuff, they should try.

to accommodate their own parenting needs in certain ways. Like I saw tons of the parents go into cloud code generating like a, know, almost like a story writer for their daughter. That’s actually my previous colleague at Wish. And it was awesome. It’s just different blanks to fill in. It’s kind of mad lips type of like a story. I do think the parents can change also their way that they are in the autonomy right now to build whatever they want to build.

Having said, it’s still a little bit of kind of like a Silicon Valley bubble type of answer, because honestly, in the world, the adoption of a cloud code is probably less than 2 % or 1%, I’m pretty sure. But I do think I would encourage parents to use AI themselves and explore a boundary, it can do, what it can does well, what it doesn’t. So then, the kind of I make a decision from there.

Grace Shao (42:45)

Yeah, no, I I appreciate that. It’s a very like thoughtful answer because it’s not just like A or B. of the day, think it’s parenting itself is so personal. It’s on how your family dynamics work, how you prioritize your time, how you want to parent. So when you want to buy technology for your children or incorporate that into their lives, it’s also a personal decision. I wanted to ask, actually, do you have any good case studies to share with us just a little bit?

Reni Cao (43:11)

We have quite a lot. What aspect, what, what type of case does he want?

Grace Shao (43:14)

Just like, I don’t know, like things that unexpected people use. For me, I mean, by default, just assume, yeah, people use it in urban areas, right? But then I think when I met you, you said, actually a lot of people use them, you know, in unexpected places, like orders come through all over.

Reni Cao (43:19)

Alright, I’ll give you one.

One of, I immediately think about one thing, one, almost like ⁓ close to 5 % of our users, they bought Dex to help with speech delay. That’s something we never anticipated, but those parents are very frustrated with all the, as we call it, sometimes autism tech or the speech therapy tech there. It’s not meeting their bar and they saw Dex, they’d be like, I would try everything right now for my kid. And surprisingly Dex helped them.

and it makes them real happy. And you can find actually all those real reviews in our review sections. Quite a few family mentioned that their kids refuse to speak certain languages or just even English, but that’s kind of necessitate the language as a fun activities. And all of a sudden, the kids start to open up and speak more, and the parents are really happy about it. This same exact story happened with my co-founder, who is really worried about his, at that time, two-year-old young son having speech delay.

But I want to disclose the name, but the song he actually first time spoken like coherent like ⁓ Chinese phrases using Dex and he caught it on a video. That was one of the most wholesome moment of our kind of like a user feedback in our channel. And right now we’re actually ⁓ volunteering to develop this special need mode. That’s kind of like, you know, customizing to special needs children.

especially like April is the world kind of like autism awareness month. And yeah, we just want to do it. And we want to donate to Dextre researchers and speech therapists to help us do it. This is a totally kind of like a side quest, but it just like give us it gives us so much kind of like energy. You’re thinking about technology can be used in a way that’s like immensely helpful.

Grace Shao (45:00)

That’s amazing.

Yeah, and something unexpected, right? Okay, I think I want to wrap up our conversation because I don’t take up too much of your time, but I do want to ask you one big macro question. With you working on whether you like to call it physical AI or not, essentially like a physical product hardware time software, how do we understand that trend going forward? Do you think AI will be essentially integrated, plugged in to more more hardware devices? What’s your view on that?

Reni Cao (45:33)

I do think there is a consensus that every wave of software technology revolution, there will be kind of like a device revolution following that. We are at the tipping point there. That’s like people starts to reimagine, where is this? this cloud? Is this the recording card? Maybe it should be a separate, like an ⁓ AI. Or this is a sort of like a little pendant that can kind of like ultimately listen to your life, help you organizing. I do think we’re at the ⁓

the dawn of a next wave of hardware. But it’s less about we’re doing the hardware because of the, I do think this is a software or technology driven type of hardware revolution out there. I do anticipate that. I do at least what I’m 100 % sure is like smartphones are not designed for children. Tablets are not designed for children. Families deserve something built with their interest.

their needs in the center of the spotlight. And I see that happening. And that’s why we started this company. And I bet there’s going to be tons more use cases there.

Grace Shao (46:33)

No, amazing. Thank you. I think ⁓ one last thing. Is there anything I missed or anything you would like to share with us?

Reni Cao (46:39)

By the way, I time. If you want to turn through all the questions, I’m happy to be here. I don’t have anything else after this meeting.

Grace Shao (46:44)

no, don’t worry. think it’s a lot of times like I use them as prompts. But you know, when we’re chatting, like we actually covered most of it, you know. ⁓ Yeah, is there anything you think we missed? But from my end, like I feel like I covered most of it. You know, we did technology, we talked about children, AI philosophy, talk a bit about your business model.

Reni Cao (46:50)

Yeah. Yeah.

Yeah.

I do think you would want to talk about. Yeah, go ahead. Go with one last one, and I have one for you. Yes, go ahead. Ask yours first.

Grace Shao (47:04)

I think one last one. You go.

So I want

to ask you one last question, which is a question I ask every guest that comes on the show. What is one differentiated view you hold? I feel like your whole thesis around devices right now on the market are not made for children is already a differentiated view. But is there anything else you think that you hold that’s non-consensus?

Reni Cao (47:29)

Yes, with this view, I got beaten up so many times, but I still got to say it, right? I believe that education should not be cookie cutter. It should be highly personalized. So is entertainment. So is the parenting software. And we’re about to enter the golden age. Finally, this is becoming the reality. And let me say it this way. You look at a school in the US, how you tell the school is good or not, you look at one ratio. It’s called a teacher-student ratio.

One teacher taking care of less, but why? Because then the teacher can accommodate, individualize the needs. I actually have a very radical view in terms of our education system is definitely lagging, significantly lagging against how our society evolves, how the technology evolves. It’s still a one size fit all and industrial way.

to handle education, handle like, you know, testing, standard testing. It hasn’t really changed in the past couple of decades, but the world is a different place now. And I guess my view is like, it shouldn’t be that. The default shouldn’t be that. The default is like every kid should almost have their personalized tutor and the playmate that deeply understand them. Unfortunately, that’s impossible before, resources-wise. But I guess we need to strive to get there.

as a race, as a humanity. Because each kids just come up, come with their own spark.

that will miss out the window to make that spark their lifelong journey. But I’m not trying to attack on educators or school systems, something like that. I just feel like there needs to be more forces from the society, especially from the tech side, to help together build this alternative, enhanced of like a system that really delivers individualized education.

Sometimes I use the word scaled homeschooling. And you cannot imagine how much people hate that. people are like, homeschooling, you’re taking away the social aspect of it. People are very constrained on the vocabulary of how they describe things. But I guess when I say homeschooling, it’s not about keeping the kids at school and hiring a teacher. And that specific process right now, I’m talking about really meet children where they are in terms of their growth, in terms of their needs, in terms of the skills they’re going to develop.

I call that a differentiator, but maybe actually lot of people will share the same views. I’ll be happy to know who shared the same view and please join us in the journey. Follow us along.

Grace Shao (49:49)

I definitely

think that view is definitely, feel anecdotally a lot more prevalent in SF when I visit. I’ve met other people like yourself, other people in the tech space or, you know, investors who are embracing this idea of modern homeschooling. And they say the same thing. They’re like, we don’t like to use the word homeschooling because, it sounds like a bit more cultish, but it really isn’t right. Like it’s really focusing on individual ⁓ growth.

I think it’s amazing because I also think it’s because Silicon Valley itself kind of harbors this kind of growth and mentality and that the fact that people can succeed without degrees, people can succeed by building different things, people can succeed in just being different and being themselves, but the best version of themselves have always been, I think, what drives a lot of people who want to go to Silicon Valley because it’s like in many ways, as a mayor, talk to see like the best version of my talk to see, right? I think

in East Asia, even as I put my kid in school right now, I find people definitely a lot less like that minded. ⁓ I don’t know if it’s a cultural thing because like, know, for you, you know, I grew up in Canada for me, I always felt like, you know, having that freedom to learn, explore when you’re young, which is more Western kind of way of, I guess, education was good. But I think a lot of peers here actually believe that, you know,

for the first like say eight to 10 years, that foundational education should be drilled in. know, ⁓ grit should be taught, discipline should be taught. But it’s very interesting because it does kind of, I guess, manufacture different kinds of stereotypes. And I think it’s fascinating. And I think one more comment on that, I know this conversation has been more personal than we thought it would be, but I love it, you know.

I don’t really get to talk about motherhood that much in my podcast. It’s usually about tech and bros and tech bros and about, ⁓ and about finance. but I think even, you know, when you have kids, people talk so much about nature versus nurture. And what I realized is I was shocked to see the nature come through, as young as like six to eight months in a child.

Reni Cao (51:36)

You

Grace Shao (51:53)

their personality starts coming through and by the time they’re one to one and a half, they start kind of babbling, start demanding things. I realize 80 % of it is all nature. It’s like their preferences for how they socialize, their preferences of even noise, even you can realize like your point, your taste. You’re going to find a six months old who just wants to sit in a corner in a play group who just wants to flip through books literally and just undisturbed. You’re going to find someone who’s screaming in the middle of the whole group.

you’re gonna find my daughter who’s rolling over everyone and just like trying to knock everyone out. And I don’t know why. You know, you’re gonna realize all of it is nature. And even I believe agency, autonomy, grit, and desire to actually succeed, that itself is nature. And I don’t think you’d be taught. And I think this is a bit controversial. But definitely I think my husband and I have been thinking a lot about this. We’re like, we can just provide them what we can. But there is no...

point of even pushing them when they don’t want certain things. the best is to push them in a direction that they want to be pushed and they will tell you. I think this is like kind of the difference in our generation of parents. yeah. Reni, thank you so much. ⁓ Yeah, go on.

Reni Cao (52:49)

Exactly.

Yeah, but can I, I know this

is over time, but can I add one last comment towards what you say? But I think what you said, especially growing up in East Asian, like, you know, education system, it has been industrial for a good reason, right? At a time where stuff like AI doesn’t exist, the most effective way,

Grace Shao (53:03)

No, of course, of course.

Reni Cao (53:19)

to develop fundamental knowledge workers, plus finishing the job of dividing the children into different segments and give them different levels of education. That education system works perfectly. I entrance exam, as I’m talking about, taking standard tests and stuff like that. But all we know that is AI is sweeping through all the knowledge works.

and specialized in knowledge works, honestly, Asian parents like favorite jobs, like being a doctor, especially radiologist, you know, and or being a lawyer, you’ve got to start somewhere as associate. Now it’s getting kind of like his hardest. The world has already changed. The tsunami already hits. But I don’t think people actually understand the level of the shit. A lot of like everyday people in the world, they haven’t felt.

this like a tsunami, right? So when you say you want to kind of like, you know, like find define your children’s nature and push them towards kind of like what they are intrinsically motivated about and give them resources to set them up for success, building grades are on the way. I do believe that I think I will 100 % agree with you that it will become the most fundamental aspect or element of education in the next like five years or even sooner to be fair.

That’s why I don’t send my... to put it in a simple term. I don’t send my daughter to Kumon. I don’t want my daughter to do Russian math. I never benchmark her against like, oh, like the other kids can read at the age of like three and a half. Why don’t you? Actually, I don’t because I fully understand that kids have their own time zone. Kids have their own spark. All you need to do is think deeply to define that, to understand that, understand why my daughter sometimes is super sensitive, understand why sometimes she got frustrated and want to hit.

Grace Shao (54:31)

I feel very validated.

Reni Cao (55:00)

Don’t take that on a surface level with the other tools you have. Go deep, understand that, and build these programs that’s personalized to her and help her. And I think like this is why I if I, talk about the word of Nei Juan a lot. If I have to dream on anything, right? I have to like a rather like ruthless compete on anything. I complete the deaf understanding of my daughter rather than anything else.

Grace Shao (55:03)

100%.

Reni Cao (55:26)

Because I actually think that’s the thing people gloss over. People must be like, education is just checklist. You got to check, check, check, check. And there is a better checkbox. Like Ivy League school, there’s a OK checkbox. There’s a worse checkbox. Forget about a checklist. That checklist is obsolete already. So I respect. I think we vibe together in terms of our schools of parenting.

Grace Shao (55:33)

Yeah.

Yeah, 100%. No, I agree with you.

parenting style. Yeah, yeah,

yeah.

Reni Cao (55:51)

But you’re so fully intuitive. don’t know whether I’m right or wrong, but this is what I firmly believe in. And I believe someone’s going to join this journey.

Grace Shao (55:58)

I think there’s more people who are aware, especially people who are more plugged in with the technology because they realize how fundamental society will change. I just thought about when we were young, I’m sure your parents also told you to go to university, go to this, go to that, right? For sure there was a hierarchy in their mind, what kind of school you should go to, what kind of degree you should get. Now I really don’t think that’s the case. Actually, a lot of my readers would even know.

my dad really forced me, well, pushed me, encouraged me to go into finance. And at one point he was like, if you don’t study finance and don’t work in finance, you’re not like following my footsteps and blah, blah, blah, blah. Right. And it was a very, it became a personal reason to do it. It’s not because I wanted to, or I was good at it. And there was actually a battle between us being like, I want to go into journalism. And he’s like, no, I was like, no, I’m going to go to journalism. He’s like, I’m not going to pay for it. You figure it out. But the beauty of it is actually found a way resourceful enough to get a full time, full scholarship.

And I still want to journalism. Again, I recognize how lucky I was. I I found the opportunity to do that. But most kids actually just end up then doing what their parents told them to do and they never, and they never actually live their best life or become the best versions of themselves because they’re doing something not actually fundamental.

Reni Cao (57:08)

you hit a very critical, I think it’s a background or context. There wasn’t an abundance before, right? Growing up, let’s say in the eighties, It’s a relatively kind of like a society. It’s relatively kind of like not that sort of like, you you wouldn’t call it abundance at the time. Let me just put it that way, right? You still need to compete for stability, compete for resources. That’s why there’s a rat race in education, which I totally understand. That’s kind of like.

It’s like a whole economy there, right? But I think that changed. No matter what we’re talking about, like, I mean, in China, I’m talking about US right now, I think abundance really will hit at some point of time. At that time, the challenge shifted from how can I avoid getting into a property or like ⁓ job loss towards kind of like, how can I find the meaning of my life? And how do I deal with this kind of like a journey?

It’s a generational theme there. It’s very funny that your dad wants you to go into finance rather than journalism. mean, for those who understand Chinese internet a little bit recently, there has been a famous influencer called Zhang Xuefeng. He almost helps everyone to pick their college major. And one of the college major he hates the most and advice everyone to not go to is actually journalism at the end of the day. Because it’s just not a...

stereotypically stable job that can make a lot of money, that can give you social status, quote unquote stuff like that. But I think that’s a lack of view of things. Our dad doesn’t know how our skill landscape is to be 20 years later, just like we’re not going to know what our kid is going to deal with. So we need to give them a more sort of generalize the resources and the skills at grit to survive. And it’s funny enough.

I made my, I named my daughter Simone because we are big fans of Simone de Beaufort, the kind of like, you know, foundational philosopher of feminism and a lot of like a sociology thoughts. So, you know, if I have, if I have to give a set of expectation for my daughter, I want her to actually do something that’s not commonly seen as a very, like a prosperous or stable kind of like career ladder. I want her to do something that’s kind of like

has a mission and some sort of like outlier type of journey. let’s see how that goes. She’s so young, so who knows.

Grace Shao (59:20)

I love it. All right. Thank you so much, Reni.

Reni Cao (59:23)

Thank you.

Listening to them, I realized how incredibly lucky I’ve been. AI Proem has grown almost entirely organically, simply by harassing 2-3 friends in the beginning and things took off on its own on . Now, with over 10k followers online and having met so many of you offline, it feels a little surreal. As you know, nothing is behind a paywall but I do provide paid private sharings sessions. Over the course, I’ve simply written about the things I genuinely care about, that excites me, and only when I felt I had something meaningful to say. [For more, see here]

Now, I rarely get writer’s block because there’s always far more I want to cover than time allows. So when I do let my anxiety get the better of me and publish something that feels a bit of a ‘filler piece,’ I apologize.

So today, I wanted to humbly ask for a small favor: If you’ve ever enjoyed any of my writing or interviews, would you please share AI Proem with a friend today? It would mean a lot. Your support is what keeps me going. Thank you.

Share

Everyone’s talking about humanoid robot marathons, and Silicon Valley’s hottest livestream is of a warehouse robot. But there is actually a more interesting physical AI story unfolding before us that is going unnoticed in the West: a Chinese vacuum company that now makes sports cars.

You know which Chinese company is making a lot of noise right now? Dreame. The vacuum company. The home appliance company. The robot vacuum company. The “AI-powered whole-home ecosystem” company. The one that debuted its huge demo event in San Francisco in late April, talking about cars, phones, smart rings, robotic arms, AI laundry, robotic lawn mowers, and a future where every consumer electronic will be reinvented.

How do you even categorize a company like this?

At Dreame NEXT 2026, held at the Palace of Fine Arts in San Francisco, Dreame staged its largest international showcase to date. Officially, it spanned smart mobility, smart home appliances, personal devices, and personal care. Unofficially, it looked like a hardware founder trying to convince America that a Chinese vacuum company could become a full-stack consumer technology platform. Dreame’s own messaging was that “in the age of AI, every product deserves a reinvention.”

The Verge’s coverage was maybe a little less generous: hundreds of influencers, some celebrities, and many products that either had no pricing or launch date, or felt more like concept theater than shipping hardware. There were phones, smart rings, a rocket-powered EV concept, and enough robotic arms to make the whole thing feel like a showroom for a future, but everything was still somewhat in the R&D process and hadn’t passed procurement yet.

Both versions are true to some extent, but I think the company is worth learning about because its pursuits and vision- not in humanoids - seem like a much more realistic immediate next step in the next stage of consumer robotics, if one were to put robots in their homes.

Yu Hao doesn’t do quiet, or lose

In today’s Chinese tech environment, as in after the whole Jack Ma saga, most Chinese tech (all rich people) founders have become careful, sanitized, or gone out of their way to be invisible. Yu Hao (余浩) has been going the opposite direction. He talks a lot, he makes public statements a lot, and he’s very active on social media. All of which is probably a headache for his PR team but somewhat entertaining and refreshing in an environment where everyone’s either hush or reading off corporate scripts. He talks about building one of the greatest companies in human history. He talks about a future Dreame ecosystem worth 百万亿市值. Now what is that? One hundred trillion dollar market cap company, and for context, that is like ~20 Nvidias. Like I said, not a humble guy. His ambition is to grow the Dreame ecosystem to the largest company in the world, essentially.

While most Chinese tech leaders of this generation have retreated from public life, they have let their PR teams handle the messaging if there is even a messaging KPI, and have stayed off social media. Yu Hao is doing the opposite, actively commenting on other companies’ high-profile social media moments, making himself the product before every product is ready. Some say this was a deliberate strategic decision due to the fact that the company used to serve a 2B audience, to pivot to 2C, it needed to recreate itself or bolster its street creds. And what gets more attention than an unhinged, unfiltered boss who embodies everything he has built. Because in a consumer hardware market where features get copied fast, and product cycles compress, attention itself becomes a moat, or at least a financing tool, a recruiting tool, and a channel tool.

Now there is also ample authenticity to this. His personality, boisterous, courageous, and arrogant, seems to have come with his success from an early age. In many ways, he even admits, he’s never not won.

Yu Hao is one of those founders whose life story reads almost too cleanly. He was 保送, which means guaranteed through every stage, meaning he was recommended to the next level without taking the grueling nation-wide/provincial-led middle school-to-high school exam zhongkao, nor the high school to university exam gaokao, or graduate entrance exams. He entered Tsinghua through the physics competition results. He studied computational fluid mechanics and launched 天空工厂 (Sky Workshop), a student initiative on campus that hand-selected its members. The student aerospace project even attracted attention and investment from Boeing. And guess what? Upon graduating from Tsinghua, he founded Dreame, and his core team was mostly from that group as well. He famously has said that once becoming number one becomes a habit, you get addicted to it. That line later became the spirit of Dreame’s internal slogan: either don’t do it, or do it to be number one. And in most categories, they have achieved that.

His ambitions and arrogance know no ceilings. Talking to the media, he once said, paraphrased: “ Every bluff I’ve made, I’ve also made it come into reality. Jensen Huang and Elon Musk created 8-10 trillion-dollar businesses, but they’re a generation older than me. I can create a new ceiling.

The Xiaomi launchpad and escape

As we know, he studied aerospace engineering, so when he thought about starting a company, there were two paths. Something about spaceships or something that might be easier to commercialize.

Looking at how it all started, Dreame did not begin with the idea of making a better vacuum. It began with flight, airflow, and motors. Yu Hao has said that when drones were hot, he decided that drones were too niche. He wanted a product that could enter millions of homes, had a real technical barrier, and had technology that could transfer into other fields. Dyson had already proven that high-speed motors could define premium home appliances. So Dreame went directly to the hard part, spending two focused years on motor technology to match Dyson’s capabilities.

And with the team’s commitment, it found an opportunity to join the Xiaomi ecosystem at the end of 2017. At that time, 2B manufacturers were largely content with making white-label products for distributors and brands. And Xiaomi’s home appliances, ranging from kettles to vacuums, largely depended on these firms.

Little did Lei Jun expect that Xiaomi would give Dreame its platform and a launch pad for its later sprawling business categories. Of course, protective measures were put in place, but Xiaomi thought that it was hard for any new players to come in and challenge its channel access, low price, premium quality, and, especially, for a young company to replicate.

Within the Xiaomi ecosystem, most companies remained in a single category. Dreame didn’t, of course, Yu Hao didn’t let that happen. It moved from cordless vacuums to robot vacuums, wet-dry floor cleaners, hair dryers, and more, often in parallel. Yu Hao wanted to expand across the supply chain. He pushed into robotics vacuums and hair dryers, which I’ve heard actually led to some friction within the ecosystem. Peers were shocked that it was breaking the unspoken gentlemen’s agreement, which is to stay in your lane and be fed comfortably by Xiaomi.

But now we know, it was all a conscious push on Yu Hao’s end; he was slowly pushing the envelope.

Yu Hao’s answer was that Dreame had built two layers: a bottom layer of long-term technologies (motors, vision algorithms) and an upper layer that combined those with external suppliers and channels. New products didn’t need to start from zero. The company could reuse the stack.

The escape strategy was clever. By 2020, Dreame’s own-brand products already accounted for about half of the company’s sales, while overseas sales were roughly 70% according to 36Kr’s reporting. Yu Hao explained the overseas-first move simply: they didn’t have enough money. Domestic promotion was expensive. Overseas markets were more open, and Chinese companies like Huawei and Xiaomi had already spent years educating the market abroad. His shorthand was: use overseas to hit domestic, use high-end to hit low-end, use new channels to hit old channels.

And here’s a detail from their early brand-building that I think is underappreciated. When Douyin (TikTok’s Chinese version) opened up to e-commerce at the end of the 2010s (2018ish), there were basically no legitimate premium brands selling on the platform. Dreame went all in on Douyin commerce early, making it a core channel for 2C marketing when everyone else was still skeptical. For a resource-constrained company competing with Xiaomi’s own distribution, that was a genuinely differentiated bet on a channel no one else was taking seriously yet.

Whether intentional or not, the success came about when a few ingredients converged. Dreame used Xiaomi to learn how to make and figure out what sells, used overseas markets to prove it could stand alone, used Douyin to build a brand when no one else was premium on the platform, and then came back to the domestic market with higher-end positioning.

Core tech vs Core customer

There are two types of companies that can expand beyond their original category, usually.

The first type expands through distribution and user ownership. Tencent can push a new service because it owns attention. Meituan adds categories because it owns local commerce behavior. Apple adds services because it owns the device relationship.

The second type expands through the reuse of core technology. Honda and Mercedes make motorcycles and cars, because the underlying technology and advantage is the capability of the engines and the ability to assemble them together. DJI moves from drones into cameras, gimbals, and robotics-adjacent products because stabilization, sensor, imaging, motors, and control systems transfer.

I’d call this Core Tech vs Core Customer. Dreame is betting on Core Tech.

Motor. Airflow. Battery. Sensor. Algorithm. Motion control. Manufacturing. Channel. Yu Hao’s argument is that these are transferable hardware primitives, and if you build them right, entering each new category becomes cheaper because everybody’s favorite word - sYnerGy. This is why he’s described Dreame as ‘没有边界的公司’, a company with no boundaries.

If that sounds grandiose, it is. The team spent the first two years of their initiation just figuring out how to optimize the motors. At the time of its inception, as I mentioned above, Dyson was the holy grail of luxury home appliances, and their goal was simply to match Dyson’s motors. Motors go into vacuums, hair dryers, fans, air purifiers, lawn robots, and maybe cars. AI goes into navigation, cleaning, laundry, and phones. Robotic arms go into vacuums, mowers, laundry machines, and refrigerators.

Now, Yu Hao, the young entrepreneur, but legend has been called a Elon Musk + Lei Jun, funny enough, the two ‘seniors’ seem to have had their encounter recently. He also famously spent 2.282 billion yuan (~340 million) to acquire control of Jiami Packaging, a listed company on the A-share market in 2025, which even fueled speculations that the company would venture into food packaging (lol I guess no borders/ boundaries right). But that claim was quickly denied as “false information” by the company, and since then, IPO rumors have not really resurfaced. Part of the reason it hasn’t been smooth for the company to list publicly is its ambitious (insane) goal.

So how do you put a valuation on a company that does everything? I mean, beyond the vacuums we talked about, it claimed to build an AI native EV to challenge Bugatti Veyron.’

Wheels before legs

So the reason Dreame caught my eye is that amid all the humanoid mania, a robotics company like itself seems to be rejecting it altogether.

Everyone wants to talk about humanoids because, well, we’ve written before 1/ ego 2/ sense of self-importance 3/ buzz and what people say “our physical world is designed for humans’ but you guys know how I feel about it all. It’s simply just not there yet. And I, for one dont understand why industrial six-axis arms don’t get more love.

At CES 2026 in January this year, 21 of 38 humanoid robotic displays were Chinese. The marathons and demos definitely fuel the excitement. The discourse follows Elon Musk’s vision of a general-purpose humanoid, and the rave followed. But Yu Hao understands that hardware does not scale as easily as software. There’s no Moore’s Law for legs, actuators, batteries, reliability, and service networks. In fact, it is said that more than half the cost of a humanoid robot is in creating the legs and the form factor. Using wheels is much cheaper. And that’s why we see Sunday Robotics or Dyna Robotics optimizing with the arms on wheels, adding a ‘head’ like sphere, I guess, for funsies, because a headless bot could turn some people off.

Someone shared a stat with me that I keep thinking about. For every step in a robotics process, assume a 99% success rate. Sounds great. But if you’re running a 30-step process, which is considered long for consumer robotics, your total success rate drops to about 74%. That’s the math that makes hardware hard. The chip isn’t cheap, the software models for robotics aren’t strong yet, and the cost structure is overwhelmingly in manufacturing.

The thing is, if you push through the noise and hype, think about what an actual practical use case of robots in a home environment is. A vacuum that can know not to suck up a sock, and elevate itself over a step, or a human-looking bot that can run fast.

And why Dreame’s consumer specialized robots and not humanoids, have actually built a business, a real, viable business. Some may say Dreame’s bots dont look ‘revolutionary,’ but a functioning bot that can clean the pool or mow the lawn really adds more tangible value than most of what we’re seeing on the market.

We are the results of our times

There’s a saying in Chinese, 时代对个人的塑造, which essentially means “How our times shape who we are.” It is often used to describe how time and place determine a person’s outcome more than their efforts. Some use it to describe their own success as they rode the wave of privatization and the opening up of the economy in China in the 1990s. Others use this phrase to showcase that, at times, even the smartest and most hardworking won’t get their break because of a dire environment.

Dreame’s founding was closely tied to the rise of OEM suppliers to big electronic brands. Dreame’s consumer- branding success may be due to the genius of Yu Hao’s strategic positioning. But now the technology itself can be attributed at least partially to the bigger movement. As the country’s decades-long efforts in building, enhancing, and grooming a generation of companies, technology, and talent in the spaces of manufacturing, lidar sensors, battery storage, and so on have definitely helped. It’s the convergence of technologies that is embodying the boom of China’s AI robotics industry now. So the story of Dreame is not really just about Dreame, it’s the whole ecosystem around the robot.

High Capacity

China's overlapping tech-industrial ecosystems

China has developed multiple tech-industrial ecosystems that overlap in terms of the firms and technologies involved. China doesn’t just have a smartphone industry or a battery industry or an electric vehicle industry. It has all of these industries and more. China’s strength across multip…

Read more

a year ago · 436 likes · 36 comments · Kyle Chan

Just as wrote in TechBuzzChina after two trips across China visiting factories and showrooms, she “came away less convinced that humanoids are close to becoming household helpers, but more convinced that China’s advantage in robotics is not any single robot. It is the closed loop around the robot: capital, factories, data, supply chains, customers, and local governments all pushing in the same direction.”

The market data reflects this. IDC’s 2025 data shows the global cleaning robot market shipped 32.72 million units, up 20.1% YoY. Lawn-mower robotics grew 63.8%. Window-cleaning robots grew 70.4%. Chinese manufacturers are setting pricing benchmarks, commercializing AI navigation, and scaling globally through e-commerce. IDC’s conclusion was blunt: supply-chain depth and rapid iteration have become decisive competitive advantages.

And you can see the category reshuffling in real time. iRobot filed for Chapter 11 in late 2025 and was acquired by Shenzhen PICEA Robotics in January 2026. Reuters reported that iRobot entered bankruptcy with roughly $190 million in debt. The company that popularized consumer robot vacuums ended up being owned by a Chinese manufacturer. Meanwhile, Chinese brands like Roborock, Ecovacs, and Dreame compete in the premium segment.

Two decades ago, German and Nordic brands dominated the premium consumer electronics space, and Japanese and Korean companies were the suppliers of many. A decade ago, consumer electronics used to be made in Japan, then Korea, then China as a low-cost factory. (Btw, even Boston Dynamics is now owned by Koreans)

The old story was the price of export but the story now is iteration speed and truly quality design. Dreame’s own numbers claim overseas revenue at nearly 80% of total sales in 2025, North America revenue up 189% YoY, products in 42 million households across 120 countries, over 10,000 patents filed with 3,000+ granted, and R&D personnel at more than 70% of the company. A Dreame-distributed release citing IDC’s tracker claimed 27.6% unit share of European robot vacuum shipments in 2025.

Even though I’ve written about how some Chinese companies are still positioning themselves as robot OEMs for global brands. That is not Dreame’s endgame. No, for Yu Hao, born in the late 80s, like many other entrepreneurs who grew up in an economically vibrant China (DeepSeek’s Liang Wenfeng), their confidence and ambitions since day one have been in the global market. They’re the 主角s main characters, not NPCs.

*Company exec just pulled out of intv ….

A Google DeepMind researcher talks candidly about what frontier AI work actually looks like, why individual genius matters less than people think, and why the AI stack may end up looking a lot like cloud computing.

As you probably know by now, I’m a fan of Zhang Xiaojun’s podcast (张小珺). It’s one of the best long-form conversation shows for researchers and tech leaders in Chinese, with episodes lasting 3 to 4 hours each, which is both the appeal and the problem. I’ve been listening to her interviews on and off for years, and I'm so glad to see it gain mainstream traction more recently. I usually listen selectively because, not going to lie, it’s hard to squeeze in 4 hours of concentrated time. But this episode with Google DeepMind’s Yao Shunyu (姚顺宇), his last character being ‘universe’ vs the Tencent Yao’s ‘rain,’ 姚顺雨, I’d actually recommend sitting through if you can. He was a very candid speaker, less polished than many business executives, but there was a sense of earnestness and humbleness from him, and most importantly, a rare glimpse into what frontier AI research actually looks like from the inside.

Quick note on the name before I get into it, because it’s genuinely confusing. There are two well-known “Yao Shunyu” in AI right now. Both went to Tsinghua and actually around the same time. One is the former OpenAI researcher who joined Tencent last year and now leads its model effort as chief AI scientist. The other one, the protagonist today, studied physics at Tsinghua, did theoretical physics at Stanford, joined Anthropic during the Claude coding and large-scale RL push, and then moved to Google DeepMind. They’re in fact good friends and used to “just hang out” when they were both in Silicon Valley. What are the odds? Honestly, even people in the industry mix them up sometimes.

For personal preference reasons or company policy reasons, frontier researchers almost never talk this openly. In a humbling way, maybe a byproduct of cultural upbringing, or just his candid view of the world, when asked what he thinks he has contributed to the projects he’s worked on, he said if it wasn’t him, it would have and could have been anyone else. At this point in life, only a handful of people can be credited for the progress of the technology, whereas the majority are just part of the natural evolution. Things would have become how they are anyway, maybe not with the exact same timing or the same people, but the direction was already there. If it weren’t him working on large-scale RL for coding, someone else would have. If it weren’t Anthropic, another company would have figured out coding agents. If it wasn’t Boris, Claude Code might not have looked the same, but someone would still have realized the next product after a chatbot is an agent harness (OpenClaw was developed independently, Gemini apparently had internal teams building similar things).

His metaphor was, AI is a wave. Whether you surf it or not, it still hits the shore.

Once the paradigm is found, history starts driving people, and not the other way around.

individualist contribution

We usually talk about AI in terms of individuals, right? Tech moguls, star researchers, transfer rumors, who left OpenAI, who joined Meta, who’s at Anthropic now. Better yet, who has Chinese tech companies courted recently? In many ways, it feels like celebrity sports league stories now, and much of the focus has been shifted to a few names and the speculated drama. (Rui has something amazing cooking here, a tracker of all the Chinese researchers)

Yao makes the opposite claim. He says that for LLMs, once the paradigm was found, the work became more industrialized. Pre-training has a paradigm. Post-training and RL have a paradigm. The industry now broadly knows what direction to go, and that was in many ways echoed by Chinese researchers too in February; they used the ‘copying homework’ analogy then, that’s all. Different labs have slightly different techniques and approaches in their trial and error process, different infrastructure capacity, different data buyouts (which you can question; they may be somewhat the same at this point), and different tastes, which is supposedly what really makes the difference now, but the general direction itself is no longer a mystery.

He argues that the transformer architecture needed the pioneers to invent it—the decision to really bet on pre-training scaling needed heroes. At every paradigm shift, key people and key teams still matter a lot. But once the paradigm is established, it starts to feel less like people are driving history and more like history is driving people. I found that framing quite eye-opening, given how much is mystified about the industry.

He basically said that AI research doesn’t necessarily need that much “brain’ anymore. What matters more now is being reliable, careful, hardworking, and responsible. In many ways, it’s about the ability to execute meticulously, which many may argue is why there seems to be a dominant percentage of Chinese researchers in the field. This wasn’t the first time I’d heard this, but I’d always assumed it was genius-talking, the kind of thing brilliant people say to sound modest. A researcher from one of the leading Chinese labs told me earlier that the way to train LLMs is already quite streamlined and that even interns, who make up the bulk of their workforce, can do it now. Kevin Xu has confirmed this, too. But the first time I heard it, I genuinely thought they were just flexing. But the reality is that many of these Chinese labs have very small teams that are relying on smart architecture and quietly upending the traditional playbook of how to train a model. The training itself is streamlined. It’s finding the next breakthroughs in efficiency and capability that requires out-of-the-box thinking, and that’s truly within the capabilities of a much smaller group of people.

‘clean’ and ‘responsible’

The next paradigm shift may still be in the hands of a few people, like one researcher at Microsoft told me, “There are a few of them, where the industry just orbits around them.” But where we are now is a mature system.

Throughout the interview, he kept emphasizing “干净” and “责任'’ which means clean and responsible, to the point that even got the interviewer to ask for clarification. Paraphrasing what he said, he explained, the difficulty is that doing the simple things cleanly, at scale, inside a giant system, is very different from academic research.

In academia, you care about your own results. You want the cleanest paper, the best individual contribution, the most impressive result you can claim. But in a frontier lab, the unit is no longer the individual paper. The unit is the whole model training system.

A researcher can find a method that looks amazing on a small experiment, but if it only works because of a better data mix, more sampling FLOPs, different constraints, or something that will not transfer to the real production run, then it may not actually help the company. In that case, the honest thing is to say: this is interesting, but it is not useful enough for the big system.

But that is a very difficult task, because it goes against human nature. Everyone wants credit. Everyone wants the result that looks good. Everyone wants to say, “this was my contribution.” But in scaled AI research, personal heroism can easily become system-level noise.

He said that a good frontier-lab researcher is not just someone who can make a metric go up. It is someone who understands whether the metric going up actually matters and what the limitations are, and has the integrity to stay honest.

Why coding worked first

The second big shift Yao described is that the game has moved from “can we make the model better” to “what should we build with the model.” And he said something that honestly surprised me: among OpenAI, Anthropic, and Gemini, no one is really worried that they can’t catch up on model capability, they’re all somewhat aware of what to expect next. Rather, they have committed to different use-cases, and what everyone is trying to figure out now is what’s the next valuable task? Can it be well-defined? Can you build the right training environment and feedback signal around it?

Coding is the perfect example, and it's what almost every lab ushered to do in its herd-like behavior. It didn’t work just because models got smarter. It worked because the task itself is unusually trainable. The reward signal is clean (code runs or it doesn’t, tests pass or they don’t). GitHub provides a vast repository of high-quality data. His point is that good programmers have surprisingly aligned tastes, and they broadly agree on what clean structure, good abstraction, and maintainability look like, but what makes something a new breakthrough is through quite literally trial and error, and sometimes a bit of serendipity.

That’s where Claude Code, Codex, and Cursor come in. The task of coding is a combination of good data, clean feedback, professional users who are willing to pay, and a product workflow where agents can do economically valuable work. What actually made it work was model, plus harness, plus workflow, plus feedback loop.

LLM business is looking more like a cloud provider model

‘Base models are IaaS. Agent harnesses are PaaS. AI apps are SaaS.’ Okay, so this is the framework from the interview that I think is potentially most useful for investors.

If Yao is right that frontier labs will eventually have similar model capabilities, with differences more about timing, implementation, taste, and productization than about some fundamental intelligence gap, then the base model business starts to look less like a pure technology monopoly and more like cloud infrastructure.

Think about it. Cloud is also extremely capital-intensive. Only a few companies can do it on a global scale. But AWS doesn’t win because it runs “cloud” in some magical way that Azure or GCP fundamentally can’t understand. The moat is elsewhere entirely: scale, cost structure, custom chips, data centers, enterprise distribution, developer mindshare, and all the PaaS and SaaS layers built on top over time. (You can apply that same thinking to ByteDance, Alibaba, Tencent, Huawei, and the SOE cloud providers) I wonder what cloud expert think about this.

I think the LLM world could end up looking similar, where base models become the new IaaS. Agent harnesses (Claude Code, Codex, the orchestration layer) become PaaS. AI applications built on top become SaaS. The analogy isn’t perfect, but you get my point.

If the base model becomes infrastructure, then selling raw tokens is not a great business by itself, and Yao basically said this. API token resale becomes a price-war business, which is why I’ve been saying I’m not sure what the long-term strategy for Chinese labs is if they keep shipping 'nearly as good but very cheap models’, it feels like a short-term solution, and thus why the hyperscalers (if they can catch up in model capabilities or swallow the labs) then would have a huge advantage as they have the full stack. Now, I should flag here that Yao works at Google DeepMind, and this framing happens to favor Google’s structural position pretty conveniently. If inference becomes a commodity utility, having your own TPUs and a lower cost structure lets Google profit where others can’t.

For OpenAI and Anthropic, the strategic question is different. They can’t be model API companies forever. They need to move up the stack. That’s why memory, agents, workflow orchestration, and enterprise integration all matter. In the cloud, the money didn’t stop at compute; it moved into databases, observability, security, identity, developer tools, and business applications. In AI, I think the same thing happens. The model layer may be huge, but it may not be where all the margin stays.

What Yao is working on now

He said he’s focused on two things, and both are worth paying attention to.

The first is AI closing the loop on AI research itself. Models already write most of the code for researchers (Yao said a conservative estimate is 90% of his code is model-generated, and for experiment implementation, his speed is something like 20 to 50x what it was a year and a half ago, which is kind of wild). But the loop isn’t fully closed yet. The next step is for AI to define the problem, design the experiment, write the code, run it, evaluate the result, figure out what worked and what didn’t, propose the next hypothesis, adjust the experiment, and run it again. The full research loop. Yao thinks this can start becoming real in the next 6 to 12 months as the entire experiment cycle compresses. It becomes a recursive loop, much like Anthropic CFO Krishna Rao was describing on Invest Like the Best, too, and that happens, the pace of everything else accelerates, too.

The second is long-horizon memory. Or as he put it, train with finite context, use as infinite context. Everyone wants AI workers, but a real digital worker needs memory. It needs to keep track of long-running tasks, know what to remember and what to forget, retrieve the right information, and maintain state over time. Without that, a chatbot can never actually become a worker. Yao’s point is that humans don’t have infinite context either, as we forget most things, but we remember important things, at least we try to. So maybe the answer isn’t literally infinite context windows. Maybe it’s context management, selective memory, and retrieval capabilities, and actually knowing what to ‘forget’. I actually think this is one of the most commercially important ideas in the whole interview, because if it gets solved, you’re suddenly talking about something much closer to an actual digital employee than what we have seen commercially available so far.

In terms of model progress, he sees no end to progress and believes that features that feel differentiated today may become default model behavior tomorrow. The question for every app company becomes: what do you own that compounds faster than the model improves? How do ‘wrappers’ survive? I think maybe vertical industry knowledge and know-how is still something defensible here.

On China

On China, he prefaced a few times saying that he himself has not worked within a Chinese lab, so it’s all observations but not from direct experience. But surely with how small the circles are, he knows what’s up, or at least is quite tuned in. Yao’s view aligns with much of what I’ve been thinking (and writing about). The model gap has narrowed, but China’s compute constraints may force genuinely different methods. Less compute means Chinese labs have to be more efficient. They have to distill, reuse, open-source, optimize, and move faster through product loops.

Yao said that there is ‘dumb’ distillation and ‘smart’ distillation. ‘Dumb distillation’ is just taking another model’s output and training on it. That’s not a moat. It may even be a sign the lab doesn’t really know what it wants. ‘Smart distillation’ is more interesting: using another model as a judge, having your own model generate answers while another evaluates them, combining multiple models into a broader synthetic data and evaluation pipeline. That starts to look like real multi-model training, and Yao even suggested Chinese labs could become pioneers here.

Not all distillation is the same, and this part was pretty funny because Yao apparently named specific labs doing each kind, but the names were beeped out in the podcast. However, he did say Doubao is not distilling for a fact - hence why everyone seems to revere them secretly. If anyone listening caught who he was referring to, please let me know.

The other big China-US difference, as always, is product. The US has this massive enterprise productivity profit pool, so everyone naturally focuses on coding, workflow, and professional software. The math is quite clean in that sense. I help you write code, it costs me $150, and I charge you $200.

China has always been better at complicated consumer products that monetize indirectly. The product might look free and unprovocative. But then it monetizes through ads, commerce, games, live streaming, payments, recommendations, or some loop that only makes sense once it’s already spinning. If you’ve followed AI Proem, you know this is something I come back to a lot, because I genuinely think China’s AI consumer path is going to surprise people. The US may produce the best enterprise agents. China may produce the first weird AI superapp that looks like a toy until it suddenly isn’t.

Share

In other news

Alibaba reported its March Quarter 2026 earnings yesterday, with cloud growth as the main highlight, though investors are questioning whether the heavy AI spend has a clear strategic path forward.

Business highlights:

• Cloud Intelligence Group’s external revenue growth accelerated to 40% this quarter as enterprise adoption of AI services deepened.

• AI-related product revenue delivered triple-digit year-over-year growth for the eleventh consecutive quarter and accounted for 30% of Cloud’s external revenue.

• Customer management revenue grew 8% year-over-year on a like-for-like basis.

We also touched on the fact that Alibaba opened up Qwen app to Taobao this past week as part of its push into agentic shopping, but it’s still unproven whether this becomes a sticky habit or fades as a fad. My personal view: utility-purpose goods may well move to agents, but for discretionary spending, the choosing and selecting is part of the experience and the joy.

Tencent just announced that users can forward chat history into Yuanbao, and users are again asking why Yuanbao isn’t embedded directly into the WeChat interface. The answer remains: security.

Beyond that, I’m watching Alibaba’s Accio for AI merchant support, and Tencent’s WeChat Mini Programs. There are now reportedly AI-powered features that help merchants build mini programs within WeChat (think Cursor-lite), letting businesses easily plug into WeChat Pay, marketing, and distribution.

China (AI) visits in vogue

And on China, everyone seems to be in China these days. With the obvious Trump visit to Beijing, AFP interviewed me about my thoughts on how AI will fit into the conversations.

The thoughtful wrote a marvelous analysis of the state of physical AI and robotics in China after a week-long tour of factories and companies. I’ve been brewing on a piece on physical AI for a while, too, but this is a banger to showcase the reality of where we are at, which is nowhere close to mass consumer adoption yet. On the industrial side, robotics have been making more strides in adoption, I believe.

The @SAIL team seems to have had a fruitful China trip and I was glad to see ’s article noting that partly why the Chinese public doesn’t seem to be as anti-AI is because the labor force makeup is different, knowledge workers simply make up a much smaller percentage of the Chinese workforce in comparison. I think that’s an acute observation and something I’ve been sharing too. ’s article gave a rundown of the cultural differences among the labs, which we’ve also touched on in the deep dives but this was an updated fun read. ’s sharp economic analysis breaks down the potential backfired strategy of the export controls, stating that US export controls have left Chinese AI labs roughly 2-3 years behind in deployed compute (and an estimated 8x gap in capacity), yet Chinese frontier models trail the US by only 6-8 months on benchmarks. That implies Chinese labs are extracting 4-7x more capability per unit of compute than naive scaling would predict.

That efficiency edge is flowing through to inference economics: DeepSeek, Kimi, GLM, and Qwen are priced 3-28x below comparable US models while reportedly maintaining 50-70% gross margins, and they dominate the open-source and on-device tier. The implicit thesis is that the controls have backfired strategically, and by forcing ruthless efficiency, they’re cultivating exactly the capabilities (cheap inference, small-model performance, algorithmic ingenuity) that will matter most as AI scales. A very interesting take.

Last but not least, came away with some humbling observations about the Chinese AI ecosystem and how maybe open source, once again, proves to be the way to propel an industry forward. In his words, “the Chinese system is designed around quickly learning from your peers and avoiding double-spending research compute — or infra effort” as the labs compound on top of each other’s efforts.

He added, “The Chinese labs, through incredibly thorough technical reports and intentional knowledge sharing across labs, effectively are de-risking ideas for their peer companies to not necessarily need to invest as many resources in.” When we talked casually offline, he said that he also felt some sense of what I wrote here about how DeepSeek is beginning to serve the whole sector as a base technological foundation. So I’m very excited to have him on AI Proem’s Differentiated Understanding podcast next week to unpack more. So stay tuned!

Oh and two other good reads this week that were insightful/ enjoyable: Viola Zhou’s The Chinese whiz kids of Silicon Valley in Rest of World and Yi-ling Liu’s op-ed The Shared Feeling of Being Harvested by the Future in the NYT.

Next week, I’m in Singapore speaking at SocGen’s annual macro investor event, so no deep dives will likely be published. Thanks for your understanding. Toodaloo~

The Korean startup capital funding landscape is shaped by three forces: Chaebols (Samsung, SK, Hyundai), the government, and VC firms. CVCs from Chaebols tend to reinforce existing semiconductor and hardware value chains rather than explore tangential innovation. To counter this, the Korean government has become a dominant LP through initiatives like “Everybody’s Entrepreneurship,” injecting capital to encourage novice founders. On sovereign AI, he believes the government’s push is less about global dominance and more about securing sensitive areas like finance and defense, though he warns that domestically-built software has historically struggled to scale beyond Korea.

Ethan is shifting focus from purely domestic champions to founders with global ambition but local execution, often Koreans educated abroad who return dissatisfied with traditional jobs. He wants to back ventures that change the world, not just build another food delivery app. He also recognizes key opportunity areas, including defense tech, K-beauty, fashion, and mental health, as society adopts AI at scale.

Subscribe now

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, founders, and product managers.

For more information on the podcast series, see here.

To find the previous episodes of Differentiated Understanding, see here.

Chapters

00:00 Introduction to Ethan Cho and His Journey

02:47 Korea’s Role in the Global AI Supply Chain

05:24 Cultural Attitudes Towards AI in South Korea

11:06 Government Initiatives and Sovereign AI

16:37 The Future of Commerce and AI Integration

28:03 Consumer Behavior and AI Adoption

28:43 Enterprise AI Solutions in Banking and Manufacturing

33:39 Investing in Founders: The New Generation of Entrepreneurs

39:39 Korea’s Future Exports: AI and Beyond

41:41 K-Beauty and K-Fashion: Cultural Exports

45:15 The Future of Mental Health in the AI Era

49:53 The Limitations of AI and Human Experience

AI- generated Transcript

Grace Shao (00:00)

As mentioned, our guest today is Ethan Cho. He has been active in the Korean VC space for over a decade with experience in the venture investing arms at Qualcomm, Google, Samsung and more. Now as a partner at the ventures, he leads a team focused on finding and nurturing the next generation of AI native startups. Ethan, thank you so much for joining us. So good to have you.

Ethan Cho (00:18)

Thank you, Grace. I’m very excited to be on the show and I would love to discuss with you more in detail.

Grace Shao (00:24)

Yeah, to start with, tell us about yourself. Tell us about venture investing in South Korea and the firm’s background.

Ethan Cho (00:31)

Sure. So I was born in Korea. I kind of moved internationally quite a bit. I moved to England when I was kid, when I was four years old. That was where I first learned my English, lived in England about four years, came back to Korea, then moved to Hungary, lived in Budapest for a year, came back to Korea again, spent the next 20 years in Korea, moved to the States, lived in New York for ⁓ my business school years and worked there for another year, came back to Korea after then. So I’ve been in and out of the country quite a bit.

I loved startup investment very early in my career, so I wanted to move towards startup investment. I actually started out as a hedge fund analyst right after business school, but I quickly found out that I’m more interested in finding good companies and good stocks. So then I moved towards the private side, started with Samsung, and moved to Qualcomm, et cetera, et cetera.

What fascinates me about Korean startups and startups in general is that everybody’s trying to change the world. I’m just such an honor to be a part of that and talking to entrepreneurs on a daily basis really excites me.

Grace Shao (01:34)

Awesome, I think it’s really interesting because you’ll definitely bring a very international perspective and not only just the Korean perspective and also kind of understand, you know, where a lot of our listeners are coming from as well. You have, you know, you have exposure to UK exposure to Europe, exposure to US. I think I want to ask you quickly, because you’ve actually worked in the public sector as in public investing, it’s kind of interesting because right now, obviously, the frenzy and the, you know, the global interest right now.

Ethan Cho (01:57)

Mm.

Grace Shao (01:59)

is in a lot of the big semi providers in South Korea, you know, focused on infrastructure layer that are public listed. At a high level, how should we think about the Korea’s role in the global AI supply chain? And then of course, we’ll shift gears into talking about the startup scene that you’re passionate about.

Ethan Cho (02:02)

Yep. I think that’s a great question. think one of the very obvious factors of AI is memory because you can only use AI based on what

you or the agent knows about you or any company. So because of that, the demand for memory is exponentially increasing. I think that’s definitely a blessing for the career semi-players and also the current industry as a whole. But at the same time, think nature has always found a way to become more efficient. So the capacity or the demand constraint will not be

existing forever. There’s going to be significant improvements such as Moore’s Law. There’s always an innovative solution that comes out every other year. So I think there’s definitely going to be a lot of exciting opportunities down the road, but it always evolves. So it’s going to be very different next year. It’s not going to be HBMs anymore. I think it’s going to be something else going down the road. We’ll have to see, but I think the trend is here, but the trend itself will also keep evolving down the road.

Grace Shao (03:15)

Awesome. So, you know, if we have a really candid assessment of what’s happening in Korea right now, what’s genuinely really strong do you think? Like the chips are strong. ⁓ What do you think that weaker maybe compared to, you know, China and the US, you know, from outsider lens, is it really the LLM labs kind of or is it diffusion? What’s happening?

Ethan Cho (03:24)

Hmm.

Yeah, I mean, it’s a complex situation. I think one thing that a lot of people quote and one kind of fate that we cannot deny is that we have a very low fertility rate. So the birth rate is decreasing very fast. We have a very rapidly aging population. A lot of people think of this just as a curse. I think there is an aspect to it that it’s a blessing in disguise because we are one of the countries that most desperately needs AI and robotics.

And because of that, I think we will be one of the more adapting or welcoming countries for AI and robotics. If we go back to the early 2000s, we were one of the countries that adapted most rapidly to high speed internet as well as mobile technology.

because we had to, we talk a lot on the phone, obviously. So because of that kind of demographic instinct, we were one of the much faster adapters to that technology. I think that’s gonna repeat in AI and robotics. As you mentioned, I think that AI and robotics is definitely not, we’re not the strongest when it comes to AI robotics in the world.

But in terms of adapting and using it for actual use cases, we may be one of the very strong countries. So I think there’s a lot of challenges and opportunities ahead of us.

Grace Shao (04:50)

That’s really interesting. think you hit something that like, you know, people are starting to pick up in the West, which is like in East Asia in general, the embrace of technology is a lot more optimistic. Some come from very realistic reasons. Like you mentioned, whether it’s in China or Japan or Korea, there is a potential labor shortage that’s coming to the next generation, right? But not only so, I think just in terms of culture and social sentiment also feels that way. So if you have to give like a high level kind of assessment on

You know, just the cultural attitude and political attitude towards AI, what does it feel like on the ground in South Korea?

Ethan Cho (05:24)

I think AI itself, I think people think of AI in different forms, obviously. I think as far as I know, China thinks of AI closer to robotics. The US thinks of it as, I don’t know, maybe a chatbot or something that they use for the industrial usage. In Korea, as far as I’ve experienced, I think it’s more about becoming a personalized kind of agent, not agent.

per se in terms of doing purchasing or actual daily tasks. We always had kind of chat bots, especially for instance, for our financial system or the banking system, we’ve always used CS bots very frequently. So we’re very used to it. So I guess there’s less resistance when it comes to adopting or adapting bots or AI featured functionality, especially in the B2C area. So although...

When we say AI just by the name, it could sound creepy. I think it’s already very well embedded in the Korean startup ecosystem and the overall society as well.

Grace Shao (06:27)

That’s interesting. So you did kind of touch on one thing, you know, the Chinese view this way, the Americans view that way, you know, definitely there’s a bit of a difference in terms of how maybe AI is being seen as whether it’s a political agenda or economic ⁓ aggregator or, you know, how it’s being diffused to the seaside. So in that sense, ⁓ my question is, is Korea building domestic AI champions or is it more right now kind of working around, you know, working on top of U.S. frontier models?

or leveraging a lot of the Chinese open source models. Like how do we understand that in the ecosystem?

Ethan Cho (07:01)

Honestly, this is a personal kind of statement and my personal observation. I think it’s all of the above. I think we were going to talk about this anyways, but the sovereign models is something that the Korean government really wants. I kind of understand it in a way. The SOTA models, the state of the art models are obviously the ones that the enterprises want to use.

But at the same time, think a lot more people and developers are looking into Chinese open source models, especially with the recent changes in cloud code and Gemini and everybody, who are basically hiking their token prices. It’s getting more and more expensive to actually do recurring work. And at the same time, think, including myself, a lot of developers are quickly finding out that the fine tuning of the models can only be achieved by very,

⁓ almost redundant loop work, which can only be achieved through open sources if it is to meet economic sense. So right now, for instance, if we use a certain American LLM model to do these hundreds and thousands of ⁓ very repeated work, that costs a lot. So from an individual standpoint, that’s not really easy to achieve. So I think everybody’s trying to find that sweet spot of mixing those three models.

Grace Shao (08:16)

That makes a lot of sense. I think for stars, especially the ones that you work with, know, the economic driver is probably one of the biggest reasons why they choose what. So I do want to save the sovereign AI kind of piece for a bit later to help our listeners understand, you know, the create ecosystem a bit better. We all hear about Chibbles. We hear about, you know, obviously the big tech like the Samsuns and the whatnots in Eskihainix right now that are getting a lot of attention, right? Help us just even understand how these different companies and the startups, how they work together. Because for example, in China, a lot of that big tech are actually the incubators and initial investors of even the startups. So even the leading LLM labs, they actually have taken 10 cent Alibaba money. In the US, it does feel a bit different. There’s vast amount of venture capital money that are kind of funding the current growth rate of OpenAI and anthropics of the world. How does it?

Ethan Cho (08:54)

Mm-hmm.

Grace Shao (09:09)

ecosystem work in South Korea.

Ethan Cho (09:11)

So I think there was ⁓ quite a few phases of evolution. when the startups were really founded, that was, I don’t know, that was like late 90s. Those were very purely internet domains, internet online communities. Like that had not a lot to do with the Chebals. But then came mobile technology and everybody was starting to invent something on mobile. That quickly got the...

Interest from the big Chebos, but actually as far as I know there were some Interests in very early on in neighbor and cacao by all these like really large companies in Korea But they never actually fully understood what that what that was and they kind of let them grow Which was a blessing for us at the end of the day. So neighbor and cacao was you know established and they grew like crazy after that the

the big companies quickly found out that, we need that DNA of innovation. They started to set up their own VC firms. They started to set up their own accelerators and everything. But as a typical CVC does, they inject a lot of money into their interest area, but not so much in let’s say, tangential areas. So because of that, there are definitely a strong value chain around semiconductors, for instance. But that kind of reinforces the

already existing ecosystem of the chaebols, which is not exactly what the startups are intended to do. So there’s kind of pros and cons there. And on top of that, after the capital was kind of concentrated into that value chain, the government kind of now is more active in kind of leading investments. large chunk of investments in Korea is led by the government.

by the mother fund or fund of funds of Korea injecting money into the ecosystem and the other funds matching to that. So there’s a layer of chaebols, there’s the government and also the capital firms who are also ⁓ acting as LPs for the Korean startup ecosystem.

Grace Shao (11:06)

Yeah, that’s a perfect segue into understanding, you know, the government’s play. So I visited Korea just recently, I think last November, and, you know, it seems like there’s a huge policy push as well in incorporating AI into the everyday everything. And, you know, it’s top down driven. And like you said, there’s capital also injection. So how do we understand the government’s current priorities in terms of embracing AI? How do we understand sovereign AI and

what kind of role it plays in the economy South Korea going forward.

Ethan Cho (11:38)

⁓ I think the government is definitely ⁓ making a very interesting and important bet. So there’s this huge initiative called Everybody’s Entrepreneurship, loosely translated into English. The government is actually injecting a lot of money into the ecosystem by giving money to...

people who want to become entrepreneurs, who are novice entrepreneurs, first time entrepreneurs. I think it’s a good thing that a lot of people are trying out their ideas at the end of the day. The downside honestly is that entrepreneurship isn’t for everybody. So there’s gonna be people who learn their lessons the hard way, but still I think all in all, it’s gonna be a positive impact on the overall ecosystem. I think...

⁓ The AI drive is definitely very serious for the government. As we mentioned earlier, there’s a labor shortage coming in. I think ⁓ East Asia most of the time has a little bit of issue with immigrants. We don’t shy away from immigration, but I think traditionally we don’t have the most welcoming immigration system compared to the States, for instance. So we’re trying to buy some time there, I think.

And I think because the strongest point of the industry, as we also mentioned earlier, is semiconductor and hardware and technology, we want to build upon that. And because of that, I think AI seems to be a very interesting and promising area for the government and Korea as a country overall.

Grace Shao (13:03)

How do we understand sovereign AI though? Like what is the, I guess, reason for like, you know, maybe the non too large company, sorry, too large economies to really start pursuing this? Because we’re seeing this kind of rhetoric in the Middle East as well. You know, a lot of the local governments are really pushing sovereign AI. South Korea for sure has been openly talking about this. think, you know, a lot of European nations are also thinking about this. Is this just for, I guess, owning?

Ethan Cho (13:06)

Hmm.

Grace Shao (13:30)

the future infrastructure or how do we understand this?

Ethan Cho (13:33)

I think that’s one point. I think owning the future infrastructure is one. But I think if people are realistic, think we don’t want the world, we don’t hope the world is going to use our own sovereign AI. I don’t think that’s the case. What I’m expecting or I believe that the government people are wanting is that to use sovereign AI in very sensitive areas, such as our financial backbone, for instance, or

You know, because Korea is technically set or on the defense part, maybe we’ll use that for that purpose specifically. I think in the early days when everybody in Korea started to talk about sovereign AI, I was actually less persuaded. The problem is, or the status is, as we see all these leaks all over the place, like even for the top, you know, bleeding edge,

builders like Anthropic or OpenAI, there’s always issues here and there. And it kind of shows. I’m not saying that sovereign AI is going to be perfect either. They’re going to have issues too. But if a foreigner comes, a foreign entity comes in and kind of screws up an operation, that kind of blame and what

what something domestic spills over. There’s going to be a different kind of anxiety in the society, I guess. So that’s maybe the angle that they’re kind of anticipating. But I mean, I am worried a little bit too, because I’ve seen multiple cases of software built in Korea domestically, which has never been successful to Korea. And it just has been kind of a wanted wonder just in Korea. I just hope that doesn’t repeat. But we’ll have to see.

Grace Shao (15:06)

Actually, this is not completely rated to AI, but just on that note, why do you think a lot of the times like Korean companies are huge? Like you just mentioned Naver and like Kakao or like, you know, Japan, and LINE and China that we chat when not like these internet companies never really go abroad. That’s just intellectual curious question just on the topic.

Ethan Cho (15:16)

Mm.

I’m a kind of linguistics buff. So I actually think the reason is in the language. The language that we speak is just different from English. because of that, think the like Naver, Kakao, WeChat and Line are all basically rooted in language. And because of that, that’s just universally different from WhatsApp, for instance. Like it’s not language per se, but if you look at WhatsApp, how they control their UI UX, for me at least, is very boring.

Grace Shao (15:27)

Mm.

Ethan Cho (15:52)

⁓ I would prefer a cacao or lime or WeChat much over WhatsApp if I could choose without being specific in which geography. So I think there’s a cultural preference, a very strong cultural preference that really is hard to translate across territories.

Grace Shao (16:09)

Okay, that’s an interesting take. Yeah, because I think it’s interesting because like, it’s basically the West has this one or even like, you know, Africa based Southeast Asia, they all fall under the American big tech kind of umbrella. And then Korea, Japan and China’s have such strong domestic players. Actually, on the note on you know, the consumer side of things, what are some interesting trends you’re seeing out of South Korea in terms of consumer AI right now? What are some companies you’re investing in that are in the consumer AI space?

Ethan Cho (16:16)

Yeah.

Consumer AI, think, is still at a very early stage of growing. I think right now the most used cases that I see with my bare eyes are actually foreign tourists coming to Korea, visiting like big K-Beauty.

department stores like Olive Young, and they go and get their skin scanned and they analyze it with AI and give recommendations to basically ads. But still, I think that’s a very clever way to scientifically analyze the customer demand. I think a lot of players, and I see a lot of players trying to replicate that into basically recommendation engines. Personally, I think that’s

clever, but it’s not good enough. It has to get better. The trade-off there is obviously privacy. So if you want a Uber personalized recommendation, you have to somehow yield on the privacy part. think we’re still not clear on where is the kind of safety line. So I think they’re still kind of struggling towards that. We have invested mostly these days in consumer brands because how

⁓ I think of it at least, is regardless of which AI becomes the winner or winners, I think as long as we have the best product in our portfolio, if the AI is clever enough, it will choose that product. So before actually deciding which AI algorithm will actually win the war, think we’re trying to get hold of the monuments before the, know, who,

before we decide who becomes the winner of the war altogether.

Grace Shao (18:12)

So you’re looking at brands as in like, like retail brands. What are you looking at? Like, okay.

Ethan Cho (18:16)

Yes, yes, for now,

yes. But at the same time on the kind of the hardcore AI part, we also have invested in sovereign AI companies like Trillium Labs, which actually develops SLMs instead of LLMs. We’ve also invested in a drone company called Bone AI, which does physical AI using drones and they’re targeting the Korean big defense industry. So that’s kind of the

more of the hardcore AI part that we’re looking into. We’re still waiting for that sweet spot where consumer meets AI. I think that’s still kind of in their very early stage in Korea.

Grace Shao (18:46)

I see. I just want to say it’s so funny you used the Olive Young example, it’s really topical. So this is really a bit of a rant, but my friends and I were saying we need to go to South Korea to do the color palette, right? And all, you know, all the girls are like raging about this right now. I literally asked Claude Coe today to do it for me and it gave me the whole like color palette assessment. I was like, wow, I just saved myself a flight and a trip to South Korea. So definitely can see like there are consumer uses in that end, but I guess what is the monetization from that bud, right?

Grace Shao (19:22)

So that I can see how that will be hard to invest in that space. On the consumer end, know, again, I read headlines about South Korea, right? And, you know, I hear about companion bots being really big. I know actually even in China, they were group bots. They were like so-called boyfriend, girlfriend bots. And to your point, you know, South Korea, China, like even like a lot of East Asian countries are in are all faced with this issue where there’s mass urbanization.

Grace Shao (19:48)

loneliness issue everyone is like, you know faced with evolution and competition so they don’t have human companion and Do you see this as a trend and do you see this as something that potentially would not be actually within the cacau’s and the lines of the world that could be a spun-off on its own and to fall off of that I was just even just kind of thinking because Korea has so much IP right now in obviously k-pop and k-drama

Ethan Cho (19:57)

Hmm.

Grace Shao (20:16)

would that potentially be a vertical where they can tap into basically creating, you know, like companion bots, based on existing celebrities.

Ethan Cho (20:25)

I mean, I think yes on both is the short answer. I think the boyfriend bots, girlfriend bots are very popular in Korea. I think my son is also using one. I haven’t talked about that openly, but I think so. And yes, I think that Kakao and Neighbor would be very cautious about adopting that technology into their existing platform. There can be some...

many opportunities to abuse that. People, as you know, once these bots are online, the first thing that everybody tries to do is abuse it one way or another. So I think a cow and neighbor would probably shy away from that and go into commerce, which is always what they wanted. We’re seeing more specialized startups that are just doing this boyfriend, girlfriend bots.

like some are more adult focused, some are more teenager focused. So there’s definitely kind of a breed that’s coming out of that. On the K-pop and K-drama bots, I think that’s something that a lot of companies have worked on for quite a while. For instance, like Weverse, is the, it’s the entertainment company that basically ⁓ relates to all K-pop stars. They have their own like personas.

So they actually provide not just only conversations on bots. I think they also give artificial voice calls. They’re already there. So you can have a conversation with your favorite star. It’s just not realistic enough yet. But I think they’re getting there. So ⁓ that’s definitely already happening. I think on the IP side, personally, think it’s more about how can you make these into really long-lasting legacies? Even for BTS and like...

girls generation, which are the you know, the idols of the day. I think they’ve only been around for 10, 20 years. Like, can we make this into like decades long, right? Like a legend, like can we actually make this into something that goes through generations, not just decades? That’s a big homework for us to figure out and make them kind of timeless.

Grace Shao (22:24)

No, I totally see that. It feels like a Black Mirror episode with the Miley Cyrus ⁓ kind of fake doll as well. But I think to your point, there’s the IP issue, there’s obviously the security issue. There’s the psychosis issue. This is like a much bigger issue. I think that we require regulators to work with businesses, right? I want to kind of move our lens to the enterprise side of things. You you mentioned just now like Naver and Kakao were

Grace Shao (22:50)

looking into maybe going to agentic AI and maybe even to commerce. Are we looking at something like what Alibaba is trying to do where you have agenda commerce through a one entry point, you you interface with a chat bot next thing you know, like a bubble teas at your door. Like, is that the future of commerce you think or what are we talking about here?

Ethan Cho (23:11)

I think, I think, neighbor and Kakao are both in fierce competition with coupon coupon is the dominant e-commerce player in Korea. ⁓ I think it’s a real headache for both of them because coupon was kind of non-existent. They didn’t have a lot of user interface and neighbor and Kakao were kind of self satisfied that they dominated the user interface. But, here comes coupon and they just basically just crushed every.

aspect of e-commerce and is by far the number one player in Korea. So that’s something that Naver and Kakao are trying to battle. The only difference, as far as I can see, that they can make is real-time purchases. If you want milk at your home the next day, coupang is much easier and much better. That’s kind of a fact. But for Naver and Kakao, because they have basically 24-7 access to your daily life, if they can actually monetize on that, I think

That’s their way to go. The competition there is also not non-existent. That’s a problem. YouTube’s there. TikTok’s coming along in Korea. TikTok’s still small in Korea, but it’s growing rapidly. that space is also... I think people think... Koreans are just so big YouTube fans. The YouTube dominance is so... They used to. Now they’re getting more used to short forms now.

Grace Shao (24:15)

Why is it small? Curious. Why is it small?

So they like long form.

Ethan Cho (24:30)

And the TikTok trend is definitely coming along. But I think most YouTubers, so-called YouTubers in Korea, are long-form originated. So the trend is changing now. So yeah, it’s a little bit slower to adopt. yeah, sure. Oh, yeah.

Grace Shao (24:40)

And I wanted some context. So coupon is like an Amazon or like it sounds like a DoorDash

and like how do we understand this just for our American audience or Western audience?

Ethan Cho (24:51)

Yeah,

Coupang is, they actually literally say that they want to be the Amazon of Korea. So they are the Amazon equivalent in Korea. Their main business is e-commerce. They have Coupang Eats, which is DoorDash or Uber Eats. They have Coupang Play, which is the Amazon Prime. So they’re basically, Coupang people would hate me saying this, but it’s kind of like the Amazon replica in Korea.

But they’re doing a fascinating job. Their killer feature is next day dawn delivery. So if you deliver, if you place your order by midnight, they’ll get the item to your doorstep before 5 a.m. So it’s marvelous for. It’s not just groceries. Yeah, so they’re very good at demand expectations. So they have a lot of warehouses in Korea. So they fulfill them in advance so that they can basically distribute almost within like four or five

Grace Shao (25:29)

It’s not just groceries, it could be anything.

Ethan Cho (25:43)

hours window, which is also honestly possible because Korea is not so big as a country.

Grace Shao (25:49)

But it sounds kind of like almost a JD.com business model as well. It’s like, but more high, high, more expensive. Excuse me.

Ethan Cho (25:52)

True. Yeah, yeah, I think I think that’s a fair. Yeah, that’s a fair comparison. Yes.

Grace Shao (25:57)

It’s a bit more expensive, right?

Yeah, so I think looking at that, then, you know, there’s also this rumor, or I guess it’s actually been verified that South Korea was, fact, OpenAI’s largest enterprise market outside of the US, which is crazy, because like you just said, South Korea is not exactly that big of a country. Why is that? Who are the people buying up all these tokens?

Ethan Cho (26:12)

Hmm. I don’t have exact numbers, but when I first heard that I wasn’t too surprised because if you look at like Koreans are very used to buying tokens online. Like that’s why Korean gaming has been

or what used to be so big, especially in the mobile era, because people were just fine with buying items online, like purchasing it like crazy, which was kind of now it’s kind of standard, but like back in the days, like early in 2000s, like it was a very weird phenomenon if you look at it from a global standard. So because of that, think people are really, really fine with just buying tokens and buying memberships, which costs 100, $200. I think that’s kind of what

the base layer, so the willingness to pay the first layer. The second layer is that Koreans love to build things. Just trying to build things so much with their own hands is one tendency that we strongly have. So because of that, I think most of that building tokens go to my stock portfolio optimizer or kind of tools like that for personal use. But I think people just like to try out a lot of things that kind of led to that consumption.

Grace Shao (27:26)

Mm-hmm.

Ethan Cho (27:33)

On the B2B side, think as far as I know, because the head of OpenAI Korea used to be one of my kind of bosses at Google Korea, he was, well, OpenAI was very aggressive making contracts with Samsung and SK very early on. So I don’t know how many tokens they’re using, but just thinking about how many employees they have, if they struck a good deal on a B2B,

business, I think that would be a very significant portion of tokens being burned in Korea as well.

Grace Shao (28:03)

That’s pretty crazy. So basically you’re saying the the enterprise side, people have pretty strong connections and reasons to buy and the consumer side are just willing to shell out subscription dollars. That’s very different from, would say, the Chinese market where there’s just like not a lot of willingness to pay from consumers. And hence why we saw all the consumer apps in China were all free. So I actually want to ask on enterprise end. So what are we seeing people spend money on in terms of AI that

Ethan Cho (28:14)

Yep.

Grace Shao (28:30)

What are people trying to build? Are we looking at like also like on the enterprise and are they trying to solve co-pilot like solutions? Are they trying to serve customer service issues, manufacturing optimization? Like what are people really focused on?

Ethan Cho (28:43)

So just based on my experience with the companies, I think one thing is the banks are very serious about building the CS layer via AI. So they want to substitute a lot of that labor force into AI. I’m not sure whether that’s like how fast that can be optimized just because people are very demanding in Korea. know, when even if you use the traditional kind of phone CS, people

end up basically talking to people. They demand to talk to an actual person instead of going through the automated call. So we’ll have to see how the ROI comes out on that part. I know that there’s a lot of AI being used for the semiconductor processing and producing process, but that’s just not public information. So we really don’t know how much is being used there. So that’s on the enterprise side. On the consumer side, think because of everybody’s

⁓ entrepreneurship program that I mentioned earlier. I think there’s a lot of people that are trying to use a lot of cloud code, for instance, or codecs from OpenAI to build programs. There’s a lot of events actually held in Korea. Maybe every week there’s an event from OpenAI or Anthropic basically, which is like the cloud ambassadors night or the OpenAI something, something night. people, lot of... ⁓

the AI builders are actually encouraging Korean builders to use their own tools by giving out a lot of free tokens actually, like thousands of dollars are given out as tokens just to nudge them into building. So there’s gonna be a lot more activity in that space for sure in Korea. And hopefully there’s gonna be something that’s really interesting coming out from that.

Grace Shao (30:25)

That’s interesting. I did want to ask, you kind of mentioned this earlier that you guys even invested in a drone company. South Korea obviously has a very strong manufacturing sector, home appliances, phones, cars. How are we seeing this whole, we have generalized this whole sector kind of lean into AI? Are we seeing physical AI being prioritized? Are we going to see? more robotics coming out of South Korea. How do I understand that?

Ethan Cho (30:55)

I think there’s still some uncertainty there because of the all of the among all the Korean robotics companies, I think the most technologically advanced one is Boston Dynamics, but that’s not a Korean Korean company, to be honest, right? Because it used to be an American company acquired by Hyundai Motor Company. So there’s that. There are quite a few robotic startups that are starting in Korea.

Just because we have Samsung, Hynix, and Hyundai, think the manufacturing industry obviously is a great application area or a market to sell to. So we are seeing a lot of robotics company coming out from the university as well as startups. The big question here is will they scale? That’s kind of the pressing question. I mean, I think the companies, for instance, for Coupang,

which does all the logistics. They’re heavily using robotics just as Amazon does. So those robotics are already deployed or are being deployed. But for instance, humanoid robots, which China is leading the way, I think that’s still a long way to go for us. And we’re trying to figure out what the application should be. So one interesting example, I think China has this too, but...

We have all these little, really cute delivery robots going down the road and trying to get food to their neighbors. That’s an experiment that a lot of companies are running right now. We also have small police robots that are also running around just to do surveillance. I think it’s a cute initiative, but can this scale is going to be a big question for a lot of us. So I think this is also intertwined with

autonomous driving landscape in Korea, which is still kind of not there yet. So I think there’s going to be a lot more of this going forward.

Grace Shao (32:44)

Yeah, no, actually on that note, I just was in Shenzhen last week and I saw one of these like, you know, street sweeping robots per se, stuck in a puddle. And it’s like, to your point, like they look cute or like, you know, you have little robots delivering your phone charger in hotels, but they’re not actually that scalable. And I don’t actually know if they’re that cost effective is the issue, right? Because, you know, sometimes hiring a person to sweep the floor, frankly, in

Ethan Cho (32:48)

Hmm.

Grace Shao (33:12)

a market like China is actually not that costly compared to even deploying a robot like that and then having to, you know, maintain it. So I see your point. Okay, I think, you know, I want to shift our focus back to, you know, your bread and butter. And I really appreciate you patiently breaking down the ecosystem for me as an outsider who don’t understand South Korea that well. But as a venture capital investor right now in South Korea, what are your, I guess, most interested areas?

What kind of founders do you really want to invest in? And are you looking at the founders more or the companies more? Let’s start with that.

Ethan Cho (33:45)

I am looking for founders. I’m looking for founders because I think there’s been a evolution of generation or a change of generation that I’m seeing. I see a lot of Korean.

like in their 20s or their 30s who are educated abroad, come back to Korea and start working in Korea, not too happy about their job and trying to figure out what to do next. I just want those people to actually start something new and I want to kind of back them. I call that like global ambition, but local execution. I think that’s something that we need more.

Until now, as you know, all the companies that we’ve mentioned throughout this conversation, like Naver, Kakao, Coupang, they’re all basically really focused on the Korean market, which was, you it’s good. But still, as we all know, Korea is not the largest of the countries. And, you know, just doing business in Korea doesn’t mean a lot, especially as we move towards AI more and more. And because of that, I just want those...

⁓ kind of people who are ambitious to really change the world in a significant way, not just build the next chatbot or the next food delivery app, but something that kind of, you know, breaks around and just changes something very significantly. That’s something that I’m really looking for these days.

Grace Shao (35:00)

That’s really interesting. Do think that has anything to do with your upbringing, just being so internationally exposed?

Ethan Cho (35:05)

Maybe, actually, yeah. think, this is kind of another personal note. think Asians are really smart in a lot of settings, but we as Eastern Asians, were brought up to be kind of modest and humility was one of our very top priorities as we grew up. And because of that, we tend to be more humble in front of people. And as we know,

The Westerners, like this is not a great word maybe, but the Europeans or the Americans are much more aggressive in PR, but we tend to be more careful about that. Back in the days, that was great when we were just living amongst ourselves, but now as we go into the global market, PR is really important and having big ambitions like shoot for the stars, land and the moon is the way to go. But sometimes we just focus on what we have. I think that’s a healthy way of living, but.

For entrepreneurship, we really have to dream bigger dreams.

Grace Shao (36:02)

That’s really interesting. I think it’s some things that I’ve even really noticed within the just generating Chinese founders as well. It’s really different. Like you mentioned coupon. I think the founder was Harvard educated, but he returned to Korea focused only on the cream market, just like the last year. He’s like the JD.com Alibaba’s and the day these are Chinese market businesses. They have global footprint, but there no one’s thinking of them as a international business, right? At the core, they’re Chinese company. But if you look at the

Ethan Cho (36:18)

Yep.

Grace Shao (36:28)

whether it’s the LLM companies in China right now, or even some of the more consumer facing ones, or even the robotics ones in China, I kind of feel like there’s a shift in generational behavior. Exactly to your point, some of them are less educated than they’re not, but in general, people are not as modest. People are actually more, not in a bad way, but they’re much more open to doing PR for themselves. Well, not just PR, but actually flexing and going more ambitious, going global.

like we said, like Kimi and Minimax, whatnot, Jiu-Jitsu, they’re used globally, right? And they’re not shying away from it. I think that’s really interesting. That’s like a phenomenon across East Asia right now. So I think for us to understand, what are some, I guess, misunderstandings or things that foreigners who are trying to invest in Korea often...

you know, get wrong or not completely get correctly because, know, obviously there’s so much societal nuances. Well, in South Korea is a country where I find, like you said, it’s not not only not that immigration friendly, but actually in some ways a bit more closed off, Much like East Asia in general, like if you’re not from there, you don’t speak the language, don’t understand formality, especially South Korea has a lot of formalities. It’s really hard to do business, right? So how do what are things that you see that foreigners might be getting wrong that they could do better?

Ethan Cho (37:44)

Hmm, I think, well, I mean, first of all, think Koreans are just a lot of time. I wouldn’t say everybody, but a lot of Koreans are just shy. They’re, they’re friendly, but they’re shy. That’s kind of our kind of default mode. So, you know, if somebody comes to Korea and nobody wants to talk to you, that’s the norm. But once you try talking to any random Korean person, he or she will definitely help you out. That’s kind of the Korean kind of way. They’re being shy because they want to be polite. That’s kind of an Asian thing, right? So.

There’s that. think because Korea has been such a small country, think people, some people think of Korea as just being focused on that very regional kind of market. We’re not, obviously. Like we want to also go global, but we just didn’t have enough chance to actually show off that.

I mean, if you look at, for instance, the Koreans working in the States, they can show you what a Korean can do if they’re put in the right setting. So I think if you’re an investor and want to work with a Korean firm, think as long as you put the resources and the human talent in the right settings, they will perform. of course, I can’t guarantee everybody will, but in general, that’s how we’re formulated.

I think one interesting factoid that I also always kind of want to emphasize is I think China is also similar to this, but because we have this crazy, crazy education system that’s like overly competitive, although we have been really stressed out throughout our teenage years, that actually made us very, very competitive when we just, you we’re put in the right settings. Like we will strive to become number one in whichever setting that we are put into. So just.

like, you know, help us get to the right market and get to the right country or what right settings we will perform. So that’s, think, the expectation that you should kind of have for a lot of Koreans and Asians in general.

Grace Shao (39:39)

just like whoever can go through the national like university exams, like they have resilience. These buddies don’t like they don’t mess up. So I think on that note, I guess I want to ask what are what should we expect Korea to be exporting that if you’re saying that you want to back companies that are going global, you want to back ambitious internationally minded creates, what should we expect? Because no one expected it. Well, not no one. But a lot of people did not expect China to suddenly be exporting LLMs, right? As like one of their hottest new technology right now. I think a lot of times robotics maybe, EVs maybe were more in the expectation over the last five to 10 years because of the strength and the slow momentum it was gaining, right? But yeah, for Korea, what should we be looking at? Like you said, obviously hardware, chips, there were a lot of synergy there. You’re trying to build on top of that, but beyond that.

Ethan Cho (40:29)

I think the low hanging fruit or the easy pick is K beauty and K fashion. That’s definitely gonna come in the next ⁓ coming three to five years. I personally think there’s a lot of interesting angle in the Korea defense industry combined with AI because honestly speaking, there’s a lot of, how should I call this? Like confusion around the American diplomacy.

policy recently because of all these international tensions. And because Korea has always been at war technically with North Korea, I think there is a lot of advancement in Korea technology wise. We have the best semiconductor in the world. I think we are one of the most flexible countries when it comes to, are you going to use US LLMs versus Chinese LLMs? Like we can do both. think we are.

We see the pros and cons there. very flexible there, so we can optimize. I think because of that, the defense industry is not only growing very fast. I think it’s a very good place to kind of experiment the new warfare technology without going into actual war. And because of that, I think the Korean defense industry will benefit a lot from AI evolution.

Grace Shao (41:41)

I see. It’s an interesting area which I’m not like I’m not familiar with at all. But it’s like kind of like you said, it’s kind of one of those areas where you don’t really hope it being really used, right? ⁓ But it’s definitely a very hot space in terms of VC investment and in the US, especially with Palantir driving over the last couple years. Again, not really to AI, but I kind of want to double click on K beauty and K fashion. Why is it like what what is it that you know, over the last week? So

Ethan Cho (41:50)

Mm. Yeah.

Grace Shao (42:08)

My husband and were trying to talk about this very casually that day. We’re like, wow, we live in Hong Kong. In the 80s and 90s, everyone was obsessed with Hong Kong pop stars and Hong Kong movie stars across Asia and then even globally. They had all the kung fu shows and then all the police shows. And then in the early 2000s, we definitely had the Taiwan wave, the Taiwan pop stars coming out of East Asia. And even I was in Canada. I was growing up Canada and people loved Jay Chow, right?

Ethan Cho (42:29)

Mm.

Grace Shao (42:36)

Nowadays, obviously, it’s all about Blackpink, right? So how does this move around? Why is it going around? And how does one society kind of, I guess, nurture or incubate a global pop star? Is it tied to geopolitical reasons or economic reasons? Or do you think aesthetics?

Ethan Cho (42:58)

I don’t know exactly the reason if I knew I wouldn’t be working in V.C. I would be another producer. But I think there’s two reasons that I think is the biggest reasons. One is we have a massive farming system, you know. Koreans train boys and girls in their early teens.

Grace Shao (43:04)

Yeah.

Ethan Cho (43:18)

to become the K-pop stars and you have to go through vicious vicious competition to actually get there. So because of that, I think there’s so much talent that’s going through that pipeline, which is a blessing and occurs at the same time to society, obviously, I think so. But there’s that. The second part is I think Korea had a mix of...

American culture very early on because of the Korean War and the Korean forces, sorry, American forces staying in Korea. So if you look at Blackpink’s music, because you quoted Blackpink or even BTS, there’s a lot of African-American music, ⁓ like features within embedded in that music line, in the melodies. It’s very some of it is reggae, some of it is very hip hop. And those kind of cultural fragments were embedded very early on because we had

more exposure to African American or hip hop music versus let’s say China, which didn’t have American troops staying in China. So there was that. Then somebody might ask, what about Japan? They also have a huge American troop there. I think Korea was because, maybe because we were a smaller country, we were more open to actually getting into and using those vibes. And because of that, think.

The Korean kind K-pop or K-beauty, K-fashion factory has become a little bit more westernized early on and that kind of made the entrance barrier a bit lower for the American market. That’s kind of my hypothesis.

Grace Shao (44:49)

Yeah, because if anything, kind of going back to your point on like South Korean and East Asian companies and people don’t really do a lot of marketing and PR, I would say ⁓ Korean cultural export has been extremely successful and has been a really, really strong soft power export. So I want to end the conversation on again, back to AI. What are some things that you think we might have not covered today? You think we’re missing? Like, what are some trends?

Or say like if we really spoke again, let’s hope not two years later, but let’s say we spoke two years later, what would be true for you to think of how society has evolved, what Korea has maybe contributed in a global AI supply chain ecosystem, how to understand how you view the future.

Ethan Cho (45:32)

One thing that we haven’t touched that I’m personally passionate about and interested in is the mental health industry. It’s going to be very different from now versus three to five years down the road. As we know, the fitness industry, physical fitness industry, has become a huge industry ⁓ after the Industrial Revolution because people started to use less and less of their muscles. I think that’s exactly going to happen for our minds and brains.

⁓ And because of that, this is not going to be driven by AI, but it’s going to be kind of a side effect or a secondary industry from the AI revolution. To keep everybody healthy, think this is something that we as a society and company as country has to work on. there’s going to be, I don’t think this, I don’t necessarily think of this as a dark scenario. I think as we go to the gym, we can go to this mental gym or something.

very on a regular basis to keep ourselves healthy mentally. I think that’s gonna be something very huge. Until now, I think we’ve focused a lot on the hows, like how are we gonna do this? How are we gonna do that? The answer to that has been AI and robotics. There’s gonna be more and more questions about what are we gonna build with this? And after that, there’s definitely gonna be questions about why, why are we doing this? I think that’s not just gonna be philosophical, but it’s gonna be a very practical question.

that will lead to a lot of business opportunities. So I think that’s something that we’ll have to question ourselves and answer and discuss on a very regular basis down the road to reach something meaningful either as an entrepreneur or an investor.

Grace Shao (47:07)

I think that’s really, really meaningful. And I think, you know, we kind of touched on like companion bots and even your you mentioned your son might be even using a companion bot himself. I don’t want to probe on a personal level, but actually on this note, then how do you view that? Like, do you ever fear that he’ll be too dependent on it or, you know, I could be creating a false reality?

Ethan Cho (47:27)

I think it really depends, right? I know this is not the best answer, but like I’m a big fan of the movie, Her. I think it was a very, very good example of how things can evolve. The ending was kind of sad and happy at the same time, but until then, he was very happy with Samantha. So it seems like there’s definitely a scenario where we can be more happy about the world, be more thankful about the world, thanks to this.

Grace Shao (47:33)

Mm.

Ethan Cho (47:52)

maybe emotional buffer that we create with our AI companion. There’s definitely that. But there’s also going to be a downside because the companion will feel real, but it’s not going to be real. So how can we cope with that? It’s going to be something. I still think it’s going to be very similar to the fitness industry just because when we do like, you know, bench presses or, you know, like all these like that pull downs, those are not actual resistances. We’re creating them artificially to strengthen our muscles.

So I think our minds should also be strengthened in that way so that we can cope with all these scenarios that we’re not gonna be able to actually experience down the road because we’re gonna live in our own world, which is gonna be safe and creepy at the same time. you know, a lot of factors that will change down the road. So kind of excited and horrified at the same time.

Grace Shao (48:42)

No, 100%. I think your point on mental health, you use like a general term, but there’s obviously the obvious fear, like what we just talked about, like psychosis and dependency, but there’s also kind of like you mentioned, touched on like, you know, if we don’t really use our brain that way, you don’t really know how to do it anymore. Just kind of like languages, you know, when you move to a country and you don’t use that language for a while, you lose it. Math, I like literally don’t know how to do math anymore. It’s pretty sad. But you know what I mean? Like if these are skills where like you kind of have

push yourself and the gym is something quite, if you think about it, very arbitrarily created for our modern day lifestyle, which obviously didn’t exist even like two generations ago. But yeah, like I think that’s a really interesting take. I don’t know if that’s actually your differentiated view, but you know, I usually always ask one last question to every single guest that comes on the show,

what is one differentiated view you hold? So something that might be a bit non-consensus, it could be provocative, it could be not, know, it could be about industry, it could be about life. Honestly, I think what you just said earlier was a bit, it’s quite insightful. It’s something not talked about in the mainstream enough, but if you have another one.

Ethan Cho (49:31)

differentiate the view. huh. I can make really dangerous comments here. ⁓ but I think,

Grace Shao (49:55)

No worries.

Ethan Cho (49:56)

yeah, this is one thing that I always think about. So I think that the creator cannot make something that the creator has not experienced. That is something that I think deeply about, and that is my personal view on the limitations of AI. How we think about AI is to become this everlasting thing that works 24-7, does only good things for humanity. But

have human beings actually ever experienced that? I don’t think so. And that’s going to be a big question because we’ve never, we don’t know how to work 24 seven. Well, of course we’ve, you know, we’ve done all nighters for sure, but can we actually think of a process that can continuously work 24 seven by thinking, not just operating machinery and also can we think of a kind of standard that is always only helpful to human beings? Like we haven’t really done that.

So I mean, I think that’s gonna be a big challenge. Like, however we construct the system or the standards for AI and robotics and all these systems going forward, there’s gonna be a loophole there. And that’s something that we’re gonna have to figure out as a society as a whole. So I think that’s gonna be something that it’s gonna be very interesting down the road.

Grace Shao (51:08)

Do you kind of, are you kind of alluding to what we’re seeing right now? A lot of people have AI fatigue where they actually make the agents just work 24 seven for them. So essentially the moment they just stops doing a task, they repeat, like they let’s restart it. That’s kind of the work, right?

Ethan Cho (51:20)

Yeah, I think so.

that definitely shows what the problem is. Because we don’t know how to operate these. So I’m facing the same. I don’t use it as much as I used to like a month ago because of that. Because I’m feeling that, this is controlling me, not me controlling that. So there’s this reverse effect. So I think it’s a good thing that a lot of people are already kind of figuring that out. people are kind of.

trying to like healthily distance themselves from all these agents. So that’s, I think, a positive sign. But I think as a society as a whole, that there’s going to be more and more things that we’ll have to think about.

Grace Shao (51:55)

No, I totally agree. And I think there’s certain things. There’s a lot of value in stopping and thinking about the action before the action respoots again. Obviously, there’s certain repetitive work that can be streamlined. so much of accessing knowledge work. mean, this discussion can go another hour, but so much of the whole argument on knowledge work being completely replaced just seems a bit I feel naive for me. Like, I feel like so much of the knowledge work actually requires us.

creating things and I don’t know maybe I don’t understand technology well enough so who knows maybe it can create things on its own. Ambanao, I really really want to thank you for your

Ethan Cho (52:26)

Yeah, that’s another. Yeah,

thank you. Thank you, that’s another hour of conversation so we can do it next time.

Grace Shao (52:34)

Yes, please. Thank you.

Having transitioned from a Stripe engineer to an AI startup founder, John shares his unique insights into the evolution of support tools. We delve into how these tools have shifted from being mere cost centers to becoming strategic assets that enhance customer experiences. John and I discuss the impact of AI on support volumes and staffing, highlighting how integration is reshaping the landscape. He emphasizes the importance of talent density and assembling high-caliber teams to drive success in the tech industry. Through his experiences, John provides practical insights into AI's current capabilities and limitations in support operations.

We also explore the strategic considerations for future AI support ecosystems. John shares his thoughts on the role of support in driving revenue and customer satisfaction, and how AI can orchestrate with human support agents to create a seamless experience. His perspective on building high-performing support organizations offers valuable lessons for anyone looking to innovate in this space.

Subscribe now

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, founders, and product managers.

For more information on the podcast series, see here.

To find the previous episodes of Differentiated Understanding, see here.

Chapters

00:00 The Journey from Stripe to Assembled

02:25 Understanding the Importance of Customer Support

05:29 Lessons Learned from Stripe

10:25 AI in Customer Support: Current State and Future

16:04 The Economic Impact of Support Operations

18:25 The Role of AI in Transforming Support Jobs

24:30 The Future of Support Organizations

26:58 Guardrails Against Fraud in AI Support

32:42 Navigating the AI Ecosystem

38:00 The Value of Long-Term Commitment in Careers

AI-generated transcript

Grace Shao (00:00)

Hey, John, thank you so much for joining us. I just recorded your bio already. It’s extremely impressive. And you’ve done quite a, you’ve had quite a few different roles now as the co-founder of assembled, right? To start, can you just tell us about your story? Like what inspired you to leave Stripe, you know, go into, you know, right now what you guys are doing, which is a software for people who run customer service support operations. You know, now you guys are pivoting into AI as well, or at least leaning into AI. Tell us about all of this.

John Wang (00:28)

Yeah, great question. When we, well, when my co-founders and I started, we were all at Stripe. We worked on a bunch of different things at Stripe. And one of the last things that my two other co-founders worked on was a support tool, an internal support tool. And I remember pretty clearly that they were making a bunch of headway. It was really, really cool. And...

They had gone to this really, really high up person and product. And this person was basically like, why are you guys wasting your time on this? Like you guys are kind of like, you’ve been at Stripe for so long, you know all these things and you’re doing support. Like I’ve got this really cool Bitcoin project that I would love for you to work on instead. And I remember my co-founder coming to me and being like, hey, like pretty bummed this is what happened.

And then I was like, wait, you just saved Stripe, you know, quite a few million dollars, increased customer satisfaction by 40%. And still they don’t understand the value of this. And that’s when we were like, hey, ⁓ there’s something here where there’s a market opportunity. So that’s what got us really, really excited about support. We were doing it at Stripe. We knew it was an undervalued place. We didn’t see any very good tools out there to do support well.

And so we decided to go build something really, really great in the support space and just like make transform and elevate support is our mission. Yeah.

Grace Shao (01:50)

Do you think it was just that stripe was too rich? They were just, and they just didn’t care about saving a couple million dollars? Or do you think it was actually a blind spot for people?

John Wang (01:59)

I think Stripe was definitely very rich at the time. think it was also a blind, it was a combination, right? Because most people, you think of support, you think of it as just a cost center. And I think recently that started to change in the sense that like, hey, this is actually a really important part of your business. But for a lot of companies, like if you look at FinTech, if you look at like a lot of health tech companies, their entire product is their relationship with their customers.

And so support’s actually really, really important for that. And I think a lot of people underappreciated that for quite a while. And now I think people are starting to understand again, hey, if you piss off your customers every time they come and talk to you, that’s not going to be a very good thing. You better be a monopoly. Otherwise, you know, they might not be coming back.

Grace Shao (02:46)

Yeah, definitely. I think I want to kind of lean into that later in our conversation as well. It’s like people are trying to replace support and customer service AI first. But if anything, it’s not the best experience when you’re frustrated with a product and you keep on getting a robot, right? But I want to kind of talk more about your experience at Stripe. You were there quite early. What do you think it taught you, you know, as a very early employee at such a successful startup now?

if even considered still startup and then like what were things that you think you learned there lessons even if soft skills that you kind of took away to to your current role like as a founder.

John Wang (03:21)

Yeah, it’s a great question. You know, it’s really funny actually. I just met up with someone where, so when I was starting out of college, I had applied for all these jobs. I was able to get a lot of them, except for this one company that I really, really wanted to go to. It was called Meteor Development Group. They built open source software. In college, I had built open source software at Ruby on Rails.

I was really big in that community. was like, wow, it’d be awesome to go and make this something I do day to day. And I didn’t get the job. I was really bummed about it. And then I was like, I’ll just fall back on my second here, which is Stripe. And Stripe was the obvious second choice because just the people were really, good. And now like 10 years later, I like think about that and I’m like, the business model is really important.

because Meteor was not a good business. Like open source frameworks is not a good business, but Stripe, really boring. Honestly, it’s just like payments. You process payments, you go talk to Visa. You literally have to like, we had a server in the server room that would send like a specific file with specific tabs and spaces in order to get it out to Visa. Really boring. Really, really core infrastructure too.

And so like the big overarching thing that I learned was like one, business model is unbelievably important because if you can just make a good product when the kind of like market is there and when there’s a really big need, then this can scale like unbelievably fast. Two was the people. I remember talking actually to a few people, Greg Brockman was maybe the second or third person I talked to.

who’s now the co-founder of OpenAI. And I remember just talking to him and being like, wow, this person is so, so smart. This is awesome. And I would talk to kind of like, I would go to the lunchroom and be talking to people at Stripe. that was just, people were talking about all sorts of things. And I think like talent density was a really, really big part of like what made Stripe successful. And

It wasn’t any one thing over time. was one, Stripe was in a great market. And then two, it iterated really, really fast on a lot of little things over and over and over again. So I thought that was a really good place to learn a lot about like what makes a company great.

Grace Shao (05:49)

Yeah, I think it’s interesting you’re talking about talent density and a lot of the AI labs I speak to actually also talk about that. But I’m curious, what does it mean when you have really strong talent? Is it like that they are technologically superior, like they can code better? Or does it mean actually that they can think outside the box, they’re more creative, they can pivot faster? Like what does it really mean to have really high caliber talent on your team?

John Wang (06:11)

I think it depends on what company or like what you’re trying to solve, right? Like talent density for Los Alamos national, like Los Alamos, like building the atomic bomb is like very different than talent density for like Bell Labs, which is very different than talent density at early Stripe, which is very different also than talent density at OpenAI Research. Like I think for Stripe in particular, the type of talent density that was there was really high curiosity.

Grace Shao (06:31)

Right.

John Wang (06:38)

really high product thinking, really technical people, and people that could dive deep on certain problems and weren’t afraid to go talk to a bunch of customers. You saw so many conversations about like, how do we make this particular API parameter better for everyone? And like hours and hours and hours of like making sure it was a really, really good product. And people who weren’t afraid to like, you know, take a week of work and just like dump it away because it wasn’t quite there. So it was like,

This combination of like, they worked really hard, they’re really smart, and they care a lot about the end result and have a high quality bar. That was Stripe’s version of kind of like talent density. But I think like, you know, if you look at the labs, if you look at different research institutions, maybe it’s just, you know, I don’t know, the raw ability. Yeah. But.

Grace Shao (07:26)

research capabilities or whatnot, right? No, that makes a lot of sense. Yeah, I wanted to ask you earlier on in our conversation, you said, you know, look, a lot of people overlook support. It’s not that glamorous. People kind of think it’s like a back office thing. But, you know, is that was that your view back then? How does you kind of, I guess, lean into this? And did your perspective or support change over the years? Now you say it’s very important, right? Did you understand the category correctly? Do you think?

John Wang (07:52)

You know, I think that when we looked at the category, we went at it from like kind of the lens of, Stripe was this company that worked in this unsexy space and did really, really great things. And we thought very similar things about support. It took us a long time to really grock support. And we talked to hundreds and hundreds of different people across different parts of the support stack. And

I think early on, honestly, it was good and bad in certain ways. It was like, we thought we could build a piece of software really quickly that solved everything. Or like, you have that problem, we can build that in two weeks. Not a big deal. And we could solve the specific problems that they had in two weeks. And I remember talking to actually a few people, which was like, the system that you’re trying to do, which is called Workforce Management for Support.

that’ll take you seven years. And we’re like, no way. Like we can do this. We can do this so fast. It’s going to be done soon. And now like seven and eight years later, we’re still working on it. We’re still uncovering more and more things. And that was probably the right, you know, that was probably the right call. But also there’s like some importance to naivety, which is like, if we had known that we wouldn’t have started. like we, yeah, like.

Grace Shao (09:06)

That’s why a lot of people say, yeah, as founders, right?

John Wang (09:09)

Yeah, so I think it was the right thing to do, which is just like start building stuff.

Grace Shao (09:14)

It’s amazing. ⁓ Why don’t we pivot into actually understanding your product bit better? So for someone who has never worked in support ops, what is the simplest way to explain to them what Assemble does? Because even between us, we had calls, we had back and forth emails. I was like, John, I don’t understand what you guys do. I’m trying to read through this material. I’ve listened to a few in the interviews. I don’t know what’s happening. Can you just dumb it down for me and explain to me what exactly you guys do?

John Wang (09:37)

Yeah, for sure. Let’s say you have like 10 people on your support team and you only do email, then you probably just staff them nine to five, right? Like there’s no big deal there. Once you start having a few more people on your support team, let’s say you have hundred people now and you might want to chat to your customers because AI chats, AI chat bots are a really big thing. Then you actually need to start thinking about

when do these chats come in and how many people do I have in order to handle those chats, right? Because like, if you were talking to a chatbot, you’re getting instant responses back and forth, back and forth. And then you’re like, I have to wait 48 hours for the human response after I get handed off. That’s a really bad experience. So the problem is, you you’ve got a bunch of people who are calling in to support, writing in, who are chatting in.

and they’re coming in at all different times of the day, they’re calling in for different types of problems, right? You, on the kind of like back end, you have a bunch of people and those people might be able to do different types of things. Like I might be a really good person to handle, you know, where’s my money kind of issues, but I might not be as good at like ⁓ fraud issues, right? Like if you’re having problems with fraud on your account.

So there’s a lot of ways in which you can actually put people to the actual incoming tickets. And what our platform does is it tries to match those two things up. if you think about supply and demand, supply is the people that you have and demand is the people, like your customer is asking for questions. And if you don’t match those up well, you’re gonna either...

spend way more money than you need to because you’re just going to staff everything way above what you need, or you’re going to have a terrible customer experience because it’s going to take you a really, really long time to get back to people. So it’s really an efficiency play. How do we make it really, really efficient for you to answer questions? In the last few years, we’ve also added AI agents, which is, you know, how do you actually respond instead of just with people, but also with AI to

go and answer a chat or answer a phone call directly using AI.

Grace Shao (11:50)

That’s amazing. I really didn’t know there was so much like science kind of going behind that. I just thought kind of like you’re on a chatbot usually you have to have your frustrating like get me someone, get me someone. I’m one of those people who like no pages pressed zero all the time. I’m like, get me a human. But it makes sense. actually once you can match the talent with like the issue, it can be a lot more efficient in solving the issue and the customer experience will be much better as well. On the AI agent side.

What’s the kind of, I guess, consensus right now? Like, are they really actually good at solving issues? Are customers complaining about them? Like, ⁓ how sophisticated are they at this point? He’s like, in my day to day, you know, obviously calling the banks or DHL for pickup or package returns, whatnot. None of those agents are really a pleasant experience, frankly.

John Wang (12:35)

Yeah, I think this depends pretty drastically on what tools you give these agents access to. I would say that the standard experience right now is fine. It will answer knowledge questions for you. And these can solve anywhere from 30 to 60 % of incoming issues, depending on how many knowledge questions you get.

the place where it really is important is when you actually give it access to say your backend database and you can like make a refund or you can look up in order or you can identify why is my what’s going on with this error, right? And that is actually the hard part that prevents most of these banks and airlines and etc agents from being very good is because like that

Access to data is a thing that they need to actually run and be able to perform actions. And then also the evals for that are really, really hard and not something that you just like launch without really thinking about it. So I’d say it’s in a progressive, like it’s in a progressing state, not at a place where it’s like, this is absolutely solved, but there are also some of our customers who have 90, 95 % of all issues who are able to be completely automated.

because they’ve spent the time to give access to all of these systems and spent the time to validate that the agents are performing.

Grace Shao (14:00)

Very interesting. ⁓ I want to pivot into the to be kind of angle. Who are you guys actually selling to? Like who are the people inside companies that are managing this? Is it the head of support operations? And when they are buying and assessing your product like yours, is it really winning on price? Is it like over, you know, other maybe large softwares? Is it winning on speed and service? Like help us understand essentially how you guys are succeeding winning over customers.

John Wang (14:26)

Yeah, we generally sell into the head of support. Sometimes that person rolls up into the COO or there’s a head of operations or something like that. But generally there’s some group that is working on unsupport related things and that’s who we sell into usually. I think generally our differentiated, like the way that we actually go and sell this is one, we know all about workforce management, which is like a really, really nitty gritty detail about how you

make your systems really good. And it can save you millions and millions of dollars. Almost actually, and this is one of the things that’s really funny, it’s like using our AI agents versus using our workforce management, we actually see somewhat similar gains across those two. Because to use the AI agents, you’re usually doing it so that you can reduce head count, right?

And in order to reduce headcount, you need to know how much can I reduce headcount without hurting my customer experience. And for that, you generally need something like Workforce Management. So what we do is we go in, usually we have Workforce Management helping you understand how your system is set up. And then what the AI agents that we can also bring in is a relatively easy sell because

our AI agents are really, really connected to kind of like how you staff and when you pull in people. The thing that you were mentioning, which is like, hey, my bank still doesn’t have a very good experience, that’s true of a lot of places. And getting access to information is really hard. So escalating to a human actually happens pretty frequently, sometimes 20, 30, 40 % of the time. So getting to the right human or the right

or figuring out when to escalate to the right human is a really, really important skill to have. If you spend a million dollars a year with me, I should escalate you much more quickly than if you are a free user and you haven’t spent any money for me ever. Similarly, depending on the topic, depending on what kinds of things you’ve already previously talked to me about, I should be able to get to different types of agents and I should be able to have different levels of thresholds.

that send me to a human. And I think because we have and handle the workforce management side, our ability to do the handle time and to make sure that you’re getting to the right person is much, better than a lot of our competitors.

Grace Shao (16:45)

So there’s an unspoken tier system then I guess with customer service as well that we don’t realize. How should we think about the economic importance of support operations? In terms of, we always think of it as like we said, back office support, but how much, do you have any proof that like basically better customer service equals better revenue?

John Wang (16:52)

There is, and sometimes it’s spoken. But yeah.

You know, that’s a good question. should probably have some specific proof here. I guess the best anecdotes I can find are usually the kind of like medium to long-term anecdotes where companies that do not invest in their customer support tend to, you know, regress to the meat, right? Like if you are really trying to bare bones your way through customer support,

⁓ Your customers will understand that and it’s not going to affect your revenue right now, but it will likely affect your revenue in 6, 12, 18 months the next time that purchase happens. have seen actually some of our customers, so in our AI agents, we have a configurable setting that’s like, do you want to be containment focused or do you want to be escalation focused? And how good of a customer experience do you want?

And we’ve generally seen actually that there’s a strong correlation. Obviously we haven’t run a ⁓ natural experiment or a true A-B test with this because it’s pretty impossible. But you see a general correlation between the customers that spend more money on support, the customers that spend more on trying to have a high quality experience, and the revenue growth of those companies.

actually most of the customers that we spend a lot of money on care so much about support that they actually have, you know, executive briefings every week about these, about what’s going on. And they’re the people who have the largest support teams and they’re the people who kind of like make the most, make the most changes with their team. Obviously this is a very biased perspective from our customer, like set of customers, but I think that there’s still something to that where if you spend money and if you want to make a

your support really, really good, that does tend to pay off with customers because they do tend to notice and it makes it easier from a product perspective to paper over all of the things that aren’t so great.

Grace Shao (19:06)

Yeah, no, totally makes sense. think even as consumers ourselves, we would be likely turned off by certain brands or experiences if the customer service really bad, right? Unless, like you said, they’re a monopoly and there’s nowhere you can go. All right, let’s talk about AI. You kind of touched on that earlier, but the naive view is that AI automate support, you know, a lot, a lot of the conversations right now about, my God, jobs are going be taken, especially the first batch is probably in roles like operationals and customer support roles.

Second batch, people are saying are maybe in like more repetitive execution roles like junior consulting roles, a lot of junior training up roles, right? How do you see that? Because at least from where I sit in Hong Kong, a lot of stories are coming out saying markets like India, the Philippines, know, across Southeast Asia where they traditionally served as those telephone call centers or operational centers, they are getting caught. Is that going to be a trend forward?

you know, how should we understand this?

John Wang (20:02)

Yeah. Yeah. I think there’s a few things. There’s like, like with all things, there’s a lot of nuance to this, which is I think your trend on seeing, you know, what we call tier one support, the first line of support who are traditionally humans outsourced. That is a place where we’re seeing a lot of change. And I don’t think that trend is going to slow down. That said, there’s a very interesting other trend that

we’re seeing, which is that total spend on humans and headcount isn’t necessarily going down by that much. And it’s kind of like Jevin’s paradox where we see a lot of our customers and a lot of customers of other AI users ⁓ who have amazing resolution rates. They’re like answering so many questions, but that’s causing actually, or maybe there’s some correlation here of

the number of tickets they’re getting and the number of chats they’re getting is like way, way, way higher than before. And I think there’s a few parts to this. One is you see way more ability for your AI agents to answer questions. And so obviously people are going to ask more questions because like, Hey, it used to be really hard for me because I had to literally type out an email to a human, wait a few days and get an answer. And now I can just like get an instant really good answer. Right. So I’m going to try asking more questions.

The second thing is, as these companies do better and better, you actually just have this natural induced demand of increasing usage, numbers of people who are asking for support. So the higher amount of support that is automated is also, the general number of how much support is coming in is also very high.

And so that actually offsets a very, very large portion of the head count. The head count is changing though. It’s not going to be the typical tier one support where it’s just like, answer an easy question. That is mostly going to go away to AI, think. The types of head count that is coming in are like, know, internal agents, people who are really good, people who can provide white glove support and like...

actually go talk to people and provide like a human experience because like our companies still want and really crave giving that experience to people. And that’s just not what the kind of BPO standard really is. So I think it’s changing in the type of what you would see.

Grace Shao (22:31)

Yeah, I was actually going to ask about that, like, as in when AI agents start resolving more tickets, if we’re just going to see reduction of headcount. And I think you answered him when you once wait, whereas like, yes, in initial stages, but later on, there will be new jobs created, right? Essentially, people who will be managing more critical issues or even managing the agents. I want to understand. So for your company right now, essentially, are you a are you like a middleman between the human agents and the AI agents and becoming the orchestration layer, like you’re providing the service, the training and the orchestration. Like, how do we understand that?

John Wang (23:05)

Yeah, that’s a great question. So we think of ourselves as how do we get you to the right way to answer your question, right? In our view, there’s kind of three main types of people that can answer a support question. One, it’s AI. Two, it’s a tier one BPO’d outsourced agent. And three, it’s an internal agent who’s like super well-trained and like super, like, you really carrying about, like really trained on customer support. And what we are trying to do is make sure that you get placed at the right area, depending on what kind of issue you have and who you’re talking to and like what is the kind of like a cue that is backing up the set of people who need calls. So for us, what we’re trying to do is really provides you that ability to choose across a bunch of different options. So we don’t actually provide any, like we don’t provide any BPO agents, we don’t provide any internal agents. All we do is provide the software that routes you. And we also provide the software that can do the AI agents, or you can actually plug into a different piece of software if you want to have your own AI agents too.

We’re trying to make sure that we are kind of third party and that we are making it really easy for you to optimize your support regardless of what specific providers you use.

Grace Shao (24:30)

So in your view, what does a well-run support organization look like in, let’s say, three years as AI adoption becomes mainstream or more more mass market?

John Wang (24:38)

I think you’ll probably want to have all of the different types of support using AI. So voice AI, chat AI, email AI. I think you’ll want to have a lot of nuance between the different types of customers that you have. You can’t generally provide the best level of support for literally everyone. Though this depends on also your customer base, right? Like a consumer customer base versus a super enterprise customer base with 100 very large customers is completely different. But let’s say for a standard company that might have ACVs that are in the, I don’t know, the 100 to couple tens of thousands range, then you’re probably going to have a combination of AI agents and human support. And you might have different tiers of human support, right? Some human support that’s really good at answering support questions and other tiers of human support, which is like, you’re just managing the agent. I think the other thing that’ll happen a lot is you’re gonna start to see more like, supporting agents acting in a simulation where right now, like the kind of typical flow is like a supporting agent gets a ticket and they answer it and it goes back. I think as the agents get more like, get more and more training data, get access to more information, really they’re only gonna come to humans for escalations. And similar to how Waymo works, if you’ve ever taken a Waymo, it’s a great experience, you’re like driving, driving, driving, and sometimes you get kicked out and a human operator in the Philippines is like, hey, I need to move you around this truck, right? And similar to support, That’s probably what’s going to happen. A human operator is going to come in and be like, hey, I can give you a refund right here. And then what’s going to happen is the AI agents are going to train on that, right? They’re going to like learn and get better. And you’re going to be able to use that whenever you have an interruption to understand like, why did I have this interruption? How do I make my model better for the future? And then you’ve got your closed loop. So I think in the future, you’re going to see much more of that happening than people who are just like, coming in and their job is to solve as many tickets as possible. I think the change is gonna be like, okay, people are gonna start to need to provide the best possible response in that particular instance so that the models can train on that and be as good as you are.

Grace Shao (26:58)

actually just on that, do you think then we’ll see more and more fraudulent activity or people trying to exploit that? like if say you know the models trained on, I say this one buzzword or one keyword and it triggers like refund. What if I just go on the call like on the phone all the time, just to be like keyword, keyword, you know, and then like how do we prevent something like that? Or do you guys kind of get involved in that building this guardrails as well?

John Wang (27:21)

Yeah, no, that is a age old question. think like, wouldn’t say there is going to be necessarily more or less of that, but I think like, it’s kind of like the cat and mouse game of like, everyone has always been doing that. And so like, and the methods always change every, every few months. I think the methods will change every few months here too. Our AI agents have a lot of guardrails put in place to automatically detect that. And we also have kind of like post-hoc guardrails which are like scanning through our logs and trying to identify situations where that might have happened. And we’re also training on those examples, right? So I think, yes, people will definitely start to exploit this and be like, hey, how do I get a refund faster? But there’s a ton of guardrails that you can put in place. For example, each account, you can have one or two, have like refunds without looking until that actually gets flagged and it needs to go to a human or.

You can set good policies, for example, like, you know, if it is within policy of 30 to 40 days after purchase, like automatic refund, otherwise, you know, flag it and do something with it. So there’s a ton of stuff that you can do to actually like reduce the possibility of that. And I do think that it will end up being cat and mouse game like over and over again, as people get more sophisticated.

Grace Shao (28:39)

Right, right. And they’ll start using AI to trick AI. That’s what’s scary, right? So as we talked about, different gender standing the podcast does not have to interview anyone related to China or Asia, but we do have kind of an Asia angle to a lot of how we view the world. So my question for you really is because you’re like out in San Fran and like your company actually has no sales in China or anything. But I actually had a curious question. How does SF

John Wang (28:43)

Totally, yes.

Grace Shao (29:05)

as a whole, the startup ecosystem kind of view the current rise of a lot of Chinese AI. And have you guys yourself or your peers, you know, tried to use Chinese open source models over the years? Is there any view on the open source models given that, you know, you previously said you were very involved in open source and I think it’s part of your philosophical belief as well, right? So just kind of like the high level vibes.

John Wang (29:28)

Yeah, our vibes might be different than at the model, like the Frontier Labs, honestly. Our vibes, we love the Chinese open source models because it adds more competition. And I think the open source models are actually very, very good. I think from my friends at OpenAI Anthropic, they don’t like it quite as much because it’s competition. But for us, we have no allegiance really to any of the Frontier Labs.

or any of the models that are out there, we want to provide the best possible experience to our customers at the best possible price. And that has meant, you know, over the years, like making changes in our models, making updates and to figure out what is that frontier of cost or performance. The Chinese models tend to perform really, really well on that, especially

Grace Shao (30:12)

Mm-hmm.

John Wang (30:19)

kind of like the latest series of models, we’ve actually spent a lot of time in the last six months kind of like pulling out a lot of our tokens. We have tens of billions of tokens per day. And a lot of it now goes to models like Quen or Kimi. And like that has actually started to really, really increase over time, mostly because you can find to them, you can do RL on them, you can...

have better latency on them, you can run them on your own hardware. There’s just like so much more stuff that you can do with it. And also, you know, the cost performance latency trade off is really, really good. Now, most of our most of the like the strategy we take is actually one where we try to understand the use case and the problem and what type of model is necessary for that. So for kind of like the main model that’s actually answering questions. We’re actually usually using a frontier model for that. But actually the majority of our tokens come from out of secondary processes, processes like detecting if I need to escalate, detecting if there’s a fraud here, detecting if there’s an adversarial intent, making updates to large swathes of data in batch, like all this other stuff where you really don’t need frontier level intelligence and where if you have a a well-tuned prompt and an open source model or an open source model plus a fine-tuned model, you can get at or better in terms of frontier performance. We’ve really seen that and we’ve actually been able to save millions on our token costs in just the last two or three months by being very smart about how we use our models. And we’ve also seen a 15 to 20 % increase in quality.

⁓ Just because like when you go and you have evals, you can make things much, much better more quickly with these open source models.

Grace Shao (32:14)

Yeah, I think that’s like the general sense I kind of get from a lot of startups, right? In a known day, it’s like, you guys are obviously more cost conscious. What is the best price to get to what you need? And there’s like a tier system where how you use the models, you might not use the most frontier models for everything. I think that makes a lot of sense, business sense, especially. Is there anything you would like to share with us that we haven’t touched on in terms of, just the overall AI ecosystem, any thoughts on, you know, where we’re going with this AI agentic push right now?

⁓ you know, are we really going to see that, you a giant moment, like just kind of some high level thoughts.

John Wang (32:49)

Yeah, that’s a good question. Recently, I’ve been thinking a lot about Opus 4.7, which got launched a few days ago. And it’s actually kind of similar to what we were just talking about in terms of this price for performance ratio. And it seems like, based on my usage, based on our evals, based on other people’s usage on the coding side, that it’s a better model, but it is also more expensive.

than before. like, you’re really it’s literally like a trade off in terms of dollars and intelligence. And it’s really interesting because, you know, a year ago, every model would just be like, this is strictly better, and it’s probably cheaper, and you’re to get more context and like, everything’s better. And you could basically just bet that you’re just going to like get better models across the board. And now actually, you’re just like kind of moving from this part of the like the frontier curve to the other part of the frontier curve without actually shifting the entire curve. And that’s happening with a few more model releases. You still see general increases in the frontier, but it’s less stark every single model release that you see that. And so I think it’s just an interesting area to look at because when you get into that world.

Gross margins has become really important. Gross margins for ourselves as a startup, but also gross margins for Anthropic and OpenAI. One of the funny things that I’ve seen, just talking to people who are working at Anthropic and OpenAI, and also people who are trying to invest in those companies, gross margins are actually incredibly important. One of the OpenAI right now is becoming a much more...

hand investment than before. And like, it used to be like six months ago, it’s like, you have it, you have shares of OpenAI, like, how do I get in? Now it’s completely different with, you have shares of Anthropic, how do I get in? And I think part of that’s because like, OpenAI wants to spend $100 billion on infrastructure. And Anthropic is a lot more measured in the way that they’re spending money. And I think gross margins actually do matter a lot right now. And that’s where I think actually

Chinese open source models are making a big difference because just at the end of the day, you still have to make money. And if you’re losing money on a per token basis, that’s really bad because if you go to infinity, you lose infinity money. And if you make money per token, great. Ramp usage up as high as you can.

Grace Shao (35:03)

Yeah. It’s just so crazy how the sentiment shifts like so every three months I feel like and then to your point like whenever I speak to investors like oh my god I got my hands on some anthropic shares and last three months earlier. Oh my god I got my hands on opening I like it’s just like and like oh no one would invest in opening right now like I don’t want to do that like people just completely go like black and white on these things it’s pretty crazy how the pendulum swings I do have a question actually on the infrastructure side doesn’t it actually make sense for open AI to eventually own their infrastructure because otherwise they have to become

continuously constantly pay the hyperscalers for all the infrastructure like so in the grand like scheme wouldn’t it make sense? I mean although obviously how much you’re spending is like absolutely crazy.

John Wang (35:55)

I think it actually does. And I think that’s like part of the problem, which is like, you know, if you think about what their compute costs are, I think actually doing all of these things that they were doing makes perfect sense. And it makes especially perfect sense if you have investors who are willing to bankroll this. But it’s almost like the, ⁓ what’s that paradox? It’s like the St. Petersburg paradox, something like that, where it’s like, you keep doing,

your expected value is infinity and you keep doubling your money basically, but at some point you need to not double your money because you don’t have enough money.

Grace Shao (36:32)

That’s such a mo- I’m like, I’m still confused when you’re saying, go back. You keep on doubling your money.

John Wang (36:36)

Sorry, So I think the I think it’s like Let me let me look this up st. Petersburg paradox is Okay, it’s a coin flipping game and You start at two dollars and with every tails you double the pot and you can basically decide to like take your money at any time, right? and so you you’re doubling exponentially as you go up and

If you compute the expected value, you should basically just like, keep going forever because your expected value is like infinite, right? Like because the doubling of the pot is better than kind of like what your losses are. You just got to, you got to run. And I think OpenAI is in this St. Petersburg paradox where it’s like, well, in theory, double everything, keep going. But in practice, you don’t have enough money.

Grace Shao (37:17)

Yeah, I see what you mean.

John Wang (37:25)

and resources to be able to do that. I think that’s actually what’s happening is like, there’s not enough money in the world, not enough investors with liquid cash who are willing to invest in a business as big as OpenAI while the gains and the returns are still, yeah, having improvements. So I think it’s both rational, but also, you know, actually practically very hard to make what they’re doing.

Grace Shao (37:41)

haven’t been proven. Yeah. totally. Okay, I want to ask you one last question, which I ask every single guest. What is one differentiated view you have? It could be on your own sector, industry, life.

John Wang (38:00)

man, have a really like, I have one that like is very controversial. I don’t know if I should talk through that one. ⁓

Grace Shao (38:07)

You’re get doxxed and to hate it after this.

Okay, tell me that one after, I wanna hear it.

John Wang (38:17)

Yeah, yeah, Let’s see. Like... I would say, I don’t know if this is differentiated now in the market or not, but the thing that I’ve been thinking about recently is that you should stay somewhere long enough where you see your mistakes through. And I think it’s like slightly differentiated right now, because like you’ve got in Silicon Valley, at least you’ve got people who are jumping between big labs, who are jumping between different startups where it’s like, Hey, I can make the next, you know, $5 million.

by going to this next thing. And there’s just a whole huge amount of opportunity and there’s like a ton of opportunity costs to staying somewhere for a long time. And at the same time, think like long-term staying somewhere for a long time is actually one of the best things that you can do for your own learning. And it gives you that a better shot to make like the long-term massive gains that you could have like $5 million.

is amazing for someone. But if you want to build your own startup, if you really want to like change everything, if you jump around between companies every year or two, like you’re probably not actually going to learn a lot. And you’re probably not in the position to make really hard decisions and then have to see those hard decisions through and then, you know, be able to learn and see that feedback from those hard decisions. Especially if you’re jumped like

Especially if you’re like at OpenAI and you’re like, no, investors don’t want this anymore. You jump ship to Anthropic. That’s like, you know, I don’t think you’re going to get that, the learning that you really need to get.

Grace Shao (39:46)

Yeah, yeah. actually agree with that. think I also took some time and like experience to realize that because when we’re all young, like you’re really excited, right? It’s like, this looks cool. That looks cool. this person hates me. hate that person. Like you take everything very personally and then, you know, we’ve all heard these stories from peers, even ourselves. But what is the threshold though? Because then the other side of the argument is that like you see people who’ve been in a job for like a decade and clearly they’re frankly not.

moving up in a very corporate structure way or even intellectually growing or even, you know, excited about their job anymore. You know, the joke is like you get the like, okay, this sounds on PC, but you know, like the 45 year old VP that’s been a VP for the last 15 years at banks, we have a lot of these. So what happens? Like when is it best for them to actually maybe jump or some say, that was like a lifestyle decision where they want to take it easy because they have some more time for kids. Fine.

But taking that kind of considerate way, wouldn’t it sometimes be better that you jump to try something new to take risks?

John Wang (40:50)

I think if you are in a place where you’re unhappy with, so I will caveat this with, have to, you should only stay if you’re excited about what you’re doing and you’re learning continuously and you’re surrounded by great people. If those three things aren’t true, yeah, it’s really hard to fly.

Grace Shao (41:05)

Which is so hard to find. you were very lucky at Stripe, right? Like you said, you were just surrounded by very high caliber, high agency people, but not everyone can get all those things at the same time. ⁓ But no, that’s great. Thank you so much, Sean. ⁓ I had a lovely time chatting with you. I still wanna follow up on what was the unspoken differentiative you later. All right, thank you.

John Wang (41:17)

Yeah. ⁓ Let’s do it. Let’s do it. Thanks, guys.

In Part 1, we argued that DeepSeek V4 is best understood as a form of shared R&D for China’s open-source AI ecosystem. The Huawei adaptation is one of the most strategically important examples of that. Making V4 run on Huawei’s stack is not just about one model. It is about creating a reference workload for China’s domestic AI hardware and software stack.

The easy headline is that DeepSeek V4 runs on Huawei chips, so China is reducing reliance on Nvidia. Directionally true, but that could be too simplistic. V4 in no means proves China has solved AI compute, and it does not prove Huawei can replace Nvidia for frontier training.

And even further, it does not mean CANN has matched CUDA. And it does not mean Nvidia’s global training franchise is suddenly impaired. However, it does give Huawei something it badly needed: a serious workload around which the domestic stack can organize and a group of developers they can work together to make CANN better.

That is why this matters. A chip ecosystem becomes real when there are workloads, developers, kernels, compilers, cloud deployments, enterprise customers, benchmarks, support contracts, and a reason for people to optimize around it.

TLDR, my main points:

1/ V4 gives Huawei and CANN their clearest frontier-class reference workload so far.

2/ The implication is strongest on inference, not training. We should not overclaim that China can now train frontier models end-to-end on domestic silicon.

3/ The three-month release delay suggests DeepSeek optimized for ecosystem strategy, not just launch speed.

4/ Jensen Huang’s export-control argument looks more rational after V4. Export controls may restrict Nvidia’s business in China, but they also create a stronger incentive for China to build a parallel stack and really push for self-reliance.

5/ The Nvidia bear case should not be overstated. Losing China's inference share is not the same as losing global training leadership.

Share

Huawei benefits from the open-source story

The Huawei story is not separate from the open-source story. It is the open-source story moving into hardware.

DeepSeek V4 was adapted for Huawei’s Ascend AI chips, for which DeepSeek granted early access to domestic companies such as Huawei, rather than sharing the model with US chipmakers for performance tuning, and Huawei said V4 is fully supported on its Ascend 950-based supernode clusters. Huawei also said its chips were used for part of V4-Flash training, while DeepSeek did not disclose whether V4 itself was trained on Nvidia or Huawei chips.

The conservative consensus reads that Huawei has not replaced Nvidia for frontier training but moved credible inference workload around onto Huawei, which the domestic stack can improve.

Serving a model and training a model are different problems. Training is where Nvidia’s highest-end systems, interconnect, software stack, and cluster-level reliability matter most. That is still the hardest part to replace, and it doesn’t look like China’s domestic options are near that. Inference is where the model gets deployed into real products, enterprise workflows, consumer apps, and cloud APIs. It is where latency, throughput, utilization, memory, cost-per-task, and reliability matter every day.

So if Huawei Ascend can serve leading Chinese open models at acceptable latency, throughput, and cost, then Chinese hyperscalers, SOEs, and private companies have a real domestic deployment path. What this could mean is that the Chinese inference market can become less Nvidia-dependent. And that itself is already a big deal.

The strategic implication is not that Huawei has caught Nvidia across the board. The strategic implication is that Huawei may now have a realistic path to becoming the domestic Chinese inference standard. Narrower claim, but still very important.

V4’s design is part of the hardware strategy

The hardware case for Ascend is not only about political will. It is also tied to model design. If China cannot get enough frontier Nvidia chips, there are two ways to respond. One is to build better domestic chips. The other is to design models that are easier to run on the chips China can get.

This is why inference efficiency matters so much for Huawei and DeepSeek. If DeepSeek can make frontier-ish models lighter to serve, Huawei does not need to match Nvidia perfectly chip-for-chip in order to become viable for a large part of domestic inference. It needs to be good enough for the workloads Chinese companies actually run. Which could, in theory, be a more realistic path forward.

This is also where the open-source layer matters. If DeepSeek figures out how to make the model more hardware-friendly, that work can become a reference point for everyone else. Other labs can study the architecture. Huawei can optimize around the workload. Chinese cloud providers can tune deployment. The model layer starts reshaping the hardware layer. The five-layer cake is moving onshore one by one.

This brings us back to the Dwarkesh Patel interview. Jensen’s argument was that the US should care not only about leading at the model layer, but also about keeping (Chinese) AI developers on the American technology stack. He warned that forcing Nvidia out of China could create two ecosystems and push Chinese AI developers toward internal architectures (and lose their leverage as an American firm)

You can believe export controls are necessary for national-security reasons and still accept that they create incentives for China to build a domestic stack. Those are not contradictory views.

V4 is basically the kind of event Jensen has been warning about (he probably knew). Not because Huawei has caught Nvidia. It has not. Not because CANN has matched CUDA. It has not. But because the learning loop is starting.

DeepSeek optimizes for Ascend. Huawei gets a serious workload. CANN gets more developer attention. Cloud providers get more deployment experience. Enterprises get more comfort. The next model can become more Ascend-friendly than the last one. That is how a parallel stack begins.

The point is not that it works perfectly today. The point is that the incentive structure is now very clear. If Nvidia is not available, Chinese labs will optimize the model around the chips they can use. Let’s say now, DeepSeek open-sources all that work, the whole system in theory, and then move faster. And this isn’t some like new epiphany, this was a logic that was shared with me by one of the labs earlier in the year as well.

What V4 means for Nvidia

As we’ve written in Part 1, the market read is that DeepSeek essentially took one for the team and delayed the V4 release by almost three months to re-engineer inference so it could run properly on Huawei’s stack before launch.

DeepSeek absorbing that Huawei adaptation burden is meaningful. It took on the wait and the cost. Running on Nvidia is just easier. The tooling is better, the developer familiarity is better, and the software ecosystem is more mature. Most other Chinese labs do not really have the luxury to spend that kind of time on Huawei adaptation when they need to launch, monetize, raise money, and keep customers engaged.

Now, it’s important to note that CUDA’s moat was not built overnight. It took more than a decade of developer adoption, libraries, tooling, optimization, documentation, community familiarity and customer deployment. CUDA became the default because developers built on it, optimized around it, debugged on it, and trusted it. That kind of ecosystem does not get replaced just because a domestic alternative exists.

CANN is not CUDA, and no one is pretending it is even close. Developers still prefer Nvidia when they can get it because CUDA is easier. But every alternative stack needs a first serious workload before people treat it as a proper alternative.

Before V4, the case for CANN was mostly strategic. The government wanted it. Huawei wanted it. Chinese customers knew they needed an alternative. But developers do not optimize for strategy in the abstract. They optimize around real workloads.

V4 changes the practical conversation. Now, Huawei can say: here is a leading Chinese open model, adapted to our stack, with DeepSeek behind it, and with real customer demand forming around it. So what V4 gives CANN is a proper workload. It gives cloud providers a reason to optimize. It gives enterprise customers a reason to test. It gives other Chinese labs a reference point. And once that loop starts, the stack can improve. Developers find bottlenecks, kernels improve, compilers improve, cloud deployment improves, and model architecture becomes more Ascend-friendly. The next model becomes easier to run than the last one.

This is maybe what Jensen Huang has been talking about with a capital E, ecosystem moat. Ecosystems can be built, albeit not overnight, but through repeated workload-driven optimization.

This is why the Huawei support matters, but the claim should stay narrow. The strongest evidence is around V4-compatible inference and partial V4-Flash training support, and def not proof that Huawei has replaced Nvidia for frontier training.

The conservative read is that the clearest evidence is around inference and partial V4-Flash training support. DeepSeek has not disclosed whether the core V4 / V4-Pro training run used Nvidia or Huawei chips. So I would keep the Huawei implication focused on deployment and inference, not frontier training replacement.

But inference alone is enough to start an ecosystem loop. Once V4 runs on Ascend, Chinese cloud providers have a reason to test it. Enterprises have a reason to deploy it. Huawei has a reason to improve kernels, compilers, and tooling around a real model. DeepSeek and other labs have a reference point for making the next model more Ascend-friendly.

This is also where Jensen’s argument looks more rational. His point has always been that the US should care not only about who has the best chip, but also about which technology stack AI developers build on. If export controls push Chinese labs away from Nvidia, then over time, they will optimize around the chips they can actually use.

The important business implication for Nvidia is therefore not that V4 is a technical death blow, but that it is an ecosystem warning. Nvidia can remain technically ahead while China increasingly routes more deployment demand toward domestic alternatives. If Chinese models become good enough, open enough, and optimized enough for domestic chips, then the question in China slowly shifts from “is Ascend better than Nvidia?” to “is Ascend good enough for this workload?” And that could be a much lower bar.

Good enough under export controls. Good enough under domestic procurement pressure. Good enough for Chinese cloud companies and SOEs. Good enough once models are designed around it. Good enough once the software improves through real usage.

This is the part Nvidia should care about. CUDA’s moat is still very real globally, but CANN finally has a serious workload to improve against. And once a domestic stack has workload, demand, and policy support at the same time, the gap can narrow faster than people expect.

V4 is not a technical death blow, but it is a market-access and ecosystem warning for Nvidia. If China keeps building models that are good enough, open enough, and optimized enough for domestic chips, then Nvidia can remain technically ahead while still losing more of the Chinese deployment market.

The US open-source debate

There has been one dominant US narrative around DeepSeek’s open releases: if US frontier labs do not build open models, US enterprises may end up running Chinese ones. That framing is being used to push US labs toward open-sourcing as a defensive, geopolitically motivated move.

I do not think the logic fully holds. If the “China threat” did not exist, would US enterprises not still want frontier-scale open models? Of course they would.

The real driver of enterprise open-source adoption is not geopolitics. It is economics. Enterprises want performance, cost control, customization, privacy, deployment flexibility, reliability, and less vendor lock-in. Geopolitics can shape procurement at the margin, especially in sensitive sectors, but the baseline reason enterprises want open models is economic and operational.

China did not invent that demand. The argument against open-sourcing is geopolitically charged. The argument for open-sourcing is also becoming geopolitically charged. But open source should not be reduced to a geopolitical weapon. Its core value has always been that it spreads capability, reduces duplication, and accelerates innovation.

If you think about it, last year, V3 shocked the world by showing how much China could do under constraints. V4 may be more low-key important because it shows how the whole Chinese AI ecosystem can reorganize around that constraint and almost unite under it all. Faced with lower model costs, there is an urge to be more collaborative, lean into shared R&D, pursue greater specialization across labs, put more pressure on standalone model companies, and push harder toward domestic inference infrastructure.

I have read through at least twenty reports on pricing and performance, but what dawned on me when I took a few days away from it all and thought through the ecosystem and disruption, is that V4’s implication on the business and overall sector that is very much overlooked right now. Thus, on a Saturday morning, the light bulb finally went off for me, and I jotted everything down. Part 1 will focus on V4’s direct implications for the industry, the signal behind the action, and how China’s open-source strategy is becoming increasingly cohesive. Part 2 looks at what V4 means for the tech stack and what Jesnen Huang keeps warning about across what he calls the five-layered cake.

Subscribe now

The easiest way to understand Chinese AI right now is that the labs are, in some strange way, starting to work like different departments inside one large AI company.

Obviously, this is not literally true. DeepSeek, Z.ai, Moonshot, MiniMax, Alibaba, Tencent, and ByteDance still compete for users, enterprise customers, talent, capital, and attention. Fundraising is competitive, API pricing is competitive, and model launches are competitive. But technically, the ecosystem is becoming more collaborative than people realize. And it’s looking more like what we heard from the researchers in February, where DeepSeek now plays an infrastructure provider role in some sense.

DeepSeek is increasingly doing the foundation-layer work: architecture, efficiency, inference optimization, and now Huawei-stack adaptation. Z.ai is leaning into coding and enterprise deployment. Moonshot Kimi is leaning into long context and agentic orchestration. MiniMax is leaning into multimodality and low-cost inference. Alibaba, Tencent, and ByteDance are leaning into distribution, cloud, workflow, and product integration. So even though these companies are still fighting commercially, the technical layer is starting to look more modular. Each lab is leaning into its own area of expertise, while absorbing what others release as open weights, disclose through technical reports, or expose through benchmarks and deployment behavior.

This is where DeepSeek V4's implication is being overlooked. The obvious read is that V4 is another pricing shock, and that is not wrong, but obviously not as big a shock to the global AI community.

The second read is that it will pressure Z.ai, MiniMax, Kimi, and everyone else trying to charge premium prices for text, coding, and agentic workloads. On top of already very competitive listing price for its API, DeepSeek is currently offering a 75% discount on V4-Pro to developers first until May 5, and then extended to the end of May, and cut prices for input-cache hits across its API lineup to one-tenth of the original price. But I do not think pricing is the most interesting part anymore, because there are already many pieces making that point. I think the more interesting point is what V4 says about China’s open-source AI system and how it is slowly justifying the question that is asked the most: why?

Open source is not just an ideology here. It is an R&D efficiency mechanism. Because Chinese labs are compute-constrained, capital-constrained, and increasingly trying to build around domestic hardware, they cannot afford to waste as many resources duplicating the same low-level/infra-level research. If one lab solves an architecture problem, an inference optimization problem, or a Huawei adaptation problem, the rest of the ecosystem can absorb it. Even if they do not necessarily get the full training recipe, data mix, serving kernels, or CANN-specific optimization code, they still get a valuable reference point.

In that sense, DeepSeek is not just competing with the other Chinese labs. It is also providing shared infrastructure for them, whether out of the kindness of its heart or from some pressure from up top.

Jensen Huang has been talking about AI as a five-layer cake: energy, chips, infrastructure, models, and applications, and reiterating the importance of being plugged into China’s ecosystem. Nvidia’s own framing is that every successful application pulls on every layer beneath it, from models all the way down to energy. And with the recent push of its open-source models, it is again signaling to the industry that open-source models will benefit the ecosystem. This seems to keep getting lost in translation. The usual interpretation is about how Nvidia touches multiple layers of that stack and thus can sell more of its compute, but I think it’s beyond that.

What V4 shows is that in the China AI sector, even inside just one layer — the model layer — the labs are already starting to behave like a more coordinated system. The model layer itself is becoming a kind of shared R&D layer that helps propel the rest of the stack forward.

That is the frame I want to use for V4. Not “DeepSeek is cheap, therefore everyone else gets hurt,” although that is partly true. The more interesting story here is that DeepSeek is pushing base-layer innovation into the open, and the rest of the Chinese AI ecosystem is reorganizing around that.

TLDR, my key arguments:

1/ Chinese AI labs are starting to look like specialized units inside one compute-constrained mega-lab. DeepSeek does the base infrastructure work; others specialize in coding, agentic workflows, multimodality, distribution, and product integration.

2/ Open source can still make money. Users are not just paying for model weights. They are paying for managed inference, reliability, latency, maintenance, routing, memory, tools, and the whole agent harness. Just like how open-source software before made money.

3/ The bigger role of open-weighted releases is shared R&D. If one lab solves an architecture, inference, or hardware-adaptation problem, the rest of the ecosystem can study the released model, benchmark it, distill from it, and optimize around it instead of wasting scarce compute rediscovering every piece from scratch.

4/ The Huawei adaptation should be viewed through this lens. Making V4 run well on Huawei’s stack took real R&D effort, even if it’s only just inference first. The read I have heard from Chinese AI circles is that DeepSeek delayed V4 by almost three months to make this work properly. If true, DeepSeek basically took one for the team.

5/ The pricing squeeze on Z.ai, MiniMax, and Kimi is real, but secondary. It may also be temporary, because these labs can incorporate DeepSeek’s innovations and lower their own cost structure over time.

Open source can still make money

There is a common question around China’s open-source / weight AI ecosystem: how do you make money if you open-source the model?

The answer to this a year ago was that there is no pressure. Then the pressure came, and the answer was through other functions through superapps. Then the thinking was to sell cloud and then APIs. But now it finally hit me, I think the answer is pretty simple. Open-weight does not mean free inference. It means zero model-provider take rate if the user self-hosts, but how many users are realllllyyy self-hosting?

Self-hosting a frontier-scale model is not easy. Users still need hardware, power, memory, depreciation, utilization, maintenance, upgrades, latency optimization, reliability, and the whole inference stack. And as AI becomes more agentic, the harness around the model becomes even more important: routing, tools, memory, retrieval, workflow orchestration, monitoring, security, all of it.

If you’re like me, you probably asked, ‘If it’s all open, then why are there quoted variations of API prices?’ BECAUSE most users are not really paying for access to model weights. They are paying for managed inference.

This is basically the open-source software model. The code can be open, but the managed service is where the money is. That is why open-source model companies can still monetize. They can still charge for API usage. They can still sell enterprise deployment. They can still sell private cloud, reliability, support, workflow integration, and model customization.

But open source changes the ceiling on monetization. If users have a strong open model available, and if sophisticated buyers can self-host or compare against a self-hosting cost floor, then the model provider cannot charge unlimited premium pricing for generic inference. So open source does not destroy monetization. It disciplines it.

That is the right way to think about V4 and what it does to the likes of z.ai, Minimax, and Moonshot. Reuters reported that DeepSeek V4’s models are available as open-source releases under a permissive MIT license, allowing companies to freely use, modify, and commercialize them, thus making the engineering behind V4 more valuable to everyone else.

Open source as shared R&D

But the more important point is not monetization. It is R&D efficiency for a whole ecosystem.

If every Chinese lab had to independently rediscover the same architecture tricks, inference optimizations, memory improvements, and hardware adaptations, that would be a huge waste of scarce capital, talent, and compute. The more efficient equilibrium is that one lab makes a breakthrough, releases it, and the rest of the system absorbs it.

This is what China’s open-source ecosystem is starting to look like. Even Open-weight models can still function as shared R&D, but the transfer is not total. Other labs get weights, architecture clues, benchmark behavior, technical reports, deployment experience, and a reference model to study.

When DeepSeek releases V4, other labs do not just get a competitor. They also get a base to study. They can look at the architecture, copy what works, improve their own inference efficiency, and spend less time repeating the same base-layer experiments. That lets them spend more time on higher-level product behavior: coding reliability, agent orchestration, workflow integration, multimodal usefulness, enterprise deployment, and post-training.

This is why I think DeepSeek is becoming the shared foundation-layer R&D engine for Chinese AI. Just as it was implied here before.

This is something I have heard directly from people working at Chinese labs. They describe it as shifting more R&D resources toward “higher-level” innovations — coding behavior, reinforcement learning, agent orchestration, workflow reliability, product-specific usefulness — while waiting for DeepSeek to open-source more of the “lower-level” infra-layer innovations.

As an analogy, imagine OpenAI, Anthropic, and Google DeepMind each focusing on one part of the research stack and then sharing their outputs with one another. That is not how the US system works. But it is, to some degree, how the Chinese system is evolving.

And we go back to “necessity is the mother of innovation.” China has less access to frontier Nvidia chips. Compute is scarce. Capital is not infinite. Talent is limited (labs are 1/10 of the headcount compared to leading labs in the US). So the ecosystem has less (no) room to waste. Open source becomes a coordination mechanism. And DeepSeek, whether purposely or not, is betting on the fact that the rest of the model companies will incorporate its engineering innovation in their next iterations.

China’s labs are specialized units

If you take a step back and look at each and every one of the labs, they look like they’re each pursuing a different strategy these days. This could be due to skewed talent, founder taste, or pressure to show differentiation as they’ve gone public. But put together, they’re like a mega lab, almost like each a business unit.

Again, these companies are not literally one company. They are still competitors. But under the open-source layer, they increasingly look like specialized units within a single large, compute-constrained R&D system.

DeepSeek is doing foundation-layer architecture, efficiency, and now Huawei-stack adaptation. Z.ai is leaning into coding and enterprise reliability. Coding is one of the few AI workloads where the process is structured enough for agents to matter, and it is one of the clearest enterprise wedges. Kimi is leaning into long context and agentic orchestration. The Kimi product experience has always been more about getting the model to handle long, messy, multi-step tasks in a useful way. MiniMax is leaning into multimodality, voice, video, and low-cost inference. It is not trying to win only by having the best text model. Alibaba and Tencent are leaning into distribution, cloud, workflow, and product integration. ByteDance is leaning into consumer distribution, Douyin, Feishu, CapCut, and short-video workflows.

So the Chinese ecosystem is not just “a bunch of model labs.” It is a set of labs, each trying to own a different part of the stack. And because much of the base model layer is open or semi-open, each lab can absorb what the others discover and then specialize further.

This feels very different from the US system. In the US, OpenAI, Anthropic, and Google each try to own as much of the full stack as possible. They each run their own experiments, hit their own dead ends, optimize their own hardware relationships, build their own inference stack, and keep most of the output private. That structure has advantages. It allows fast private compounding. It protects the research frontier. It lets the winning lab capture more value. But it also means a lot of repeated work.

China does not have the same luxury, if you must, so the system is becoming more modular. DeepSeek pushes the base infrastructure forward. Other labs build on top.

I think that is the bigger picture that is being overlooked. When I started this piece, my first thought was that DeepSeek hurts its peers commercially by squeezing their Anthropic-like tiered performance/ usage pricing models. But as I processed the information, I realized that’s only the story in the short term; rather, in the longer term, DeepSeek will once again propel its technical capabilities, and the pricing model will readjust.

Huawei adaptation is also shared R&D

Now, if you view the Huawei part of this story through the same lens. We can see that making a near frontier model run well on Huawei’s stack is not just a political move or a ragey response to export controls, nor is it a clickbait press release. It takes real engineering work. It means dealing with CANN, kernels, memory layout, numerical formats, inference bottlenecks, latency, throughput, compiler issues, and all the tedious software-hardware co-optimization work that developers normally prefer to avoid if Nvidia is available. This is what Jensen was saying: that most developers want to use CUDA and that it took them 10+ years to build that kind of stickiness out of habit, due to quality of service, and so on. It’s pretty safe to say that most developers think running on Nvidia is just easier.

So, the read I have heard from people in Chinese AI circles is that DeepSeek delayed V4 by almost three months to ensure the model would run properly on Huawei’s stack before launch. If true, that is a very important signal.

Most other Chinese labs do not have the luxury to do this. They need to launch, show progress, monetize, raise money, keep customers engaged, and prove that their model is still competitive. If they have a strong model ready, the rational commercial move is to ship it. Spending three to six extra months trying or even attempting to switch to Huawei adaptation is too costly.

In this sense, DeepSeek kind of took one for the team. It absorbed the painful adaptation work, then released a model that gives the rest of the Chinese ecosystem a reference workload for Huawei. That is not just good for DeepSeek. It is good for Huawei, for CANN, for Chinese cloud providers, and eventually for every Chinese lab that wants to reduce Nvidia dependency.

As we know, a key change from earlier DeepSeek releases is that V4 inference was adapted for Huawei’s most advanced Ascend AI chips. It is also reported that Huawei said V4 is fully supported on its Ascend 950-based supernode clusters, and that the entire Ascend supernode product line now supports the DeepSeek-V4 series models.

So this is where it goes beyond the model layer and crosses to the next layer in the cake, the hardware layer. In Jensen’s cake metaphor, the model layer in China is now actively pulling the chip and infrastructure layers forward. DeepSeek is not just releasing weights. It is helping create the workload around which the domestic hardware stack can improve. This again explains why open-source? Now they can all try it out.

The pricing squeeze is real, but not the main story

As I said, this piece was initially 80% about the pricing squeeze, but then I realized that it is only temporary. But this does not mean the pricing angle is wrong.

V4 still pressures Z.ai, MiniMax, and Kimi, especially in text, coding, and agentic workloads. Any lab trying to charge a premium for generic reasoning or coding now has to explain why customers should not route more volume to DeepSeek, or at least use DeepSeek as the reference price. Before V4, the Chinese frontier labs were trying to move pricing higher - quite literally, all of them did it. Compute was tight, demand was growing, and investors wanted to see monetization. Some raised prices to make more, some used price as a way to filter out users because they could not serve as many as they’d like to. Not to be too repetitive, but because V4 is open, other labs can study its architecture, adopt its inference optimizations, and reduce their own serving costs over the next few months. So DeepSeek is compressing peers’ pricing today but also giving them the technical path to lower their own costs tomorrow, and raising the ecosystem’s technical floor.

Furthermore, the question is not only whether V4 beats Kimi, GLM, or Qwen on a specific benchmark. If V4 has better inference efficiency, the rest of the ecosystem can learn from it. If V4 has better agentic behavior, the rest of the ecosystem can post-train around it. If V4 is optimized for Huawei, the rest of the ecosystem now has a stronger reference point for domestic deployment.

The US contrast

This kind of partial switch from CUDA to CANN takes serious effort. For example, Anthropic has spent serious effort making Claude work across multiple compute platforms. It announced an expanded Google/Broadcom partnership for next-generation TPU capacity, and it has separately described deep technical collaboration with AWS on Trainium, including writing low-level kernels and contributing to the AWS Neuron software stack. Amazon also said Anthropic will secure up to 5GW of current and future Trainium capacity, and that Amazon and Anthropic engineers communicate on everything from low-level optimization work to high-level architectural decisions for next-generation chips.

That is strategically rational for Anthropic. The difference here is that no one would expect Anthropic’s Trainium / TPU optimization work to become shared infrastructure for OpenAI.

In the US right now, hardware optimization work becomes part of each lab’s private moat. In China, DeepSeek is doing the Huawei adaptation to not just benefit DeepSeek. It potentially becomes a reference point or even a standard for the rest of the ecosystem.

To be clear, I am not saying China’s structure is automatically better. The US system still has deeper compute, better chips, more capital, more mature cloud infrastructure, and the leading closed frontier labs. But China’s system may be more R&D-efficient per unit of scarce compute. And it is working more closely like a team, and again, under constraint, that can make a huge difference.

Why DeepSeek can behave this way

Most independent labs cannot behave this way, obviously, nor are they that just ‘kind-hearted.’ They need to launch models, acquire users, raise capital, show revenue, and monetize API demand. If they have a working model, the rational move is to ship it.

DeepSeek seems to have more room to optimize for ecosystem strategy rather than immediate monetization. That is why it can delay a release for Huawei adaptation, publish foundation-layer improvements, keep API prices low, and still operate as if the broader ecosystem benefit matters.

We’ve written about how the founder is philosophically quite committed to open-sourcing frontier technology, but in many ways, we can interpret DeepSeek as becoming somewhat of a national strategic asset in function, even if not formally nationalized in ownership. Since it’s backed by High Flyer, the quant hedge fund that is no short of dough, and it is said to continue to serve large enterprise clients. It is under no pressure to make a quick buck now, with its recent round of valuation at ~20billion USD, it is absorbing some of the infrastructure burden for the whole system.

Big tech benefits from this

Now, how do the big tech players fit into this? Tbh they’re kinda in the background and not that important for this strategic piece, except for the fact that they’re all financially linked to the labs (rumors are Baba and Tencent will invest in DeepSeek’s next round). What they are good at is that they are structurally better positioned to diffuse AI than standalone labs.

Alibaba, Tencent, and ByteDance do not need to monetize only through tokens. They can use cheaper models to improve cloud, commerce, ads, payments, enterprise software, productivity tools, games, short video, search, browsers, mini programs, and developer tools. If DeepSeek makes the model layer cheaper, that is not necessarily bad for them. It can actually be useful.

They can integrate whichever model works best at the time. Tencent did this with DeepSeek before, and it can do it again. One caveat is that Tencent’s latest HY3 preview was released in the same week as DeepSeek, and the limelight seems to be completely stolen.

Independent labs are in a harder position. Z.ai, Moonshot, and MiniMax have to compete more directly on model performance, coding, agents, hallucination rates, private deployment, or specific consumer usage. They do not have the same built-in cloud, payments, commerce, social graph, or short-video distribution. So that makes them more exposed to model deflation.

The duality of DeepSeek

So the real lesson from V4 isn't just that DeepSeek is cheap. It hurts other Chinese labs as competitors, but helps them with infrastructure. I think the bigger story here is that China’s open-source AI ecosystem is becoming increasingly coordinated, and each has a role to play in propelling the whole sector forward, given compute constraints, in efforts to optimize R&D.

DeepSeek pushes the base infrastructure forward. Other labs absorb the work and specialize further up the stack. Huawei gets a serious reference workload. Standalone labs face pricing pressure today, but also have a path to lower their own costs tomorrow.

Anyway, Part 2 is about V4’s implications on Huawei and Nvidia and what this symbolizes across the AI stack (the five-layered cake).

Some relevant reads that helped me in understanding the pricing economics better and DS analysis:

• DeepSeek V4 preview release. DeepSeek’s official release page on V4-Pro and V4-Flash, open weights, parameter counts, and 1M context.

• DeepSeek V4 technical paper. Architecture, hybrid attention design, MoE structure, post-training pipeline, 1M-context efficiency, and the 27 percent compute / 10 percent memory comparison versus V3.2.

• South China Morning Post. Huawei and DeepSeek collaboration, day-zero Ascend adaptation, and DeepSeek’s throughput caveat until Ascend 950PR ships at scale.

• Reuters. DeepSeek V4 adapted for Huawei chips, V4-Pro and V4-Flash positioning, Huawei support on Ascend 950 clusters, and Huawei’s claim that its chips were used for part of V4-Flash training.

• Bernstein, “China Internet: A primer on AI token economics and inference margins,” Robin Zhu et al., 21 April 2026. Q1 2026 compute price hikes, GLM-5 vs M2.5 cache pricing and hit-rate data, Qwen 3.5 cost-vs-intelligence trajectory, Alibaba Token Hub, “low-cost agentic back-end” commoditisation framing.

• Bernstein, “China Internet: AI inference margins addendum; DeepSeek-V4, HY3-preview thoughts,” 27 April 2026. V4-Pro and V4-Flash blended pricing, 75 percent discount within 24 hours of launch, MiniMax throughput reconciliation, Tencent Muse Spark framing, DeepSeek fundraising base case.

• Jukan / Citrini, X thread. The argument that V4’s architectural pattern aligns with Nvidia’s Rubin and G3.5 roadmap rather than simply eroding it.

• Jensen Huang on the Dwarkesh Patel podcast. The argument that US export controls accelerate the development of a parallel Chinese chip ecosystem.

• Tencent HY3 preview official release. HY3 architecture, open-source release, Tencent ecosystem integrations, CodeBuddy / WorkBuddy improvements, TokenHub pricing, and inference efficiency.

Matt lived and worked in China from 2010 to 2016 and later led China tech research at the Paulson Institute’s MacroPolo. He’s the author of The Transpacific Experiment. He speaks Mandarin, and he turns complex policy into plain English.

In this episode, he helps us understand China’s AI governance, about how Beijing is thinking through the social and political consequences of rapid AI adoption. We focus especially on a shift that became more visible in early 2025: rising concern inside China’s policy community about AI’s impact on jobs, worker anxiety, and social stability.

Matt explains why China’s AI labor question is different from the Western debate. We also discuss how the Chinese government is trying to balance support for technological progress with the need to manage public anxiety, clarify labor rules, and avoid social instability as AI becomes more deeply embedded in the economy.

He broke down the myths, explained the jargon, and the regulatory bodies in China. Our conversation started slow, but it became very, very heavy, what they call 干货满满 substance heavy. Also, a shoutout to ’s work in helping us better understand the open-source ecosystem and ’s for helping us understand investing in China AI!

Subscribe now

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, founders, and product managers.

For more information on the podcast series, see here.

To find the previous episodes of Differentiated Understanding, see here.

Chapters

00:00 Introduction to AI Policy in China

03:10 Matt Sheehan’s Journey into Chinese Tech Policy

05:55 Shifting Perspectives on AI and Labor

09:02 Public Concerns Over Job Security and Government Responses

15:09 Education and AI: Preparing for the Future

17:50 Regulatory Landscape of AI in China

34:00 Navigating China’s AI Regulatory Landscape

40:58 Misconceptions About Chinese AI and Government Funding

43:57 Understanding AI Safety and Security in China

52:03 Global AI Governance: Cooperation or Parallel Paths?

AI-generated transcript

Grace Shao (00:01)

Hi Matt, thank you so much for joining us today. I’m so, so happy to finally have you on the pod for people who are listening. We’ve been trying to make this happen for like six months, but between us, there are like three little children running around with a bunch of viruses and have just not been able to make this happen. I’m really excited. ⁓ A few months ago, what really caught my attention about your work again is that you shared something on WeChat saying you were dissecting the new Chinese AI safety paper, like the big national one. ⁓like verbatim in Chinese. And I was like, wow, this is extremely impressive. It’s not an easy task. I commend you for doing that. So I really wanted you to help us understand the nuances of the AI policy world, especially how people are perceiving AI in China. I think there’s more more interest in how China’s governing AI ⁓ while we were hearing the backdrop of how the Chinese government is trying to push on AI diffusion, right? And then on top of all of this, like where areas where China’s AI governance seem to be leading, because in many ways it seems like

China’s AI regulators are much faster to respond to how fast technology is evolving. But to start, we would love to hear about your personal story. Tell us about how you ended up studying China, studying Chinese tech policy. We met in Beijing years ago, maybe a decade ago. ⁓ Yeah, so tell us about that.

Matt Sheehan (01:11)

Sure.

Yeah.

Yeah, sure. Sort of stumbled into China stuff. I hadn’t taken Chinese or really knew anything about China until about halfway through college when I ended up getting a summer job in Beijing. I was just kind of like instantly fascinated and knew I wanted to move back there after I graduated. So took a little bit of Chinese my senior year, moved to Xi’an, taught English, kind of followed what at the time was a very like typical

know, trajectory of like, go there, teach English and then go study Chinese at university and then get a job and get a slightly better job. And eventually I was able to kind of wiggle my way into journalism. And so I was a China correspondent for a publication called The World Post at the time. And that took me up. was there from 2010 to 2016. So kind of like the hinge period before and after she came to power. Pretty interesting thing to see. And

When I moved back to California in 2016, I started working on a book about China-California ties. I’m from California and this was like the period of kind of explosion in cross-border investment and Chinese students come into California in the Silicon Valley-China relationship getting even more like twisted and complicated. China-Hollywood. So I wrote a book about that and as I was doing it, the kind of the tech section, the China-Silicon Valley, China-U.S. tech connections kept growing bigger and bigger and I ended up

working a little bit with Kai-Fu Lee on his book, AI Superpowers, which was kind of my turn from like, it was like all things China, China, California, China, Silicon Valley, China AI. And since 2017, I’ve been working almost exclusively on AI issues in China. Maybe the first three years of that, like 2017 to 2020, was very focused on comparative capabilities. This was kind of a period right after the National AI Plan in China when there’s a big explosion in activity. And I think...

This is kind of was like the first time America kind of got freaked out about Chinese AI capabilities. And so I spent a few years being like, okay, let’s try to like ground these assessments in some data. Let’s get like an actual grounded sense of where the countries are with each other. ⁓ And then starting in 2021, I sort of turned into focusing on Chinese AI governance, Chinese AI regulations. That’s when they first started rolling out their regulations or recommendation algorithms and sort of deep fakes. And I was kind of making a bet that I think

If China continues to be at or near the frontier of AI, then how they choose to regulate it domestically is going to have huge implications for China’s own ecosystem. And then it’s going to really ripple out internationally on safety, security, growth, all this stuff. So I spent the last, now it’s like five years, ⁓ just deep in the weeds of Chinese AI policy and regulation.

Grace Shao (04:04)

Oh, great. I think I definitely want to double click on all the algorithm security and the kind of, you coined the answer versus what’s called security and versus what’s the other Chinese word? Yes, yes.

Matt Sheehan (04:17)

Anshun safety security. ⁓

Jeff Ding was talking about this very early on, but yeah, it’s a, it’s a constant thing that we have to negotiate for people who don’t know it’s the word Chinese, the Chinese word Anshun ⁓ means both safety and security. whenever you’re kind of translating documents on this front, you have to know, you talking about AI safety, which is kind of a different thing versus AI security, right. I think both you and Jeff definitely are some of the more nuanced scholars I follow. And I do want to kind of double click on that later on. But to start, I think I want to talk about something that’s top of mind for a lot of people. You just wrote a piece that you said was not super serious. It was just your scattered thinking put together on subset. I thought it was very well written about the growing anxiety around potential job losses. ⁓

your perspective is that you know there are more and more people voicing this kind of concern and I wanted to hear a perspective on that and I kind of wanted to share a bit of my different share my different perspective on this and what I’m hearing on the ground and kind of have a conversation around that as well. Yeah why don’t you start with sharing like what you found yeah

Matt Sheehan (05:22)

Yeah, sounds great. Yeah. So, I’m not an AI and labor person. That’s not been my focus for a long time, but I’ve been monitoring it for a long time and just lightly. And starting around, I guess it was early 2025, I just started to hear a lot more out of the Chinese policy community about worries about AI’s impact on labor and jobs. And this was kind of a surprise to me because ⁓ just a little bit prior to this, say early 2024,

I had, I sometimes ⁓ in my job, I run these kind of like informal surveys or almost like a, what do call it, focus group of American and Chinese AI policy people and asking them like, you how would you rank these different risks? How concerned are you about ⁓ job risks versus privacy versus military AI? And we have both sides that like rank the risks and then talk about the results. And when I ran one of these in early 2024, it was very striking that the Chinese side

I think it was at the time we had seven different risks and the Chinese side ranked labor impacts a second to last as six out of seven. And so my sort of baseline was like, OK, for a variety of reasons, this isn’t really too on the radar of China’s ⁓ policy community or wider policy community. And then starting around early 2025, some of those same people who I had been talking to about this before had really changed their thinking. They were saying that there was a big change in thinking within

China, maybe especially within China’s of policy and government circles, but then I think also a little bit wider. And so that sort of sparked my curiosity. And for the past, now it’s like over a year, I’ve been just sort of tracking when does this AI and labor question show up in state media? When does it show up in kind of the online discourse? When does it show up in policy documents? And sort of the TLDR is like, I think this is really, really ramped up a lot. Over the past 18 months. It’s been maybe the single biggest change in how China perceives different sort of risks as it relates to AI. And I think I’m looking forward to sort of discussing how maybe like the policy world or the government’s perception of this differs from ordinary people or certain categories of people. ⁓ But I think it, from my perspective, it’s sort of it’s infused into both. think there’s been a fair amount of public concern.

the policy community picks up on that and they want to both respond to like the actual problem, know, actual job losses, but they also really want to respond to people worrying about job losses. That’s kind of maybe the thing that actually made me write this piece now was discovering an interesting piece in state media. I think it was in science and technology daily. That was the headline was like ⁓ AI must be controllable, but people’s ⁓ anxiety about AI must also be controlled.

And it was all about how sort of OpenClaw has triggered a lot of anxiety and a lot of people about, are they going to get replaced? You need to be building your own AI agent in order to not be left behind. And they’re sort of trying to tamp down those concerns in a few ways. So that’s what sparked the piece itself.

Grace Shao (08:33)

Yeah, I think definitely what you saw and you wrote about is like definitely kind of playing out in the China AI policy ecosystem that I see as well. And I think for sure, the open-claw frenzy have kind of opened the eyes to lot of the even average people what AI could potentially do. However, I guess my argument, not against it, but it’s just like, you know, we kind of cite each other’s work on subset. But my point was kind of saying, you know, this is a reflection of a relatively elite group of people end of the day, because the knowledge work economy in China end of the day is only like only 30 % of the workforce actually are the knowledge working economy. And end of the day, even though it’s 30%, because China has such a huge population, the mass, the sheer scale, it feels really large. However, I want to bring it back to the idea that like anyone who’s lived in China understands that the government’s like top top priority really is about social stability which leads to what they call social harmony, right? And I just think that, you know, the rising anxiety of job control a lot of times maybe is because there’s a fear of if there’s a lot of disruption to jobs then people will lead to social unrest which obviously gets a bit more sensitive but you know, a lot of what they do comes from that I guess thinking so I agree with you top down definitely have to understand what’s happening with technology and how advanced AI has become in 18 months.

⁓ have given them kind of, I guess even fear mongered a little bit internally, right? ⁓ But the nuance here is that I think ⁓ the rest of the 70 % of the Chinese workforce actually don’t work in anything structured that we know. I think even probably even 80%, you know, people in China, they most of them are actually like, you know, service providers ⁓ from, you know, rural areas and urban areas. A lot of people work in factories, even the entrepreneurs, right? They run like say hospitals, clinics, factories, bottle cleaning, like factories, whatever, right? ⁓ Car logistic rental businesses, these people aren’t actually ⁓ trained in the way that maybe the West by default think they are. They actually just run it from a grassroots way and they don’t have very, very streamlined processes. They don’t have documentation. They don’t run like what we think a corporate has run. So in that sense, I think it’s very hard for AI to replace any of their workflow.

because it’s actually not a, we can’t really provide context and a lot of things, business is done is through one C, is through a wink, through a look, through a gesture, through, you know. So a lot of that, I think, in fact, will be harder to replace than even maybe some of the more mature businesses in the West where there are structured processes and everything. So that’s kind of my, I guess, a more nuance, I think, push on that. Yeah, wonder what you think of it.

Matt Sheehan (11:20)

Yeah. Lots of thoughts. And I think the sort of the fundamental distinction that you’re pointing at is very valid in that like, you know, companies in the West, in the United States, they’ve been like big companies have been running sort of digital databases for decades. They have like decades worth of data. They have pretty advanced like enterprise software. It’s just a much more I want to say something like a bias, but it’s a little more like put together sort of official structured

⁓ technological backend and not just technological, but like a process backend. Whereas in China, it’s just things have developed really quickly. A lot of it’s on the fly. lot of it is, know, enterprise software is just not, there’s not really a market for that in China in the same way. It’s a lot of stuff is pirated or they’re just not digitized in the same way. And it’s actually very interesting. This is in kind of the early days of the like China, US who’s ahead. ⁓ know, a lot of the debate focused on data.

Matt Sheehan (12:18)

And there was this idea of China has a billion people, so it must be this huge advantage in data. But my pushback on that was always kind of what you’re arguing, is like the US actually has very structured data, and it’s owned by corporations, it’s deployed by them, they’re already doing type of sort of lower end market intelligence type stuff. ⁓ So I think that that sort of backdrop is very real. think ⁓ maybe from there I’d like differentiate out to potential risks or debates. And it’s kind of what you were pointing out as well. There’s like the actual question of how many jobs are going to be impacted. How many people are, what is it going to do to people’s wages? What’s it going to do to aggregate employment? And then there’s the question of like, ⁓ how do people think about that? What are the fears due to sort of the Chinese social stability, even if the things haven’t manifested. So I think separating those out, definitely the government is... ⁓

their sort of initial response is related to public worry about this. So in the piece, I detailed the way that ⁓ sort of a robo taxi incident in Wuhan was in many ways the spark that really like ramped up the government thinking on this. And this is something that I heard from a couple of different Chinese policy people who both pointed to this incident, which I had totally missed at the time and wasn’t like major international news, but that had a big impact. And basically what it was is that

Baidu was rolling out its sort of fully autonomous robot taxis throughout the city of Wuhan. There was this kind of like public letter, ⁓ open letter released by a taxi company that was kind of railing against, you know, both ride hailing platforms and autonomous vehicles as, you know, stealing the iron rice bowl or just smashing the rice bowl of taxi drivers and of companies. And even though it was a kind of like a small thing, a couple of days later, a Baidu taxi actually hit pedestrian, I don’t think they were seriously injured, but it kind of fed into this overall, a big kind of online reaction and discussion about like, what’s going on with AI? Is it going to take people’s jobs? And it, it’s one of those things, it’s funny to explain to people because it sounds like nothing, but it did lead to a pretty significant ⁓ imprint on the way that the Chinese government is thinking about it. So it was coming from, in many ways, public discussion of it. Like the discussion was happening online. This discussion might be happening among, you know, elites.

chronically online people. ⁓ But it’s something that the government definitely picks up on. So I that’s one element. They’re worried about the worries and they want to, like among their sort of policy reactions in a ways, thus recently has been sort of directing platforms to say like, you kind of need to tamp down these articles or these viral videos that are telling everybody, like if you don’t adopt open claw, you’re going to be left behind. There’s been this kind of rash, both in China and here in the US of like,

you know, kind of like fear mongering people into clicking and taking your course on building agents or just subscribing, whatever. And one thing the government is doing is telling you like, chill on that. Like, don’t be putting that narrative out there. ⁓ And so that’s part of this kind of like public opinion management thing. In terms of the actual impact on jobs and who will it hit?

It’s a huge open question that I personally have gone back and forth on for years. I first kind of did a deep dive on this way back in 2017 when everything was still so speculative. At the time, was pretty not worried in part because of the reason you described it. I’m like, there’s just a lot of friction. There’s just so much friction in this economy. And just because an AI system can theoretically execute a test doesn’t mean it’s taking a person’s job. And for me personally, my thinking on this has changed a lot in the last

year to 18 months, mostly because of how capable agents have proven to be. I expected agents to essentially be hitting a lot more roadblocks while they’re being deployed online. They haven’t been. They’ve been operating much smoother, or just they’re more relentless, and they can break through these bottlenecks. ⁓ It’s interesting that in China, the inciting incident was not about white-collar workers. It was about taxi drivers. ⁓

I don’t know this as a fact, but if I had to guess, I would guess that a much larger portion of the Chinese population’s job is driving a car or driving a scooter or something like that. And that maybe that’s a vulnerability that they might face depending on how self-driving vehicles roll out or delivery robots and stuff like that. The knowledge workers, yeah, it’s such a messy and unclear thing, but I think the government has at least started to take it seriously because it’s not, their policy responses are not just this like,

public opinion management stuff. They’re also talking about, ⁓ like one of the more interesting pieces that I highlighted is, and maybe the most concrete thing they’ve done so far is, ⁓ according to Chinese labor law, there’s sort of reasons why you can and cannot fire a person. There’s like legitimate and illegal reasons to fire someone. And when someone is fired and they object, this gets taken to like a labor law mediation ⁓ body that’s under the Ministry of Human Resources and...

forget what the second part social ⁓ security. ⁓ Yeah, Ren Li, Ziyuan, Shouhui Bao Zhang. Yeah, that’s what it is. ⁓ And in the last year, one of the things that really made a lot of made kind of a big splash is that those mediation bodies declared and it was echoed in like the biggest state media that saying that you replace someone with AI that AI can now do this person’s job is not a legitimate reason to fire someone and those people have to be reinstituted into their jobs.

That’s a concrete policy thing that’s actually directed quite clearly at like actual impacts. ⁓ Is it going to work? I really don’t know. It might just be a little bit of friction and, you know, maybe China’s kind of doing what it always does, which is like, we’ll figure this out. We’ll kind of muddle through this. We’ll put some friction here. We’ll grow a little more here. But I think the concerns are real, whether they bear out, ⁓ whether they hit faster in the United States or China, which country is better positioned to sort of roll out a more redistributive welfare system. think these are all open questions, but I think the concerns at least are real.

Grace Shao (18:24)

Yeah, and I think you hit something that I feel like it’s being kind of missed in the headlines, which is the government actually cares more about the general mass, which are the people who are driving the scooters and the like the DDS cars more than the knowledge workers, which is kind of different from the Western kind of conversation right now, where a lot of the whether it’s fact, frankly, the power that can lobby and the power that the voice that have the voice are all really concentrated in.

the white collar elite jobs that are very much concentrated in Silicon Valley and whatnot, right? And I think it really reminds me of the time when, you know, during the Hulianmang Shidae, like the internet era, you know, like the big tech only really got clamped down when the average consumers felt like they were really being pushed to R-Shrine Egypt 2-1. So that’s when they had to add the monopoly to probes. And then soon after, only maybe two years after the probes happened, there was a common prosperity rollout, which basically all the big tech in some capacity had to like showcase that they had a CSR aspect to them. I think this is something that we don’t really see in the US as much with all the big techs, because it’s kind of like they’re doing what they need to do. They have their profit driven interests. And then of course, everyone has a CSR, but it’s not really allegiance to the government CSR mission. It’s more like, we believe in ESG. We believe in climate. Amazon is going to have some like carbon footprint reduction plan, right? Whereas like the common prosperity thing rolled out. ⁓

you know, it kind of died on its own, like no one really talks about it anymore. However, during that phase, when it did get rolled out, it was like an understanding where, OK, if the government wants the, frankly, the poor or the middle lower class to feel protected, then you as a very large ⁓ moneymaker in the economy need to showcase that you are somehow ⁓ part of this kind of support. So I wonder how this will play out for the big tech in China when, the job protection policies really get rolled out in practice, like what you’ve mentioned. And obviously they can’t really say, you’re being replaced by AI. At least there’s that superficial guardrail there, I think.

Matt Sheehan (20:24)

Yeah, I think the sort of the political economy of these questions is going to be super interesting in both countries. know, essentially like business in many ways, it inverts the sort of technological impacts on employment that have been around for so long. Normally, like greater technologies integrated into the workforce, it hits sort of people working maybe low end manufacturing jobs or jobs that would be considered sort of repetitive and, you know, quote unquote, low skilled jobs, even if they’re not. ⁓

And, you know, in the United States, we’ve seen like three, four decades of this. And the people who are concerned about that basically didn’t get hurt because they are not, like you say, part of these influential classes. AI is going to, yeah, in the United States is going to be very different. This is going to be the first time that you have the way that my sort of mental model for it is like, if you’re a senator and you have kids or nieces and nephews or your friends, kids like what, what are their problems?

And like how close do those feel to you? And you know, if you’re a 60 year old senator and you’ve got like a 23 year old niece and she just graduated from college and she got a degree in, you know, something that’s like is legitimately employable normally like in marketing or something like that, and those jobs just aren’t there, that’s just going to feel very close to home for people in power in the U.S. in the ways that it hasn’t felt in past waves of technology impacts. In China, I...

partially agree with what you’re saying, but I think there is also going to be a significant element of this sort of the same dynamic as the United States. mean, yes, common, you know, Xi, common prosperity. He’s focused a lot on sort of eliminating extreme poverty and, you know, the CCP, it’s in its bones that like the rural, the working class are in many ways kind of the long term support base of them. But I mean, also if you look back at Chinese history, like a lot of the biggest and for the government most dangerous protest movements came out of elite schools, came out of students at elite schools who ⁓ either couldn’t find jobs or were facing inflation issues or had, for ⁓ ideological reasons. I think that stuff does hit close to home. think some of those same dynamics, if you’re a deputy director at the National Development and Reform Commission.

your family, the people that are close to you, are going to be the type of knowledge workers that are going to be impacted by this. And I think that just can’t help but kind of like compress in on the thinking on this. ⁓ You know, how AIs can... Yeah, yeah, and like how...

Grace Shao (22:58)

Yeah.

everything becomes personal in the end. Like in the end, it’s like

politics is still personal. Yeah. Sorry.

Matt Sheehan (23:08)

Politics

is personal and yeah, mean, like cities are where social instability is the most dangerous. Like cities are where people gather and you can have potentially dangerous incidents. These are the people who are very online and are sort of sparking or leading the conversation as much as that can be controlled and manipulated via censorship regimes or public opinion guidance. Like these people are gonna be vocal. yeah, I think if I was at the CSPI, I’d be concerned about

Grace Shao (23:39)

I want to kind of go into on education. Like you kind of touched on it, right? Like, you know, there’s been draft rules about children’s interaction with AI in China as well. There seems to be more guidance and obviously concerns around that ⁓ from at least from the top down ⁓ about their mental state, their dependency, or even what constitutes as an AI companion, how we should draw the line on that. We know like

famously a couple years ago, China installed this rule where like, you know, kids under 16 cannot actually play online games on their own without the parents consent. However, that you know, there’s obviously loopholes in practice. But again, it goes back to there are, you know, rules and laws in place to try to protect minors. How do you view all of this? ⁓ Because in the with the backdrop of China trying really hard to diffuse AI into the real economy. And then there is this pushback like you just mentioned on

concerns about AI taking jobs. I feel like there’s also almost like a ironic kind of contradiction happening where, you know, Tiger moms are like, okay, now we don’t need to learn math, we know how to learn AI. And Tiger moms are like saying, how do we optimize getting into, I don’t know, Harvard with AI’s help? And how do we get AI into the education, education, ASAP? I mean, honestly, we don’t even know what the education system might look like in like two decades.

from our kids, but at this point, seems like there is like embrace. I don’t know. How do feel about that?

Matt Sheehan (25:13)

Yeah, a couple strands there. One just on the sort of regulatory side, like this is a long term strand in Chinese tech policy and tech regulation. They always put a pretty heavy emphasis on like how are kids using technology. They have, ⁓ they sort of mandated having like a minors mode on various ⁓ apps. ⁓ This is the regulation I think that you’re referring to as the newly passed. ⁓

I translate as anthropomorphic AI. That’s the word that’s the official translation. ⁓ So it basically means AI that, you know, behaves like a human. This could include AI companions that are, you know, literally like a character pretending to be your friend. They could also include, you know, the way that people interact with chat, GBT or Kimmy or whatever, you know, the phenomenon of having AI boyfriends and girlfriends and all this stuff. So there was a new regulation on this that was just finalized, I think, last week and

It has some protections for everybody, specifically around ⁓ self-harm, addiction, and stuff like that. But it has really ramped up protections for minors and for elderly users. So there’s all these kind of specific add-on requirements. For ⁓ minors, it involves permission from parents. Parents can review at least some. They can set limitations on how the child uses the system. They can review.

conversation that might have got toned down a little bit in the new version. ⁓ But I think, this is many ways it’s the same concern that surfaces in the US and elsewhere, like California just passed a just passed last year, passed a similar regulation on AI chat bots that I think also had believe it had extra protections in there for kids ⁓ on the education side of things. I guess there’s a couple of things. One, there’s just like the yeah, you say the tiger mom’s like this is ⁓

It’s a booming industry of like, I’m going to teach your, you know, four year old AI so that they can use it because this is going to be how they get a job and how they get into school and how they get a job. ⁓ you know, a lot of it is bogus. Maybe most of it is bogus, but it’s very attractive to parents who have grown up in a really, really cutthroat competitive education system where you’re looking for every single edge that you can find.

And so that’s, that’s a piece of it on the, from the policy side, they have both sort of AI and education policies, AI plus education policies that they’re pushing in a bunch of ways. I have some friends who are working with teachers over there who are described to me pretty like sophisticated and interesting ways. The teacher that are using AI to lesson plan, to create like really interesting games that keep the kids engaged and learning stuff like that. So you have those, and then on the labor.

the labor side of things, they’re also viewing AI, they’re also viewing education as something of an antidote ⁓ to AI fuel job disruption. This is in the, I in the five year version, it is in the five year plan, it’s in the AI plus plan, it’s in a few other places where they say, we’re really gonna prioritize lifelong education. So maybe you used to be an accountant, you lost that job and ⁓ you’re gonna retrain as something else. ⁓

which I think is a good, you know, it’s a good attitude to have. If you’re a person, you should always, I’m always trying to learn, you know, books. Um, think they’re great, but I don’t know if at a totally like a, you know, macro population, I know if you’re going to get 500 million people to be constantly staying one step ahead of AI in terms of what jobs it can do now. I mean, a lot of the things that we would have told you go back like two, three years and say, what jobs are going to be disrupted by AI? A lot of the recommendations would have been totally backwards.

People would have thought that, ⁓ coding jobs are great. jobs involving creativity, ⁓ illustration, ⁓ stuff like that. AI can’t do those things. It can’t be creative in that way. And it’s like, that’s actually kind of what it’s best at now in some ways. I mean, you can argue about the level of creativity, but like generation of content, generation of images, videos, language. So I’d say it’s a piece of the Chinese sort of response on labor concerns.

a fad, but maybe like a ⁓ useful fad within like the sort of education industry. But I’m a little bit skeptical of this as like ⁓ an actual antidote to the disruption that I at least imagine is coming.

Grace Shao (29:49)

Yeah, it’ll be really hard to be like upskilling, like re-skilling like hundreds of millions of people. Like, it’s just, you don’t even have the capacity to do so if there’s actually mass disruption. alright. ⁓

Matt Sheehan (30:00)

mean, this was always the response, in the United States on coal miners. We’re going to teach them to code. everybody, all these manufacturing workers, we’ll offer like a job retraining program. I’m like, ⁓ maybe, yeah.

Grace Shao (30:05)

Yeah.

I’ll take generations. I’ll take generations for things to shift, know, resources to shift, people’s mentality shift, you know, for a while, like, when many, when remember the first wave of like, a basic rural kids no longer wanted to work in factories and wanted to go to urban cities, there was a surplus essentially service providers and then like everyone eventually became a DD driver or a food delivery man and then

Now we’re seeing a reshuffle in that population again, where people want to move back to their rural, cities. so I think based on how society is evolving, opportunities will arise without even us realizing anything, hopefully in the best case scenario, where people will find opportunities to reskill. ⁓ But I want to talk about something that’s a bit

Grace Shao (30:57)

I guess not heavy, but actually not many people understand, even including myself. So you really look at the government and the policy structure of ⁓ China’s regulators in the cybersecurity space and whatnot. ⁓ There’s so many players. There’s the CAC, then there’s the NDRC, there’s the MMIT. I can keep on naming acronyms, but can you give us a really, really quick high level understanding of who’s regulating whom? ⁓

how do they actually work with each other and are their KPIs aligned before we get into more about how you know, how China’s policy is shaping the technology and AI ecosystem.

Matt Sheehan (31:36)

Yeah, maybe I’ll do it. ⁓ I’ll introduce a couple of the players and I’ll do it somewhat chronologically in terms of like when have they become important or rise and fall in importance. So AI policy, like the first really big policy document was the 2017 National AI Plan. It was released by the State Council, effectively sort of China’s cabinet, sort of the highest level of government. But ⁓ people who are in the know say that that was largely sort of drafted and pushed by the Ministry of Science and Technology. So this is

really like the policy wave of like 2017 to 2020 more or less. And it’s the Ministry of Science and Technology and it’s also the Ministry of Industry and Information Technology, MIIT. So these are really the organizations whose job it is to promote science, promote innovation, and MIIT is more of like the industrial applications of the technology. So they were kind of in the driver’s seat in that period of time. They were the most relevant actors. They were the ones who were driving real activity.

starting in 2020, 2021, you had the CAC, the cyberspace administration of China, really like come to the center and become the most important actor in AI policy. The CAC, it’s basically the internet regulator. It was created in 2014. It was largely created to kind of like get the Chinese internet under control from a sort of political content ideology perspective. They’re connected to the Ministry of Prop, or the propaganda department.

publicity, as they say now. ⁓ So from 2021 through 2023, the CAC was the one rolling out these binding regulations on recommendation algorithms, on deepfakes, on generative AI. And these are the regulations that actually force companies to do things. They actually force companies to register their models, to do pre-deployment testing, at this point to label AI-generated images in different contexts.

There’s for that 2017 to 2020. It’s kind like the go-go period. Let’s just like push this industry forward. You have the Ministry of Science Technology, MIIT. And then from 2021 to 2023, it’s really the CAC. This corresponds roughly with the tech crackdown of 2020 through the end of 2022. That was a period when the CAC, the CAC is kind of at least historically, it’s kind like the bad cop of tech policy. They’re the ones who are like telling companies like come in and drink tea and we’ll tell you what you’re doing wrong or, finding companies in different ways. Cyberspace Administration of China, yeah, CAC. ⁓ Some people call it CAC. ⁓ And then sort of one of the more significant changes from 2023 to now is the rise of the NDRC, the National Development and Reform Commission, Chinese Fagawei. ⁓ And they are a macroeconomic regulator. They are like what grew out of the sort of state planning apparatus. And they’re

Grace Shao (34:01)

This is a cyberspace administration of China, right?

Matt Sheehan (34:30)

really powerful, they’re kind of a super, super ministry within the bureaucracy, but they’re not sort of directly, there aren’t that many direct connections to AI. They deal with, they deal a lot with money. They have money to give out for projects that funnels into compute projects and stuff like that. ⁓ But they wouldn’t be like who you would think of as the go-to AI regulator. What I was told and what I feel pretty confident ⁓ is what happened is that in some time in, I think, 2023, maybe mid to late 2023 and then into 2024, the top leadership in China essentially said, hey, we need a little more balance in our AI policy. The last three years it’s been led by the CAC. They’re kind of a bad cop. They’re really focused on controlling the technology, controlling the sort of output, the content, the ideology from it. And that is important. That’s kind of their first priority. But we need to rebalance this a little bit. We need to move out of our total tech crackdown era. And now we realize like our economy isn’t doing great.

We realized we’re behind the US after CHAT GPT came out, and we need to balance this out. And so they empowered the NDRC to be a of a coordinator across AI policy, someone who is intended to take the input from the various ministries, from Ministry of Science and Technology, MIIT, CAC, and to try to make it little more coherent and balanced. And so that’s kind of the role that they have played for the past few years. The details of how that works out are

shrouded in secrecy, you know, you hear little tidbits here and there. But there have been like visible manifestations of it. They had not, when they released these regulations, usually there’s a sort of a lead regulator on it or a lead policy document person on it. And then various other ministries, they co-sign it and they’re like listed below. NDRC hadn’t been on any regulations prior to 2023. And then starting in 2023, they were listed second as like the second sort of most important organ.

policy body on these things. essentially we have this kind of like 2017 to 2020 is this like go-go period. Let’s diffuse. Let’s push the technology. Let’s push innovation. 2020 to 2023 is this more constrictive crackdown. Let’s build the regulatory infrastructure for things. And then 2023 to today is just like, let’s balance this out. Let’s not be purely focused on the content and ideology concerns. Let’s also be thinking about development. Let’s be thinking about employment. The NDRC is actually allegedly one of the groups that is very concerned about the employment impacts. you know, tons more details that I will love to go in on, but maybe that’s a starting point.

Grace Shao (37:06)

No, I think it’s super, super helpful. I just understand the nuances of like what their actual KPIs even are and like, you know, who does what, how they work together. I think that’s really helpful for lot of listeners and even investors who are trying to follow the space and just confused by acronyms. But help me understand now, like you say that 2023 to now essentially is in the same kind of era. However, I feel like at least from the capital market perspective, you know, the last year might have seen a bit of a

shift again, you know, it was a bit of a let’s go AI, big tech AI, all the labs, let’s go, let’s go IPO. Then obviously the deals, some of the deals didn’t come through, some of the IPOs didn’t come through. ⁓ There seems like you even said people are being told to tamper down their excitement a little bit. Is that aligned with what’s happening with the policy side of things? Or is that actually more a reflection of just, frankly, you know, the AI space not being that exciting right now, you know, since the Gentic ⁓ kind of breakthrough. We’ve not seen more consumer and breakthrough. Also, there’s a lot of talk about, you know, there’s no obvious proof ROI on all the spending from all the big tech right now. Help me understand all that, I guess.

Matt Sheehan (38:15)

Sure. Yeah. When I was breaking down those errors, is largely its policy, but it’s already kind of like government attitude towards it. It’s like which, you know, they’re always in some ways swinging back and forth, going back and forth on the seesaw between, you know, control development, control development. And that 2023 to now being one era is sort of in that sense. It’s the period of rebalancing more towards development. There’s tons of sort of wiggles in that process and

things they’re pushing more and retreating on. But from a positive perspective, that’s the overlay. ⁓ In terms of like the capital markets, investments, I mean, I think this is kind of at least for people in the United States, it’s kind of like the one of the most misunderstood things about the Chinese AI ecosystem is that it is really like cash constrained, that it is not like the United States where, you know, open AI is just like sucking in.

the tens of billions of dollars from a huge variety of investors are just spending huge capital outlays, which people talk about, is it a bubble? Is this going to come back to bite them? That’s an open question. But in China, you don’t have the concern about that bubble because there just is not the same level of infusion of cash. when a couple of the companies did IPO recently, Z.ai, formerly Jerpool and Minimax IPO in Hong Kong, and I think I’m not

Grace Shao (39:28)

100%.

Matt Sheehan (39:39)

really an IPO guy. think the IPOs were like modestly successful, but the valuations are just, yeah, the valuations are, yeah, not even close. And ⁓ it reflects a lot of things, but it largely reflects like a funding environment, a business environment, a macro economic environment, and the general sort of attitude towards risk investment. think I was just reading something that ⁓ Ray Ma from ⁓ TechBuzz.

Grace Shao (39:42)

So that’s six to eight billion dollars. The valuation is tiny compared to American peers.

Matt Sheehan (40:06)

China was writing on this. She’s always very good on these topics. yeah, it’s just people kind of assume that there’s like infinite money in China. They’re like, yeah, the government, whenever they want to, they just like turn on the taps and then, you know, it’s like, no, that’s not how it works. And like the VC ecosystem is much smaller, much more new and immature. And so it’s a different story.

Grace Shao (40:28)

on that note you know I was just in SF like last month and I met with quite a lot of investors and people’s kind of I guess misunderstanding was often twofold. One is exactly your point, people are just like oh China’s so rich the government just gives money all these AI companies are backed by the Chinese government I was like 100 % no first of all like there’s some other issues happening in the background but like the government doesn’t even are you know it’s kind of cash constraint and not even that much right now second all these companies are definitely not being backed by the government in any sense in fact

Most of don’t want to take municipal governments or provincial government money because you get kind of tied into, you know, what we’re seeing is, you you get forced into working with the government and it constrains your profitability and commercial goals. On the other hand, another really big misconception was, I thought quite funny was that people often ask, was open-claw frenzy because the Chinese, average Chinese consumer or user were really, really concerned about privacy issues. So they wanted everything on edge.

I was like, hmm, like again, it’s kind of like just not, a major conversation people have. Like I think I hate to generalize, but I think because of how the internet ecosystem is in China, people by default have kind of ceded to not thinking too much about privacy or personal data issues as much. So that definitely isn’t. So I kind of want to bring the conversation that this, you know, like

What are some biggest misconceptions you think people have and how do we help them understand and bridge that gap a little bit better?

Matt Sheehan (41:59)

Yeah, I think yeah some of the stuff that you point out is correct like If you’re if you’re a if you’re a startup if you’re like a small medium company You I’ve talked to these people they’re like actually like we do not want to take government money if we can avoid it not just because we get kind of in mesh but like Entrepreneurs are legitimately afraid that if they take government money and then their company doesn’t work out and they lose the government’s money like they could end up like on the hook like in jail this

legitimate fear that it was stated to me by someone. like, you know, is that happening to entrepreneurs everywhere? No, but it’s like you don’t. ⁓ The government. It does a certain amount of sort of VC-esque investing, but there’s not really that VC mentality of like high risk, high reward. Like we know that most of this is going to go under. It’s kind of local governments at least have been trained on like real estate investment, which is like 10 percent, 10 percent, 10 percent every year.

And this idea that most of these companies that you invest in are going to fail is not really ⁓ deeply embedded there. I do think some of the companies do rely on government funding in different ways. ⁓ mean, Z.ai, Drupal, one of their biggest, maybe their biggest single revenue stream is from ⁓ building custom models, but custom applications for ⁓ state-owned enterprises, local government, stuff like that.

Matt Sheehan (43:28)

When you listen to them in interviews, they’re like, it’s not that big. Maybe it’s 40 % or something like that. But it’s a significant revenue. It’s part of their business model. So there’s that type of a connection to government. With DeepSeek, that’s a company that’s kind of quite mysterious. And we don’t know exactly where all their money comes from. Is it all earned? I think the government got more hands on with them in the sort of aftermath of the DeepSeek moment. You had reports about the government taking passports away from

people who worked there to make sure like you guys stay local ⁓ or the government was like vetting investors was another this is reporting the information. ⁓ But the idea that like these companies are just they just have the kind of like the hose of government money just flowing in at all times and therefore they don’t have to think about anything else is just not it’s just not real. ⁓ They’re they’re much more constrained cash constrained. ⁓

terms of like trying to misconception on Chinese AI regulation, AI policy, this is like my, you know, much of my job is like first getting across like, the trend does actually like seriously regulate the technology. And then, you know, the next layer being like, it’s not all people think, you know, it’s an authoritarian system. She didn’t think he must just kind of like sit down and just like write the regulation. So like nothing matters except what he thinks. And we don’t know what he thinks. It’s like, no, like he doesn’t. There’s this actually very complicated and sophisticated policy ecosystem of,

legal scholars and the companies are doing their lobbying and their thought leadership and, you know, they’re responding to public outcry over things. And I think that’s a, know, you can get this across to people, but it’s certainly not the people’s default mental model of how China works on policy is ⁓ just does not reflect the kind of sophistication in this zone. And it’s somewhat understandable. Like there are policy areas where Xi Jinping just like makes a decision and that’s.

that’s where things are going. Like I think we saw a lot of this in the kind of 2020 to 2022 era. But as an AI policy, COVID, AI policy, it’s not that way. certainly things are not, people are not gonna like, you know, actively push things that are totally against the will of the top leaders, but they are within the constraints of like,

Matt Sheehan (45:52)

the direction of travel, what the CCP is good with, what she wants to do within that kind of very wide lens. It’s really individual people, scholars, bureaucrats, companies that are filling in all the details on this. And it’s a very sophisticated system because they’ve just had a lot of ⁓ had a lot of bites at the apple. They have like passed, I don’t know, eight, nine different A.I. already.

The regulators at the CAC have been getting documentation from AI companies for three, four years. They’ve been building evaluations. been like, and they kind of, got their reps in with AI policy and it leads to a more sophisticated ecosystem.

Grace Shao (46:33)

⁓ yeah, so the last kind of section I want to focus on is just getting into the nitty gritty about, you know, the policy and the security and safety kind of side of things we touched on in the beginning of our conversation. you are one of the few in the West, I think, and talk about the nuance of the word, which you just explained, it’s security, but also safety. ⁓ help us understand.

how to interpret that when we read about that. It actually even helps us understand a little bit of what’s happening in the West. Like, I feel like there’s the governance people, the security people, the safety people, but from someone who might not be in that ecosystem, people are conflating it a little bit. And I just want to understand, you know, how do we understand each of their objectives, again, KPIs, or even their goals?

Matt Sheehan (47:20)

Yeah, yeah, basically, it’s really complicated. It’s very context dependent and it’s always changing. ⁓ I think maybe the first key thing to understand here is like the very particular meaning of AI safety in the West like that. The West AI safety ⁓ largely refers to kind of a specific camp ⁓ of AI development and policy people that are, you know, believe that AI is going to achieve human and superhuman capabilities.

And this could pose like serious, maybe catastrophic risks to people. like, that’s a somewhat coherent community in the United States that has a certain amount of power. Their power kind of ebbs and flows depending on things. But like when you say AI safety in Washington, D.C., it means one quite specific thing. ⁓ In China, that community, it has started to emerge, but it’s much newer. It’s much more recent. It doesn’t have the deep roots that it has in the West.

And ⁓ the way that the word is used in policy documents is both confusing and has changed over time. So a lot of times when ⁓ just to kind of put a little color on the terminology, Anquan, when ⁓ when you’re talking about cybersecurity in Chinese, you say Wang Luo Anquan. So it’s like network Anquan, network security and cybersecurity means something very different from

AI safety from super powerful AI systems posing risks. And so there’s one sort of category of mistakes, which is to be very naive and to read all the Chinese policy documents. And every time they say a word that’s translated as safety to believe, wow, they’re talking about AI safety, they really, really care about this. That is a very naive and incorrect reading of things. ⁓ But in the past, I would say, 18 months, two years,

you have seen a pretty significant uptick in the way that people sort of in and around the system and to a certain extent in and around the companies, their level of attention to what we would call in the West, AI safety to these more kind of large scale, potentially catastrophic risks from powerful AI. I’d say this is, there’s like sort of levels and degrees of this. There’s people talking about this. There’s it showing up in government documents in one way or another.

And then there’s actually implementing this either through like binding regulations or through sort of companies doing their own testing and evaluation to try to their own sort of safety research and their own safety testing. I’d say what we’ve seen so far is a large increase in rhetoric, a large increase in sort of awareness within the policy community about these safety issues. We’ve seen it to start to show up in more significant documents. I think the one you’re referring to early on that I was working working on analyzing is called

They call in Chinese the AI safety and governance framework 2.0, which is in many ways put out by some organizations underneath the CAC, the internet regulator. And it’s kind of ⁓ their attempt to diagram and do an initial discussion of how they see different risks from AI ⁓ and how are they going to mitigate these risks. Oftentimes they’re focused on technical standards as a mitigation. And there was a AI safety governance framework 1.0 in 2010.

fall of 2024 and there was a 2.0 in fall of 2025. And just between those two documents, you can see real increase in the frequency and to a certain extent, the sophistication of the discussion around these risks in China. I’d say it’s pretty significantly below the sort of the AI safety discussion in the United States, but it’s on the radar. will counter, they’re like, okay, that’s great that they’re talking about it, but are what, you

Are they just trying to trick us? Are they trying to make us believe they believe in safety? Are they saying it but not doing it? And ⁓ they’re like, we have not seen much in the way of like, we certainly have not seen like concrete binding regulations that sort of implement safeguards on this front. And in terms of what the companies are doing, it’s quite opaque, but ⁓ we don’t think that they’re doing the, I’d say.

pretty confident they’re not doing nearly the level of sophistication or intensity of safety testing as you see at places like OpenAI and Anthropic. To me, this seems somewhat normal. This is kind of a process. Chinese companies have been behind. The government has perceived itself as being behind. When you’re behind the frontier, you’re not as worried about frontier risks as you’re like other people are going to get to those first and we need to catch up. ⁓ So I see this as kind of like a long-term process. And I think that the sort of increase in discussion about this is

you encouraging if you’re concerned about these issues. But it’s a really don’t want to be just kind of reading the documents and say every time we see Anquan being like, wow, China cares about AI safety. Look at all this stuff. It’s much more ⁓ nuanced and evolving, ⁓ evolving quickly, I would say.

Grace Shao (52:21)

I know like when we spoke a couple months ago, just catching up, you were saying a big part of your job is also trying to help, you know, bring the two sides together. Obviously, you know, ⁓ it’s been challenging given the geopolitical backdrop, but how do you think the global AI governance space can work together? ⁓ Are we going to see, you know, kind of the two world superpowers and two super AI powers, ⁓ you know, guide?

in different directions or do you think there are certain issues where they need to come together and they will come together and are coming together? ⁓ For example, to your point on safety issues around protecting humanity, protecting children, are these things that you are seeing collaboration?

Matt Sheehan (53:08)

Yeah, I’m kind of ⁓ both an optimist and a pessimist on this front in that, like I said, I have very, very low expectations for the United States and China to work together on anything. I have very, very low expectations of any type of a binding agreement or some type of detente where we both shake hands and kumbaya and we’re both going to be very safe with AI and we agree and it’s great. I just don’t expect that. ⁓

So in that way, I’m pessimistic. I think attempts to try to sort of preemptively create these global governance structures that are going to bind both of the countries in advance so we never reach these dangerous thresholds. ⁓ That’s just not where I’m putting my bets. I think it’s good. We need to make all kinds of bets on this front, and it’s good that people are working on this, but that’s not where I’m putting my bets. Where I’m putting my bets is on a much more

limited kind of narrow bore, but I think potentially highly effective form of ⁓ engagement. wouldn’t even say cooperation. I wouldn’t even necessarily say coordination. ⁓ my sort of the term, my mental model for it is something I call AI safety in parallel, which is that like the two ecosystems are going to be moving somewhat in parallel. They’re both going to be pushing the technology forward. They’re both going to be working through safety issues from a technical perspective, from a policy perspective.

And as we kind of move forward in parallel, we’re not going to be telling each other what to do. And we’re not going to be like, okay, I’ll do the safety thing because you are. You told me you’re going to do it, so I’m going to do it. We’re not like sort of moving in lockstep on this, moving in parallel. And we need to have these touch points. We need to have touch points where the two sides develop some form of mutual understanding of what the other side is doing. They understand how other side is thinking about the issues. They understand how they’re perceiving these risks. That’s one of the reasons I do the risk ranking.

stuff and in some cases trying to share best practices, explain kind of explain what we’re doing and why we’re doing it and have the Chinese side explain what they’re doing and why they’re doing it and then where possible share good ideas that we think are sort of uniformly good in the U.S. and China. If we think that we have a policy intervention maybe it’s around ⁓ certain types of pre-deployment testing. ⁓ It’s good to communicate that.

to the Chinese side. And it’s good to have the Chinese side communicate some of ⁓ the reasons and the sort of the specifics of what they’re doing on these fronts. We’re not here to just like trust each other. I think a lot of people are very worried that the Chinese side is going to, is they’re going to trick us. They’re going to say they’re doing it and they’re not, which is like legitimate concern. know, that’s, this is high stakes like geopolitics and powerful technology. So you don’t take anybody’s word for it. But when you have these conversations in talking with people, you can get a,

a sense of their level of sophistication when talking about the issue. If someone is talking about AI safety and they’re like, yes, humanity first, protect the humans, control the machines, that’s our policy. And it’s like, okay, is there anything more to that? They don’t have more than you kind of know that they’re actually not really thinking about it. But if you can get into a more ⁓ deeply engaged discussion, you can see like, actually, yeah, they’re working through these problems themselves. You can see it in the way that they’re.

discussing it. You can see it when they talk about their specific regulatory mechanisms. You can see sort of the connection between sort of action and outcome or thinking and action. And so my model for this is like, we’re not going to agree on things. We’re not going to sort of trust each other. But there are ways that we can both be moving forward at the same time and comparing notes, checking in, getting a sense of what the other side is thinking and doing that I think could contribute to safety in a meaningful way.

Grace Shao (57:02)

I think that’s fair and I think the word you kept on using trust is quite interesting because I feel like whenever I speak to people in the industry, ⁓ there’s just such a lack of trust even within whether you want to say countries or communities or beliefs and value systems and a very, very optimistic, naive way, I really hope that there can be a bit more consensus on certain things like that need to be protected in practice, like such as children, right? And how we go ahead with that. ⁓ But okay, I don’t want to end on a super somber note or anything, but... ⁓ The takeaway is trust nobody. That was...

Matt Sheehan (57:34)

It’s optimistic in a way. think this can’t... When you do see... Well, trust nobody, but

talk and see if you can share some good ideas along the way. think there is real... ⁓ I’ve seen some real sort of traction from these type of things and I think it’s limited. We’re not going to get some kind of hard guarantee that China is going to be perfectly safe or we’re both going to...

Matt Sheehan (58:04)

do the right thing. But within with those low expectations, with those kind of pessimistic expectations, there are there’s progress that can be made.

Grace Shao (58:13)

⁓ I do want to ask one question that’s kind of been happening around right now. That’s been kind of happening like the whole idea of Chinese open models seem to take a little bit of a sidetrack and starting to kind of only release their most frontier related models and close weights. ⁓ Obviously from a very like, you know, capital perspective, where I study it is I feel like it’s a lot of it is because they need to see our eye. They cannot keep doing this because they’re not making money. API sales not enough to, you know, sustain the kind of

long-term ⁓ business as well as research costs. Are you seeing anything from the policy side? Like do you think there’s been a policy shift? That’s also kind of why I asked earlier if this year somehow, know, last year there was a public embrace by the government saying we should open source our technology. Has there been a shift?

Matt Sheehan (59:04)

So ⁓ I’ve seen sort of little tidbits around ⁓ sort of public policy concern about open weight models, but not enough that I would call it a shift. ⁓ In that document, the AI safety governance framework 2.0, it was interesting because it was the first time that there was a fair amount of ink spent on potential risks from open source models. The risks they were primarily talking about were

essentially if there are vulnerabilities in these models in some way, either maliciously inserted or just a vulnerability mistake, those could proliferate throughout the ecosystem because you have all these downstream models and that could lead to impacts. There’s a little bit of a mention of like, maybe open models will be used by criminals and stuff like that. I certainly don’t see this shift in specifically Alibaba strategy as

Matt Sheehan (1:00:04)

in reaction to a significant policy shift. mean, it kind of makes, yeah, corporate decision. It makes if you’re going to be spending tens of billions of dollars building models or at least hundreds of millions, billions, billions of models, giving it away for free is a. That’s a choice. I think there’s reasons why it’s advantageous for China to do that, or at least it was for a stage like it was going to be pretty hard.

Grace Shao (1:00:09)

corporate decision then.

Matt Sheehan (1:00:34)

to get people around the world to kind of believe in ⁓ Chinese models if they were only going to be able to access them through API. ⁓ You know, a lot of American companies that are, you know, deploying ⁓ Quan, Alibaba’s model, I don’t think they’d be doing that if they, I don’t think they would have at least made that leap initially if they had to sign a contract with Alibaba and they believe that maybe their data was going back to China or, you know, the model was more of a black box relative to them. So maybe the open model wave was a very good period of publicity. ⁓

that might pass with time, but I don’t know. I think it’s being more nuanced. I’m not the expert on this. Nathan Lambert, runs the interconnects sub stack, and then Kevin Shue, who runs the interconnected sub stack, are much, much better and more sophisticated on this, and they have good writing on the ecosystem. But from a policy perspective, I haven’t seen a shift. I think it just kind of makes sense.

Grace Shao (1:01:35)

Yeah. Nathan, I actually spoke about this ⁓ a week ago, and I think both of us kind of feel like it’s really more of a capital and business constraint that’s really driving this. ⁓ But I just wanted to hear from you if maybe there was some kind of a top down initiative as to but it doesn’t seem like it. Right. I have one last question for you, which is a question I ask every single guest that comes on. ⁓ What is one differentiated view you hold? And actually, I try not to limit the question to just about AI. However, most people do want to talk about that.

Matt Sheehan (1:02:02)

Mm.⁓ Differentiated view. ⁓I mean, is... I don’t know where most people fall on this, but I think one thing I’ve been thinking about lately is just language learning and what’s going to happen with language learning in the era of simultaneous really good translation either through hardware, devices, or software. And I’ve been reflecting, why have I been spending 16 years or more learning Chinese and just in the...

Matt Sheehan (1:02:40)

mud of trying to learn and remember this language. And I guess my maybe differentiated view on this is I think it still is really important. And I, I waver, you know, I don’t want to just be like, ⁓ you know, justifying to myself why I’ve spent all this time. And like, I, don’t know that I would tell like a 16 year old, like I want you to invest, you know, 10,000 hours on a language when, when you can have it all translated for you, but they’re

there just is something quite meaningfully different about reading these policy documents, about listening to people and hearing the original language and just knowing how that language is used in all these different contexts that gets flattened by translation. I use machine translation all the time. I throw documents into it to get either a first draft or to get something I can share with people. But I guess my view is that I really think

We need people to keep like actually learning Americans to actually keep learning Chinese. And I also just think it’s so much fun. just, was earlier today, I was thinking it was like, why is this, it’s been such a, like a joy in many ways, extremely painful, but kind of a joy to like really struggle with a language over time. And so that’s my take.

Grace Shao (1:03:59)

No, I actually agree with you and I think if anything, I’ve thought about this a lot. So I’m raising a trilingual child by nature because we speak English at home, our parents speak Mandarin, the environment she lives in, they live in speaks Cantonese, right? And I think to your point, in many ways I’m like, wow, it’s actually, it would be so easy for them to travel the world and communicate with people. But the reason why I pushed them to learn the language is really to communicate, to understand a culture and the people and more.

nuanced way, even for myself, like my parents pushed me to learn Mandarin. It like to your point, it’s so painful. But the ability to speak to my mother in her native tongue and understand her is so much more complex and you appreciate much more when you’re older, even though my parents speak English, obviously, but when we were young, we would speak English to him. As I got older, I actually enjoyed speaking to them in Chinese much more because you hear about

Matt Sheehan (1:04:55)

Hmm.

Grace Shao (1:04:57)

It’s also your personality changes, right? Like you kind of get to the core of who they are in their native language and their native way of expressions. So I think for sure, I agree for certain languages, there’s still such value, if anything, even more so to understand a human connection, human connectivity. then for pragmatic reasons, like I took two years of German, I remember nothing. I probably wouldn’t do that again to myself, especially in my class. a bunch of third gen German kids where they spoke the language at home but they can get away with saying they were doing beginner’s German, you know? But yeah, so I appreciate that. Thank you so much Matt, thank you for your time, I really appreciate it, we finally got together to do this episode.

Matt Sheehan (1:05:34)

Yeah, thanks for having me. That was fun.

Today’s post is a long read, and it is not just limited to “China AI.” I’m not here to fearmonger nor to spread doomerism. Let me know your thoughts…

Share

The context and the dichotomy: to be replaced or not to be replaced

For a long time, the anxious ones have said that AI will take all our jobs. But as agentic AI gets integrated into everything from customer service to consulting to even fundamental research and equity investment, what people are realizing is that, 1/ like what Jensen Huang has said, don’t conflate a task to a job. 2/ And even for those who first warned of negative growth, Professor Alex Imas seems to have changed his mind, “a vision of how the AI economy could work out not so badly.” The economist and professor of behavioral science, economics, and applied AI at the University of Chicago Booth School of Business, said on Bloomberg’s Odd Lots recently, just because a job is exposed to AI, it does not mean it will be taken by AI. AI agents may be able to do, let’s say, 50% of the tasks, but then the human is freed up to do the other tasks that require the human touch, the ability to connect the interrelated tasks, and make it into a complete loop. An AI native force may free up the tedious tasks that certain jobs require, but then, actually, these humans who are left to “quality control” will end up being paid more.

This argument has been echoed by various industry leaders (Jensen) and investors, too; some even go so far as to say there will be more lawyers, more software engineers, more demand for what is already scarce because the volume of that challenging work will increase, and more people will be required to manage the agentic workflows.

Imas makes the broader economic case in a recent Fortune profile. His framing is structural change theory, meaning that the economics of what happens when technology makes one sector dramatically more productive. The example he gave is that around 1900, 40% of the American workforce farmed. Today, that number is under 2%. People didn’t stop eating; they just stopped spending most of their time making food once it became commoditized. “The economics of scarcity won’t disappear; it’ll just relocate.” His implication: as AI commoditizes more of the economy, spending and employment will migrate toward what he calls the “relational sector” — things with a distinct human element. Middle-class consumption patterns tomorrow will look like those of the wealthy today.

So the argument is that, as we all anticipate disruption, the value may shift to ‘relationship roles’ while those tedious initial steps and the drafts of their knowledge work output could be completed by AI. And the disruption may not be apocalyptic; rather, it will push some traditional business models to evolve…

Professional services may be first in line

A slight side tangent on software, but it is context for my point later down. After many conversations with investors across software (oomph, been rough) in recent weeks, I realized that the disruption won’t be as in-your-face as some may assume because, beyond the AI-supercharged bubbles, most industries will pose many hurdles to AI adoption at scale. Software will not be replaced as smoothly and swiftly as the market’s reaction suggested.

But much of that “software will be completely eaten” argument is flawed. Software was never just code. It’s maintenance, security patches, compliance, integrations, support — an entire service layer that has nothing to do with writing the first version. Ben Thompson wrote about this recently in “The Beneficiaries of AI-Written Code” on Stratechery, and I think he nails it. “Companies — particularly American ones — are very good at focusing on their core competency, and for most companies in the world, that isn’t software.”

1/ Software often is not just about the actual action; it’s the embedded value of service, natural plugin, ecosystem, debugging, and so on. 2/ If AI-assisted work will increase productivity, then the number of activities will increase, and these activities require actual software to complete. Agents are “completing tasks” by accessing your browser, email, CRM, and so on. They’re not going to do it on thin air.

The vibe coding discourse on X makes it seem like every company is about to build its own internal tools from scratch. Some will. But writing the original app is the easy part. It’s everything after that which people underestimate. And it still goes back to human judgment and safeguarding.

However, the argument that AI will disrupt the workforce may be a bit more nuanced: AI-assisted work will drive more of the change in the workplace, rather than completely replacing work (only in certain categories). You see my thread here?

In the coming years, if not months, the first major disruption from AI and agentic systems will likely hit the industries that built their economies on scarce human synthesis, process, and billable time. Not actually the code itself.

So think law, consulting, accounting, and so…

Think about it. Much of why law, accounting, consulting, and advisory can charge such high hourly rates is because clients believe they are buying elite training, specialized synthesis, and reliable processes. They are also buying the human judgment and human credibility behind the deliverables. Much of the signaling credentials translates to reliability; it is hedging against the possibility that something does go wrong. While, of course, much of that value is real, until recently, turning messy information into a defensible answer required a lot of expensive, smart people, which meant a lot of money for the businesses.

Now, with generative AI, those steps and the first drafts of their knowledge work output could be done much more cheaply. Agentic AI can make research, document drafting, and many of these workflows cheaper, at least its more junior responsibilities.

The gatekeepers

For years, the modern knowledge-worker economy rested on a bargain. Customers paid high fees because good judgment, high-quality synthesis, and reliable processes were considered scarce. AI does not destroy that bargain all at once. But it does something just as consequential: it exposes how much of it was built on the old cost of intelligence and how much of that is human judgment.

The old guards made money from controlling two kinds of scarcity.

The first was access scarcity: it was hard to get the degree, hard to enter the profession, hard to learn the playbook, hard to know how a proper prospectus, memo, contract, or board deck was supposed to look and sound. Entire industries were built on that gatekeeping. Clients were paying not only for the answer but also for access to the institutionally approved way of producing it.

The second was judgment scarcity: even after the work was produced, clients still needed someone credible to stand behind it. They needed the senior banker, the law partner, the consultant, the auditor, the person whose name, reputation, and liability sat behind the document.

And we’re seeing that AI is eroding the first scarcity much faster than the second. It is making elite formats, workflows, and first-pass analysis much easier to reproduce. The gates around production are coming down. It’s no longer about whether you attended the right schools, were trained at the right firms, and learned the right vocabulary. More people can now produce contracts, documents, research, and presentations that look close — at least closer — to institutional-grade.

Judgment, credibility, trust, and accountability still matter. Goldman can still charge for the prospectus and the advisory not just because only its bankers can physically produce the document, but because its name helps place the deal, frame the narrative, and signal quality to the market. A senior law partner can still charge to review a contract because clients are not only buying the words on the page. They are buying the partner’s judgment, credibility, and willingness to stand behind the advice if something goes wrong. Kate Zhao, a strategic communications leader, made a similar point in a recent FT Chinese op-ed: “IR is evolving from information delivery to value interpretation and trust building.” The deliverable can be in law, in IR, in banking advisory, in M&A advisory, or whatever, but the deliverable is no longer what people are paying for. It’s the human sitting behind it.

Where the repricing is already visible

For years, many firms have been able to charge not just for expertise, but for the gates around expertise. As intelligence becomes cheaper and more abundant, those gates start to look less defensible. And service providers cannot only just lean on their ‘rules’ or old ways of doing business. As they already have the intent and willingness to purchase, all they need to do is evolve; otherwise, a more modern provider may offer a cheaper, more efficient alternative.

In a recent SemiAnalysis report, “Dark Output: When Wages, Prices, and Output All Lie”, they actually even broke down the cost of labor and the cost of tasks. See this excerpt.

“A junior associate at a midsize law firm billed 6.2 hours last Tuesday reviewing an NDA for a Series B startup. Standard work. The partner signed off. The client paid $2,480.”We built a benchmark — 1,105 professional tasks, each priced against frontier AI, each quality-gated by adversarial evaluation — and ran the same NDA through it. Same document, same review criteria, same output spec. Cost: three cents. Time: eleven seconds. Quality gate: pass.

Replacement ratio: 10,257 to 1.

That associate still has a job. Most of them do — for now. But the price anchor for NDA review has moved from $400 to four cents, and the firms that know it are capturing the spread quietly. This is not a story about one associate. It’s a story about what happens when the cost of professional work collapses by four orders of magnitude and the instruments we use to measure the economy can’t see it.”

That’s exactly it, isn’t it? That is what SemiAnalysis called the “Dark Output”, it “is labor displacement by AI that is invisible to official statistics. GDP measures spending, not output; when AI replaces a $300 human task with a $0.50 AI task, GDP sees $299.50 less spending even though the work was still done. The displacement is real. The statistics record silence.”

Adopt or Else Be Repriced

SemiAnalysis, as the bro-voice they are, may not sound sensitive, but just being real here. “AI eats the cheap work first. When the $400 wills and the $175 NDAs vanish from the transaction data, the only legal services left in the price index are the genuinely hard tasks with big dollar signs attached. Measured prices go up. The actual cost of legal work went down by four orders of magnitude.”

The reality is that we’re already seeing repricing across three layers.

A. Intelligence is becoming abundant. Drafting, summarizing, synthesizing, coding, reviewing, comparing, brainstorming, translating, and first-pass analysis — all dramatically cheaper. This is the layer most people see when they say “AI changes everything.” They are right that it is changing. They are wrong that it is the whole story.

B. Workflow is still scarce. Someone still has to define the process, own the systems, manage permissions, integrate data, handle exceptions, and make the output usable inside an organization. Tech analyst Benedict Evans argues that software is never just code. It is an institutionalized workflow. “The hard part is working out that that problem exists and then working out the right way of solving it and then going out and building a go-to-market.”

And in their conversation on Stratechery, Ben Thompson makes the complementary point. The typical large U.S. company runs 350 to 450 separate SaaS applications. Evans calls every single one of them “a thin SQL wrapper.” Just a database. But we know, the value was never the code. It was the workflow someone designed, productized, and supports. Code is intelligence. The product around the code is workflow and accountability. Those are different tiers, and they’re repricing at different speeds.

C. And then accountability is becoming scarcer still. When the answer is wrong, incomplete, illegal, non-compliant, politically sensitive, or strategically foolish, somebody still has to stand behind it. In a world where the first draft is increasingly machine-generated, the premium on the person who says “I take responsibility for this output” only grows.

The Consultancies Are Already Feeling It

If this repricing thesis sounds abstract, look at what the Big Four and top consultancies have done in the last twelve months. The numbers tell a clear story: they are simultaneously cutting the old labor model and spending aggressively to build a new one.

PwC eliminated roughly 3,300 roles between September 2024 and May 2025, primarily in audit and tax — the most process-heavy, preparation-intensive functions. It cut graduate hiring and missed its 100,000-person headcount growth target. KPMG cut approximately 330 employees from its U.S. audit workforce, about 4% of the division. McKinsey laid off hundreds of people in technical roles — design, data engineering, cloud, and software. For these professional services, their assets are their people, and when remuneration around people’s time, execution, and judgment are being challenged, they’re faced with the reality: transform or die.

So with cutting, there are new hires to facilitate that transformation. EY has committed $1.4 billion to its EY.ai platform, with over $10 billion in total enterprise AI investment over three years. The Big Four collectively are investing north of $4 billion in AI. EY’s AI-related revenue grew 30% in fiscal 2025, and the firm deployed over 100 internal AI applications with another 1,000-plus agents in development.

McKinsey’s own research estimates that AI could automate up to 30% of consulting work hours by 2030 — primarily in research, analysis, benchmarking, scenario modeling, and presentation preparation. That is not a competitor saying it; that is McKinsey publishing it about its own industry. The firm has deployed thousands of AI agents internally, and its CEO expects roughly equal numbers of AI agents to human employees by the end of 2025.

The consulting business model is shifting from labor arbitrage to technology leverage. And there’s a human reason why professional services feel it first: executives will always have more qualms about firing in-house teams than cutting vendors. Most Fortune 500 companies would be more ruthless about reducing service provider spend than disrupting internal morale. Professional services firms are, in that sense, the path of least resistance.

Case Study: Harvey and the legal vertical

At the March Morgan Stanley TMT conference in SF, Harvey’s CFO described not a company trying to “replace lawyers” in the cartoonish sense, but domain-specific AI that works because it sits close to a high-value workflow. He described a company investing heavily in security, context, the integration of internal and external knowledge, and built-in product evaluation. His point: differentiation is not just the model. It is everything built on top — domain proximity, workflow design, legal engineers, onboarding, evaluation, collaboration, and memory.

Now, Harvey’s CFO has an obvious reason to frame the moat this way. His company is raising at a multi-billion-dollar valuation, and his job at a Morgan Stanley conference is to convince TMT investors that the application layer, not the model, is where value accrues. So take it with a grain of salt.

Fortune Magazine profile of Harvey CEO and co-founder Winston Weinberg

But it is still clear that Harvey does not need to remove the lawyer to break the economics of the old process. It only needs to remove enough of the drudgery to compress turnaround time, reduce staffing needs, and shift the human role upward.

And there’s something else Harvey gets right that I think gets lost in the broader AI-will-eat-everything discourse. Winning enterprise is fundamentally different from winning consumer. There isn’t enough fragmentation for a generalist AI to walk in and take over. Enterprise means compliance, edge cases, institutional memory, client-specific preferences — a completely different set of expectations. Harvey picked one thing — legal — and is nailing it. Twenty percent of their employees are lawyers. They built evaluation benchmarks specific to BigLaw. They went after Allen & Overy as their first customer, not the easiest one. That is the opposite of the vibe-coded, do-everything, ship-fast consumer playbook. And it’s why the moat, if there is one, will be built by companies that go deep into a single workflow rather than wide across many.

[And this is becoming a clear trend. I just met with some former hedge fund investors who are trying to build an AI-native equity research product. The product can only be enterprise-grade ready if they are built by people who understand the workflow and incentives of the job, thus helping to complete tasks. They’re hoping to sell to the single-PM shops first, then the platforms. Emphasis* these are former (top-tier hedge fund) investors building tools for their peers.]

So, as much as professional services are about a team of people ‘servicing’ the client, it is very much about the senior partner’s know-how, which offers assurance to the in-house person. It’s a whole chain of command built on checks and balances, but the process can be extremely inefficient. The junior layers do the reading, sorting, summarizing, cross-checking, marking up, formatting, red-flagging, and first-pass analysis that make that assurance possible. AI can replace that preparation layer first. It does not need to be perfect to be economically disruptive. To make economic sense, it only needs to make a ten-person workflow workable with six for a business to pull the trigger.

The Barbell and the Apprenticeship Gap

The same logic that we mentioned that Thompson applies to software applies here. AI does not eliminate the need for legal, accounting, or consulting expertise. It eliminates the need for as many people to do the preparatory work around that expertise.

At one end, a smaller number of highly leveraged professionals will operate with agents, copilots, memory layers, and internal tools — producing far more per person than before. At the other end, organizations will maintain a smaller but more accountable layer of operators, reviewers, relationship managers, and domain specialists who handle edge cases, manage risk, and absorb ambiguity.

The people who are ‘junior’ will be AI-native running AI tools and agents, and the people up top will review and quality control. This raises a structural question that I think is underappreciated. The labor pyramid at most professional services firms — BigLaw, the Big Four, large consultancies — often depends on junior billing (insane #) hours subsidizing partner economics. If AI compresses the base of that pyramid, the margin model that sustains it comes under pressure. PwC missing its headcount growth target is not just a hiring story. It is an early signal that the traditional pyramid may no longer scale the way it used to.

And there is a deeper problem beneath the surface. If the junior work gets automated, where do future seniors come from? The document review, the late-night diligence pass, the first-draft memo that gets redlined six times — that was not just grunt work. It was an apprenticeship. It was how people learned pattern recognition, institutional judgment, and client instincts that made them valuable ten years later. AI may not fully replace the expert. But it can absolutely compress the work that used to train the expert. That question will show up in accounting, consulting, research, recruiting, underwriting, due diligence, and large parts of corporate coordination.

The thing is, the integration of AI into our existing workforce requires us to revamp our business models and hiring strategy because AI does not really discriminate between different forms of professional knowledge based on their current economic value. Legal knowledge may be more valuable than, let’s say, grocery purchasing knowledge today — at least it is more life-or-death — but AI mainly cares about tokens consumed.

From intelligence to agency

There is one more shift I think a lot of people are sleeping on that reframes the whole discussion.

We keep debating which model is smartest. But the real question is quickly becoming: who builds the best agent layer? This is why people are talking about harness engineering. Not just raw reasoning, but the orchestration — identity, permissions, memory, context, multi-app coordination. This is the harness thesis again, but applied to the enterprise knowledge stack rather than the cloud platform stack.

We’ve written about the tension between Tencent's and Alibaba’s token hub strategies. If agents reduce the number of humans doing knowledge work, and your business model charges per seat, you’re literally pricing against your own product’s success.

The broader point is that the debate is shifting from “which AI is smartest” to “which system can actually get work done inside an organization.” When intelligence is commoditized, the new scarcity is agency, and the ability to act across systems, with the right permissions, in the right sequence, with accountability attached. That is workflow. That is what the refineries are really selling.

This applies to professional services firms as much as it does to software or cloud companies. The consultancy, law firm, or accounting practice that builds the best agent layer on top of its institutional knowledge (ie. Harvey) — its client histories, matter precedents, compliance frameworks — will have a durable advantage. Not because of the model underneath, but because of the workflow, context, and accountability stacked on top. Harvey’s investment in “memory” — user-level, matter-level, client-level context within ethical walls — is an early version of exactly this play.

If models continue to commoditize, and I think the direction is pretty clear, value will accrue less to whoever owns the best brain and more to whoever owns the best route from question to workflow to accountable outcome. Incidentally, this is also why vibecoding won’t replace industry-specific software — context isn’t something you can prompt-engineer.

Where we’re headed

Thus, are we all going to be replaced? I think not? I hope not? But it may be something more immediate and, in some ways, more destabilizing: a structural shift in who captures value and how. Some of the incumbents in professional services will hold on because they have trusted relationships, proprietary distribution, regulatory permissions, or deeply embedded workflows. In some cases, the top of the profession may even gain leverage as trusted advisors become more productive. But the old bargain underneath many firms will change. What looked like a moat may, in part, have been a labor-cost structure the market mistook for defensibility.

As you all know, I’m a natural optimist, and that’s maybe why I like writing for investors and seeing the world through a ‘growth’ and ‘potential’ lens. Call me crazy, but I do get very emotionally impacted when surrounded by negative thoughts. So, as I have written before, about the tension between AI’s extraordinary capabilities and the real-world frictions that slow its adoption, such as governance, budget reallocation, and organizational absorption. Those frictions are real, and they will buy time. But they will not reverse the direction. The industries, firms, and professionals who use that time to redesign their workflows from first principles, rather than simply bolting AI onto the old process, will define the next era of knowledge work. What will become the real moats for all of us will be our judgments.

Because when intelligence becomes cheap, the real scarcity is no longer the answer. It is everything required for the answer to matter.

And on that note, I just had an awesome conversation with at CEIP for my podcast Differentiated Understanding that will be released next Monday morning EST. He will walk us through his recent writing on China’s rising anxiety about job loss to AI and the broader policy picture. As always, thank you for reading, especially if you actually finished reading this super long post today. Get in touch, I’d love to hear from readers!

But despite all this, Tencent remains notably committed to the mass consumer market. The OpenClaw frenzy has already led to five different Clawbot-style products emerging across its ecosystem. Joining me today is Shuyu Zhang, Senior Product Manager of QClaw, to break down the thinking behind that frenzy, from the cultural logic to the business rationale to the product design choices shaping it all.

QClaw is to be accessible to everyone on April 21. It is the first consumer-grade AI agent built on OpenClaw. No technical setup, scan a QR Code, and the agent will be live in 3 minutes.

Product link: qclawsg.qq.com

Waist list: https://docs.google.com/forms/d/e/1FAIpQLSeIfEzlOV8jq_tGMbV5mqTSALyufE0kZ933XqE3Fnha1_CRfA/viewform?usp=publish-editor (Founding Claw — limited 20,000 slots)

Share

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently. Season two will host a series of guests from early-stage investing, as well as builders, founders, and product managers.

For more information on the podcast series, see here.

To find the previous episodes of Differentiated Understanding, see here.

Chapters

00:00 Introduction to OpenClaw Frenzy in China

02:30 Shuyu Zhang’s Journey and Insights on AI Accessibility

05:27 Cultural and Societal Factors Driving AI Adoption in China

07:48 Understanding OpenClaw’s Popularity and Usage

10:36 Exploring Tencent’s AI Product Ecosystem

13:33 QClaw’s Integration and User Experience

15:45 The Philosophy Behind QClaw’s Design

18:14 Raising the Claw: The Concept of Personal AI

22:48 Incremental Value in AI Products

23:33 User Experience as a Priority

27:55 Understanding User Needs and Safety Concerns

33:07 Business Model and Global Expansion

Transcript (AI-generated)

Grace Shao (00:00)

Thank you so much for joining us today, Shuyu. I’m so excited to have you on the first day. I understand it’s such a pleasure to have you here. I’m really excited for today’s topic because it’s something that’s kind of been at the top of mind for a lot of people. Why was there a massive open call frenzy in China? Why did it take off? The thinking behind it all. How did WeChat go about opening up to this whole new era of agentic AI within WeChat development? And there’s no better person to talk about this topic than you. So first of all, for the audience, please start with telling us about yourself, your role. How did you end up here?

Shuyu Zhang (00:35)

OK.

Hi, everyone. I’m Shuyu, product lead of Qclaw. Also, the architect behind it’s overnight growth story in China. I achieved breakout success with zero marketing spend. I got a master’s degree of finance from Washington University in San Luis, then worked for Alibaba Group as head of AI product in Sanyo for four years. We mostly crafted AI product for business there. Almost everything we did was about AI for work. But one day, I decided to do something different. I want to get more exposed to consumer side.

to experiment with the chemistry of AI for common people. Tencent is famous for its consumer side products I want to work with and learn from consumer product experts here. So I came here last September. Yeah.

Grace Shao (01:15)

Really cool. And I think one thing that really kind of resonated with me is when we’re talking about AI agents right now, how a lot of times, you know, the products still don’t feel that intuitive for non-technical people. And you yourself joked and said, you know, look, you’re like a social science and liberal arts student. You’re not a technical person yourself, but you’re able to lead the product design for this. And your mission is really to make QClaw more accessible to non-technical people. Tell us about that and the thinking behind it all.

Shuyu Zhang (01:43)

Okay, okay. Actually, the story starts when I was working for Alibaba. Initially, we worked for the engineers to help them improve their working efficiency with AI. But later I found out that this group is overly served. They only account for like 1 % of all of the people, but

Every day there are lot of products designed exactly for them. And I think this is still the major theme of AI revolution since 2023, where people think the way to AGI lies. But these groups also are easily unsatisfied. They raise a lot of questions about the services that the AI provided, and at the same time, they’re afraid to be replaced. So I think the vibe is strange. So when I’m moving to a new environment to Shenzhen, when I was hanging around, I found a lot of interesting common people are using AI.

And they’re getting a lot of happiness and convenience, even though the product capabilities for them are not the most advanced and designated for their needs. They’re still satisfied. There are two interesting stories. The first story is when I was getting home on the plane during spring festival, I met a 12 year old girl. This girl looks smart. She was playing with some toys. And after that, she was playing with a chatbot during waiting for the plane to take off. I was shocked.

Because in the past, when I am younger, we usually play games during while we’re waiting for the plane to take off. But the youngest generations are playing with chatbots. This is interesting. And she’s using it to either cut the photos and even make phone calls with the chatbot. And when I talked to her, he also told me that when their friends are hanging out, the 12 years old girls are hanging out. They’re playing with the chatbot too. And I was really shocked by that story. And the second story is during the early January, Yang Liping was doing opera in Shenzhen. And I went there too. And a 50 year old lady sitting next to me, I cleared at her phone screen. The phone storage was actually out of use.

But she didn’t delete the chat box applications in her phone as well. She only reserved WeChat, TikTok and the chat box applications in her phone, even though it’s already full because she might not use the very advanced phones. Yeah. And then I found out that for common people, the requirements or the needs for AI exists as well. And problems widespread for the problems, not widespread, for the problems widespread, but not complicated. But the supplies for them are far from enough. And I know there are a lot of people are chasing higher and stronger AI, but the gap between the bottom and the ceilings, 90 % of the people are between them. I want to make a product that can feel and satisfy these people’s needs. Yeah, that’s my story.

Grace Shao (04:30)

That’s really interesting. think it really ties into a theme that we’ve been writing a lot about AI-prone, which is also about how China is really embracing it as a full on mass market product versus I think in the West right now, AI is still really used by a select group of knowledge workers or certain kind of demographic. I do want to double click on that, which is what do you mean by the young girl is playing with AI? What was she doing actually?

Was it interactive with their friends or were they trying to build products or what were they doing?

Shuyu Zhang (05:00)

Okay. The young girl, when she was playing with the chat bot applications, she actually sent a photo of her roughly taken, not in a very good light or in a very good background. And she just asked that, asked that application to curve it for me to make me look prettier or make me look funnier. She’s not actually a, because I asked, I also asked her a very interesting question. I asked, do you post TikTok shorts or Instagram?

She said, I don’t because I don’t like to show enough my life to the public, but I just like to see my photos in the funny way. don’t even though I use AI to, you know, process my photos, I’m just enjoying it by myself. I don’t want to it to other people. And the happiness of the AI processing of the photos is already enough for her. And this is the first scenery she’s using. And the second scenery is that she actually stays at school all the time. She didn’t go home during the from Monday to Friday. So she told me when she missed her mother, because her mother worked in Shenzhen and is a working mother, her mother doesn’t have a lot of time to, you know, FaceTime with her or the teacher doesn’t allow that as well. Because when she go back to the dormitory, the roommates are silent. They can’t do that, but she can always talk to that chatbot. It’s like a companion. And that also makes me feel warm actually, but also little bit sad for her. And the third scenery for her is that she told me she would call the chatbot because the chatbot never blames her and the chatbot always holds her words because sometimes for I don’t know, for the young generations, a lot of their topics are hard to get for the friends, but She said, chatbot is always a good friend because the young generations, they don’t actually care about the or they don’t know about the appearance of people or the words behind the words. But the chatbot is always blunt and sincere and always happy to chat with. Yeah, this is the three scenarios she’s using it.

Grace Shao (00:00)

Following up on our previous conversation about why Chinese people seem to have a much more optimistic approach to AI and why did the open claw, why did open claw take off in China, like such like wildfire.

Shuyu Zhang (00:15)

Okay, so I think Chinese people in general, embrace technology with open arms. They have a strong, better self mindset. They believe that new tools can help you learn faster, work smarter and live better. People want to upgrade themselves. And the second point, and also this is a key, for over a decade, Chinese tech companies have been quietly lowering the barriers using technology. They make complex things simple. You don’t need to be an engineer to call a DD or take out on Meituan or buy anything online. It just works. So people naturally expect that new technology will make life easier, not harder. That’s exactly what Qclaw and Tencent’s lobster products did. They took OpenClaw’s powerful but geeky core and wrapped it into a simple IM plugin. The barrier to entry dropped from weeks of learning to 10 seconds. That’s why lobster caught fire in China. not because it was the most advanced AI in the world, but because someone finally made it useful for everyone.

Grace Shao (01:13)

I think that’s a really interesting take and I think in general it kind of ties together to the bigger kind of sentiment as well where overall the reputation of Chinese big tech such as Alibaba, dance and tents that still are perceived quite positively by the average person to be an employee there is something very prestigious, ⁓ very like sought after. Whereas in the US, I think in the last couple of years, there is a bit more contention or negativity around the big text, whether it’s monopoly or behavior or even the capital allocation that they’ve really received that’s unfair compared to the rest of the country. overall sentiment is a bit different. I think that really did contribute to this as well.

Shuyu Zhang (01:58)

Yeah, exactly.

Grace Shao (07:00)

I think that’s really eye-opening and I think I’m not trying to hijack this whole conversation, but to me when I hear that, I think I want to like you said, the companionship is really great and the ability to help the child feel more connected to her mother is great. But at the same time, I do feel like there is some concerns or worries about that. ⁓ I did notice that the Chinese regulators recently pushed out some regulations around actual child use of AI.

which we’ll have an expert to join us one day to talk about this. But I think it’s interesting to showcase a phenomenon of China really embracing this at a mass scale. On that note, I don’t want to go too deep into the child use today, but on that note, I do want to ask you, why is it that China seems to be so amazed and enthused about AI, and especially this time with the open-claw embrace?

Obviously at Tencent’s headquarters, saw pictures going viral where people were lining up and getting open-clawed saws. Various big tech, whether it’s Alibaba or Baidu pushed out similar products like yours, like Qclaw. Could you explain to us from the big picture, is it cultural? Is it societal reason? Is it a top-down policy reason? Is it commercialized as a business reason? Is it product design, like you said? What is it that really drove like everyone going AI.

Shuyu Zhang (08:16)

Okay, so the first background of Chinese AI is that since 2023, since the chat-chip goes out and after that Baidu, Alibaba and also Tencent and also most importantly DeepSeq, it’s widespread of AI, their LLM makes AI widespread in China already because during last, the one before the last Spring Festival, almost like 200 million people use DeepSeek every day. So the basic foundation of people knowing AI in China is already widespread. And why OpenClaw is also going wild in China this year? Because OpenClaw’s capability is quite different from other products that people are familiar with currently. So this is the first decision point. more about the culture thing.

Firstly, China is a fast developing country and Chinese people are diligent by nature. And every generation, people of every age want to be a better self. So during this race, anxious middle-aged actually is a very big contributing factor. And also fear is also a big pusher. They always fear, you know, when they’re getting old, they will be lagging behind. So they want to, you know, learn more things.

learn what the new generations or the techie guys are doing. Yeah, this is the first decision point. And the second point is that hiring a personal assistant or secretary here is not common, but everyone wants to be an emperor because there are so many operas and TV series and soap opera shows recently about how the past generations, how the ancient times, how the emperor times. Yeah, everyone wants to be an emperor. So texting a message, gets the people done your job. Everyone wants it. So as long as you make the product simple enough and convey strong similarity with being emperor by sending messages through WeChat, really suits people’s taste. Yeah, I think this is two big factors about culture and society reasons here.

Grace Shao (10:18)

It’s interesting because basically they’re saying involution itself needs you and has made everyone want to adopt a new technology faster. It’s something I’ve never thought about. On the second point, am curious, like jokes aside about the Emperor thing. What is it that like the average person, like what are they using OpenClaw for though?

Shuyu Zhang (10:36)

Okay, actually there are two big categories. The first category is still for the common people because they don’t, even though they know OpenClaw is wild, they don’t know what to use it about. So they still use it like Yuanbao or Doubao or the other phone. They just like asking, yes, yes, yes. They still use it like the chat bots. the second part, they already use it in a more advanced way. They used it to earn money.

Grace Shao (10:53)

which are chat bot products. Yeah.

Shuyu Zhang (11:04)

Like for example, ask QClaw to seek jobs for them, like scanning the boss or the Liepin website and apply for the job.

Grace Shao (11:12)

Which are LinkedIn, Chinese LinkedIn, Craigslist, or indeed kind of like websites. Yeah. Okay.

Shuyu Zhang (11:18)

Yes, yes. And the second part, they use it to operate the social media account like Red Note or they even use it to operate X account, get some posts from X and watch it to write on other social media platform. Yeah. And the third condition is that they actually, trying to use it for like investing suggestions, how to invest in some stocks, what’s the price to get in and should they keep it or sell it.

Yeah, these are the main categories, but they are all about making money.

Grace Shao (11:49)

I see. That’s what fascinating. I want to get into case studies a bit more later. But before we get into this further, I want to help our listeners understand, can we, provide a base framework? Right now there are five claw bought like products within just Tencent ecosystem. And Qclaw is one of them, right? Which falls under WeChat, the product WeChat. Can you help us understand like these products first?

Shuyu Zhang (12:13)

Yeah, sure. The five claw product here is different for the users and are different between the target users. First, Qclaw and also WorkBuddy, we’re targeting at consumer and the Lighthouse, they’re targeting at enterprise side needs. And also there is a product called Claw Pro. They’re targeting at enterprise for the enterprise who want to make their stuff. Everyone has an ⁓ enterprise size claw and also the cloud desktop they’re targeting also at the enterprise side. Yeah.

Grace Shao (12:44)

I see that that’s just a good framework to have. So, okay, let’s get into the product pieces and your strategic intent then. OpenClaw is the open source framework ecosystem, while QClaw is Tencent’s package localized layer built on top of it. Can you actually help us understand how that works? What does it mean to have an OpenClaw integrated into a Tencent ecosystem?

Shuyu Zhang (13:04)

Okay, sure. Okay. So open cloud is actually a package of codes. If you install them on your laptop and connect it with LLM APIs, you can have a personal assistant already on your desktop. But that were required to handle like command lines, which is very technical skill. Even though I’m a product manager, I don’t know how to run command lines before after my engineers taught me to. Yeah. So purchasing and also purchasing APIs from providers is also not familiar for common people.

and also connecting with WeChat channels or like the other channels is also not that so easy before we do it there. OK, so we made all of these coding execution into visible and simple product features, which are already educated to common people. For example, we made the channel connection by making it just scanning a QR code and you can already get a QR code onboard on WeChat.

Grace Shao (13:33)

Mm-hmm.

Shuyu Zhang (13:59)

And also we make all of the LLM API purchasing processing invisible. We don’t need them to purchase a game. We are reincarnate in the product. And also the installation part in the past, people might need to, know, NPM run open claw, but now they only need to download the applications and double click it’s on.

Grace Shao (14:19)

see that that’s really helpful for people who don’t understand technologies, understand how this works, including myself even, I was a bit troubled. So I want to understand the thinking behind the QClaw product, right? You kind of talked about it, the frankly, the more cultural aspect of like how this came about and your own personal mission. What was the business decision really for WeChat? Like why did WeChat push out QClaw?

Shuyu Zhang (14:43)

Okay, so the thinking behind QClaw, how did that come out? Actually, aside from the thinking that I want to build a product that is easy enough for common people to use it, I still have the following thinking. The first is, ⁓ what’s the vibe of the product we should use? Is it work or life or both? Because the chosen, the choice of the vibe will be different for different people. How do we categorize that the work it can do for us? Because in the past, I think all of the AI products categorization are hard to get because most of time you just categorize, for example, something like finance or.

⁓ work usage or something like information gathering. my God, who knows that? So, and also there are a lot of times that work and life are mingled together. So if you’re designing an agent or a product features for everyone, if you’re trying to do that, that’d be hard. For example, if I want to build a finance agent, common people might just want to, you know, like search the stock price of something for me.

or recommend whether I buy or sell. But if for the professional users, their requirements will be higher. For example, they want, they wanted to, for example, write a quantitative trading strategy for me or something like that. Actually, the depth of capabilities providing would be hard to define here. And also it may discourage users if not handled properly because ror example, if for the pro users you’re designing it too easy, they would think, this is useless. This is far from replacing my interns or something like that. And if it’s designed to be too complicated, the common users would think, my God, this problem is not just designed for me. I don’t deserve to use that. And people are born to have fearness towards their unfamiliar domains. So I think technology or products should

undermine these fears or lower the barriers here. So in that way, we actually divide the categories of QClaw in three ways, which will be in international version. We categorize in three ways. First, QClaw it up. For the things you don’t want to do, but I have to let QClaw do it for me. That’s QClaw it. And also,

QClaw daily for the things I need to do every day, but I don’t want to forget a break. And the third QClaw up for the things I can achieve by myself and the expertise support. We divide the categories in these three ways. So everyone would have the it daily and up requirements. And we will also be more flexible or more concentrated focus on what types and what level of capabilities we’re providing.

This is the first thinking behind that. Yeah, because I think that the categories are complex and I want to make it simple and direct and focused to the people’s needs. They will know this, oh, this is for me. This is not too hard or something I don’t deserve. This is something I deserve to use and it would really help me. And the second thoughts behind that is, it’s a line we draw in selecting the building clause for agent.

since we already raised a lot of full grown claw. Because in China, a lot of people find it hard to raise a claw. They have to educate it. They have to do a lot of configs. So it’s hard for them to raise. But we...

Grace Shao (18:05)

Sorry, one moment. Explain the

context of what raising a claw mean. Like in Chinese right now, the buzzword is 養龍蝦. Explain to people what that means.

Shuyu Zhang (18:14)

Yes. So raising the club, a lot of people for common people, they were thinking like feeding all of my knowledge, what I know, who I am, what I want, what I like to it. They take into your input and they know what to generate in this mind. And also there is a mechanism called dream during the dream. They were, they were rethink about everything you, you told it today and they would generate something called memory.

And in the later usage, they will use this memory to know you better. So you will know that after daily’s inputs, after daily’s talk with it, your claw will know you better. And every instructions you give it will be better than the common AI products that don’t know you. That is called Yang Longxia or raising the claw.

Grace Shao (18:59)

It’s so funny. It’s

basically providing the technology, the context, but then when it gets better, better people say they’re like lobsters are growing, growing. It’s a funny analogy. I don’t know how it caught up, but it’s hilarious. Yeah. But yes, please continue. Thank you for that context.

Shuyu Zhang (19:12)

Yeah, sure.

Yeah, yeah. And I also think Yang Longxia or raising the claw is interesting because it’s like raising a kid or raising a pet. Because in the past, I think everyone still remembers there is a product called QQ Pet on QQ of Tencent. And I think back in my days, I was like seven to nine years old. I also raised a pet by myself, even though it died twice. Yeah.

Yeah, I really enjoyed raising a pet, like feeding it every day or just bring it up on the website and see it on my desktop every day. I think that is interesting for me. And I think a lot of people would want that this kind of companion of AI and they also enjoy the feeling that something is getting smarter or clever because of them. There is connections between them and the AI. They will bring a great joyness here.

not just something, for example, something that is already very supreme, high end package well and bring it to you and you just use it. You don’t feel connections with it. I think this is a very different feelings here. OK, so keep going with my point. ⁓

Grace Shao (20:22)

No, I think it’s funny

because I do think you touch on something basically like how Chinese tech companies gamified as well. So that’s how it also helped the mass market adoption that we were talking about earlier. And you just reminded me like when I was young, we all had Neil pets. think any millennial people in the West would know that. And like, even though they’re a virtual pet, you actually had a strong emotional connection with it. So I kind of see what you mean by this where

A lot of people might have not even found a purpose or use case originally, but even just building that context, that relationship with the AI that actually helped, you know, adoption rate. then sooner or later you try to find ways to make it more useful. Right. But yeah, please continue.

Shuyu Zhang (21:03)

Yes.

Yeah, exactly. OK, so after I saw that trend and also I saw the problem here, I think if we’re giving them some ground-claw, we should bring incremental value or incremental user experience to these people. And because this is important for the product perception, because if we’re the same like the chatbot about what we can do, people wouldn’t think or people wouldn’t take it seriously.

people will still think, okay, this is just another chatbot, but I want to bring incremental value to them. For example, before, in the past, when people want to make some travel plans using AI chatbot, they can only say like, I’m going to like Shenzhen for three days trip. Can you design a trip for me? Because in the past, like even though AI might be different in every answer, there are 80 % of the answers are similar.

or told you like to go to some park or some supermarket or some something like that or the hotels to say but if you’re using claw it would told you okay based on your fondness based on your habits based on the things you told me before i think some blah blah blah hotels would be better for you and some restaurant which is closer for example to your hometown or the flavor is similar to what you have told me that you like

This would be the incremental value. also the claw can also like book the flights, book the hotels, or just, you know, make transactions with the restaurants and ask for example, I’m going to spend my birthday there. Can you arrange something for me? All of the things is incremental value of QClaw AI can bring to people compared with the common air products. And I think this is the second thought behind it.

I want to do something that brings incremental value. Yeah. And I think these are the considerations here.

Grace Shao (22:48)

So I think one thing that’s quite interesting is Qclaw is obviously built out by the Tencent team. However, it can be accessed not only through your own WeChat Wecom, which is the WeChat Enterprise product. It can also be accessed through ByteDance Lark, which is like the Slack product within ByteDance. What is the thinking behind that? Why did you open up to your competitors essentially?

Shuyu Zhang (23:06)

Okay, so these are all channels, channels where people are already living in. We want QQL to keep company with users either in life or work or any interface. So limiting any channel will bring inconvenience to users. For us, the user experience is ultimate mode. So we don’t really care about the other, you know, so-called the business consideration. I think the user experience is the most important thing for us. Yeah.

Grace Shao (23:33)

That’s a very WeChat answer. feel like Alan Zhang has been kind of known to always prioritize user experience over any other kind of thinking, whether it’s commercialization or even, you know, sometimes functional adjacency within other products within the Tense Umbrella. So interesting. OK, so I have more questions. Something we talked about prior to recording was that you said WeChat is the default entry point for QClaw.

And that’s still a huge mode or advantage for you guys, especially certain features that you guys introduced, such as like scanning the QR code for downloading or installing a claw has been a big selling point. Walk us through kind of why is that and why WeChat, QCOP being built within WeChat is something so powerful.

Shuyu Zhang (24:22)

Okay, so actually there is a very simple reason before that and after that I will explain a more complicated one. The very first simple reason is that during the initial launch, we only had five engineers and me in the very first place. And we started to develop this product after spring festival, but we launched it in March 9th, I remember. So the time is very limited, but there are so many things to be productized.

of OpenClaw. What is the choice here? Because I want to make it simple. I want to make it user friendly. I want to make it widespread. So the first two things need to be adjusted by us. We need to do deletions. And the third part, we need to do adding items. Because before that, actually WeChat is not supported by OpenClaw officially. So we just

watch all of the files on the WeChat open platform and found out that there is actually a way to connect WeChat to OpenClaw. Then we just implemented it. And actually we don’t have more time to do that. So I think, okay, if I only have one time to make a decision or I only have one chance to select the channels, what would I choose? Of course I would choose WeChat because not everyone use some other working messaging applications.

But almost everyone in China use WeChat, even though whether you are like four or five years old even, or you’re 70, 80 years old, everyone use that. so that is why I choose it for the first channels. And the second reason is that people are mostly adjusted to use WeChat. And WeChat can be also connected through scanning QR code, which is the user experience other products can give them.

because other products, you see it on the light, like the tutorial, you need to go to the open platform of that product. need to copy your user ID, which is a very long link, and you need to copy the token or pin or something that is so technical terms. People would get scared by that. But we already have a very simple user experience, ⁓ user interface method that is getting QR code.

Why don’t we just do that? And also, also through the past years, mobile payment through scanning QR code is also widespread by Tencent. Tencent made scanning QR code and making payments widespread in China. And everyone, either they’re like the merchandising, the business, big business or small business, they can have their own QR code. And everyone is

used to scanning the QR code and connect everything. That is why we choose WeChat as the default selection.

Grace Shao (27:04)

I think that’s interesting because it’s like basically what headlines been missing. A lot of it is also just the native user experience and the ease of people to even access this kind of new technology because I think like you said, a lot of people are not maybe not scared, but it’s intimidating, right? It’s intimidating to try out new things and a lot of what’s out there in the market right now feels very technical. And if you’re not a technical person, you’re not following the progression of AI like closely. It’s

It feels intimidating to even try these new products out. Yeah, so I want to bring the conversation back to the real life usage, right? Because that’s a thread we kind of been talking about throughout our conversation. You really focus on bringing AI agents to the average show, the average person to the mass market. What is it that people really, really want out of this? Is it?

purely for gimmicky use, it’s for fun. You know, the case that he brought up with the young girl is obviously very interesting, but I would assume that’s not the mass demographic, right? Is it for consumer convenience, prosumer productivity? Is it for small medium sized businesses automation? I wanted to understand that. And then I want to expand beyond that, which is, are they not concerned at all about safety or privacy when they’re using these products?

Shuyu Zhang (28:22)

⁓ actually I think, yeah. Okay. So, ⁓ the, the first question I’ll explain here is what is a real target use case here? Either it’s consumer convenience, prosumer productivity or small business office of my automation there, right?

Grace Shao (28:23)

So two parts.

Shuyu Zhang (28:36)

Okay, so for me, it’s still consumer convenience, or we don’t categorize in this way, because we look at people as the subject, what the people need to do and how we can make it smooth and convenient. People can have different requirements in different conditions, either it’s on life or it’s on productivity, or some more serious or related to the business. We make integrations and push the ecosystem to provide the rest.

We will, like for example, if you see QClaw international version, you will see already put like some, I mentioned before the QClaw and QClaw Daily about for example, either you’re seeking jobs or you are operating your ex account or you are, for example, your career pop fan and you want to search for the concert tickets or chasing all of the information behind the hero. This is the conditions we provide. And we will also push the ecosystem.

For example, we already have a lot of ecosystem supporters here who provide the doc intelligence, like providing, for example, scanning contracts, the receipts, scanning something like that. And also there are ecosystem friends who already provide video generation, something relevant to, for example, creative parts or the designer shop. Yeah, we have a lot of ecosystem friends.

providing here. So we are still starting from what the people needs in aggregate, what people needs aggregately. so this will, the second question about the safety problem here is that we actually provide an incarnate safety features called the AI gateway or in Chinese Longxia Guanjia. That is something relevant to our team because my team is a Tencent PC manager.

which is a very, very, very old product. I think many people who knows this product might be like 30 or 40 or even older. Yeah, this product was built in 2004. remember, might not be correct. So it’s still in the computer-sized format, but in the past, it actually used a lot of safety capabilities.

Either it’s like preventing prompt injection or skill security or a lot of file security. It already has a lot of capabilities in it. And it’s also vetting the possible cyber attack over the internet. So we will know what’s the risks here. So we already provide a gateway in the product that everyone can be protected under this gateway.

they will not be exposed to the risks on the internet already. And also there’s point that, yeah, and actually.

Grace Shao (31:18)

I see. that actually protects

them more than someone directly installing an open claw themselves, right? Like going through cue claw is a lot safer in that sense.

Shuyu Zhang (31:28)

pardon?

Grace Shao (31:29)

So it’s basically a lot safer for the average person to use Qclaw than installing their own open claw on a Mac Mini or whatever. Because there’s not that kind of safety guard rail built around it, is what I’m trying to say, yeah.

Shuyu Zhang (31:37)

Yes, yes.

Yes, yes, because common people actually initially if they’re as long as they’re on the internet, they’re exposed to these risks. That is worse when they’re using open cloth without any protection. But

Grace Shao (31:49)

Mm-hmm. Yes.

Shuyu Zhang (31:55)

There’s also interesting part is your computer actually don’t have a lot of information and you don’t have a lot of money. So you’re not actually a target. Yeah. But we provide enough to safeguard for these people. And for the more important issues or more severe issues, part of this will provide higher levels of security. Yeah.

Grace Shao (32:04)

Yeah.

I see, Yeah, I think, you know, that’s a really good big picture on just QClaw’s build out and why China really took on like QClaw at a mass scale. I want to understand your business model and long-term implications. Obviously, I understand, you know, you guys fall under Tencent. It’s extremely lucrative business in the cloud side, the gaming side, obviously, we chat advertisement, etc.

You guys might not face the pressure to make money from this product, but I still want to understand how does it work? Will QClaw be free basically forever? Will you guys introduce a subscription model? know, I know recently you’ve even talked about you’ve been traveling around the world a lot. You’re in the States. You’ve been in Europe for some time. Are you trying to go global with this product? Are you trying to sell globally? Is it to enterprise? What is the kind of business behind thinking behind this?

Shuyu Zhang (33:08)

Okay, so the firstly, it will not be free forever, but we will always provide some free tokens for the first time users because they deserve to know what it can bring you, what increment value it can bring you. So we will provide some free tokens here, but we also provide different subscription plan here to support different layers of requirements. But I don’t think the token fee will be the only monetization methods here if we bring people’s whole life here, just like the...

WeChat strategy as well. Yeah, because selling tokens, tokens currently is really expensive. Even it’s a huge company, will still face the pressure here. Yeah, but a lot other modernization methods here, but we will also be very cautious here to trading between the experience and the modernization here. And this is the

Grace Shao (33:46)

Mm-hmm.

Shuyu Zhang (33:57)

the answer about the subscription plan and the charging plan. And also we are traveling abroad. We are trying to go in global. And this month, exactly April, we’ll be launching internationally and we will be starting from some main regions, North America, of course, and Asia Pacific. And then we were spending into more areas. Why? Because I traveled a lot in the last years. Everywhere I went, I would stay there for like a month.

And I would thoroughly experience the real life there and talk to the people there. I found out that even though there are differences between people’s life, of course, but people everywhere share similarities in needs, and they also have curiosity towards others. So people are bringing a better way to live with themselves with QQLO, like QQLO A, QQLO WAP, and QQLO Daily. So they can share it with others.

on Qthaw and benefit from other people’s sharing. That is why we are bringing it globally. And every states we go, we will co-create with people there and pass on merit to more areas.

Grace Shao (35:00)

Yeah, so I think it’s really interesting because obviously QQLA has a huge advantage in China because it leans into the WeChat ecosystem we talked about a lot today. But what is your advantage when you are going global? How do you compete with international competitors?

Shuyu Zhang (35:16)

Okay. I think the first important part is still the use experience because I use a lot of global applications because I, my job is AI product manager. I would find a lot of product even complicated for me. I think the product experience is not, hasn’t been done very good yet. There are still a lot of bugs. There are still a lot of, you know, complex, complex items, complex terms or complex workflows here.

I want to make it all simple. And integration would also be a great part here. And the third part is the community advantage. would, you know, because we are not so intimidating in the image, we will not be like, we are the tech guys. We are the advanced ones. If you don’t know it, you’re the stupid one. You’re going to learn from us. No, we will not do it like that. Yeah. Yeah. We would co-work with the creators. For example,

Grace Shao (36:05)

You

Shuyu Zhang (36:11)

⁓ For example, when we’re spending to Japan, will co-create with some local, like fan, big fans there, or the ones who knows well about the food there, or the one who knows about the job marketing there. And we will make it more localized and make the people who really use it create that. We will co-create with them and bring it.

to their community because that would be the way that the community gets the concept or gets the usage of AI in fastest way. Yeah, I think so. basically.

Grace Shao (36:44)

Yeah, I see what you mean. Like you marketed a much more accessible product than other peers maybe on the market right now, which are targeted for again, relatively niche demographic. So that brings me to the next question, which is like, do you think then the messaging apps with a chatbot like kind of interface will become, will remain the default or will we see a new kind of interface layer for agentic AI and how we call on them.

Shuyu Zhang (37:16)

think, actually, I don’t know about this answer because messaging apps can be, they can evolve as well because they have a lot of engineers as well and they have a lot of ⁓ intelligent people there and they care about the user experience here and Asian products can also evolve. The true interface layer, I think is dynamic and there is no fall or lamb in the current arena. So I actually don’t know the real answer of the, you know, who are the final interface layer, who is the true one? I don’t have the answer yet because I see a lot of interesting apps evolving from AI agents, but they’re trying to, you know, cutting into the messaging apps interface. And also there are a lot of messaging apps there. for someone like the X or I don’t know what’s his future plan, but I also know there are a lot of messaging apps who are cutting to an agent domain. the answer is dynamic, I think.

Grace Shao (38:13)

I appreciate the honesty and the humility actually. I think no one really knows the future right now. The speed of evolution of industry is insane right now. And I always hear people who really like, you know, like even the Ben Evans and the world, they’re like saying, if you think you really know the industry, you don’t really know the industry because you can’t possibly, you know, have a strong grip on what’s happening because things are changing so fast and so much happening constantly.

Grace Shao (39:05)

I have the last question, which is a question I ask every single guest that come on the pod. What is one differentiative view you hold? This can be anything about the product we talked about today. It could be about the industry. It could be about anything in life.

Shuyu Zhang (39:08)

Yeah.

Okay, one differentiated view I hold is that I think the most profound function of a superior AI, like the clock, is not to solve problems, but to reveal them, to reflect back to us the questions we’ve been unwilling to ask ourselves. Think of it as an archeologist of behavior. We narrate our lives, edit our memories, even lie to ourselves without knowing. But something like the clock observes

what we actually do, what we choose or what we linger on. It doesn’t judge. It mirrors. In the end, what it shows us isn’t its own intelligence. It’s us, our contradictions, our desires, the fractures in our collective consciousness. So the most meaningful conversation isn’t whether AI is becoming human, but whether we are brave enough to look clearly into this vast, mirror it holds up. And finally, see...ourselves.

Grace Shao (40:18)

I think that’s super interesting think it helps us. You touch on something, it’s like really helping us recognize blind spots. I just want to share a personal anecdote as well, which is like, think, you know, when I started AI Pro nearly two years ago, it was really hard for me to sometimes seek help from other people’s opinion because it takes people’s time, right? And then for reviewing of my work and I didn’t simply want a grammar.

like, you know, copy editor kind of grammar fix. So now what I did was I built an editor council and what I did is train the council basically to have certain perspectives, follow certain guidelines or, you know, ⁓ angles of the world and then critique my work and really help me recognize ⁓ blind spots I’ve been missing or my logic or my thinking that are not, you know, synthesized clearly or not.

flowing smoothly, things like that, that I just found it so helpful. In fact, in some ways more helpful than a human editor at certain tasks, because exactly to your point, humans have biases, humans have judgment, and not like intentionally, but just by nature, we all hold biases based on our own knowledge, whatever. But the machines basically are just can be very critical if you tell it to be critical and can kind of show you a 360 view of your thinking. that’s super interesting and I appreciate your sharing on that.

Okay, thank you so much for your time today, Shuyu. If anyone wants to reach out to you, how should they find you? If they want to learn more about QClaw, where should they go?

Shuyu Zhang (41:50)

my LinkedIn. The name is Shu Yuzhang. And they can also follow me on my X account, which I can share with you later. It’s also called Shu Yuzhang. Okay.

Grace Shao (41:59)

Perfect. Thank you so much. We wonderful conversation with you. Thanks again.

Shuyu Zhang (42:03)

Thank you, thank you Grace. Bye.

Many of you may know Zixuan from this earlier episode recorded for Differentiated Understanding. So if you’re curious about their business model/ the first LLM company to go public, check out the episode below.

Announcement

Btw, if any of you are in Singapore in June, please join us at SuperAI June 10-11/ or catch me in town June 9-12. Representatives from Chinese labs and big tech will share the stage to discuss their businesses and development. (And of course, I’ll be moderating;p)

Find out more about SuperAI and the speakers’ information here:

SuperAI is the largest AI conference in Asia, bringing together over 10,000 attendees and 1,500 AI companies from more than 150 countries at Marina Bay Sands in Singapore. The conference is where East meets West in AI – spanning robotics & embodied AI, frontier models, AI infrastructure, biotech & healthtech, finance, and AI’s global impact. SuperAI takes place 10-11 June 2026 as the anchor event of Singapore AI Week (8-14 June).

Find out more about SuperAI and the speakers’ information here.

ANDANDAND DON’T MISS THIS: 10% discount through my link!

A few quick thoughts on recent announcements and rumors

Tencent

QClaw, one of Tencent’s clawbot products, will launch its international version tonight, and I recorded an exclusive interview with their Senior Product Manager, Shuyu Zhang, which I’ll release tomorrow around 6:30 am EST. She dives into why QClaw, the product design behind it, the safety guardrails, the advantages Tencent and thus WeChat’s ecosystem have, and demystifies in a cultural and social context why OpenClaw took off in China.

Alibaba

Ok, and on to the next. It was confirmed that Alibaba’s former Qwen leader, Lin Junyang, had a philosophical clash with Eddie Wu, Alibaba's CEO, and decided to leave. Eddie wants to push on AI commercialization. It was said that Lin Junyang firmly believes in open sourcing for the ecosystem, so he left. However, Alibaba released an impressive video model last week - code name happy horse.

For many of the labs now, they’re under pressure to make money, and especially for Alibaba, a publicly listed company, it has a fiduciary duty to its shareholders. However, on Lin Junyang, I guess if he truly doesn’t vibe/ agree with this direction, I understand his decision to leave. I’m sure the discussion was ongoing for a while and not just a sudden heat-of-the-moment decision.

DeepSeek

Based on multiple conversations, it is widely known that DeepSeek will release a new model in the coming weeks. The expectation is that it will take the number-one slot on the usual leaderboards, but the margin over number two will be narrower than last cycle, when DeepSeek sat roughly 15 points ahead of the next open-source model. Worth noting alongside this is a separate news that DeepSeek is fundraising at $10 Billion-Plus Valuation per The Information’s reporting.

To read: Nvidia’s Jensen Huang warns Huawei chips for DeepSeek AI models would be ‘horrible’ for US and China’s AI firms scaled up on open-source models. The next phase may be different, both by the SCMP.

Subscribe now

Breaking down the monetization framework

So for the labs, the revenue model is really simple, as confirmed by the labs and investors; the thinking is simply Price x Quantity. The battle in the end is to find the optimal balance of who has pricing power and who can sell the most. Both variables have increased simultaneously, thereby explaining public disclosures suggesting triple-digit revenue growth at some Chinese labs, including MiniMax and Zhipu, though public ARR disclosure remains limited.

When I asked whether they are not scared that if the price goes up, then demand goes down? The analogy was awesome. “We’re no longer selling 二锅头s, we’re pivoting to 茅台s, so if we lose some customers, so be it. (the margins will be much higher).” How do I translate this? Basically, the former is a white liquor that burns in your insides and tastes like diesel, and the latter is what is used at state banquets and fancy people's weddings.

Is pricing power everything?

On the price side, China’s leading AI labs have been quietly nudging prices upward, especially Z.ai, which has explicitly announced. But it reflects a general trend in many of their coding subscription plans and the per-token rates they charge developers through their APIs. They are doing this from a position of scarcity rather than strength of marketing: demand for their models is outrunning the supply of GPUs they have to serve it. That changes who gets prioritized. Customers willing to pay the full sticker price, and especially those willing to sign multi-year commitments, are being prioritized or preferred over those hunting for day-to-day discounts. In plain terms, according to an investor, the labs are choosing better-quality revenue over bigger-looking revenue, which is the kind of move a business makes when it is trying to protect its margins rather than chase growth at any cost.

The subtler point, and the one most outside observers miss, is that the headline price of tokens is not really what generates the revenue in coding use cases. Most of the billing in a coding workload comes from something called “cache hits.” Here is what that means in plain English, which was explained to me: when a developer works within a large codebase, the AI has to reread much of the same code repeatedly as it answers follow-up questions. Rather than charging full price each time, the labs charge a much lower rate for that repeated content, and that is the cache-hit price. Because so much of a coding session involves re-reading the same files, those discounted cache-hit tokens end up being the bulk of what customers are actually paying for.

This matters for how you compare Chinese models to Western ones. Depending on whether you compare input, output, or cached-input pricing, top Chinese models can range from modestly cheaper to dramatically cheaper than Sonnet. On output pricing, GLM-5.1 is roughly 70% below Sonnet 4.6; on standard input pricing, the gap is closer to 50%.

If you judge Chinese labs by the sticker price, you will badly underestimate how much money they are really making per customer and how healthy their margins actually are.

The open-source game theory

Despite a growing mix of proprietary flagship releases from Qwen, Alibaba has not abandoned open source. It is now running a hybrid strategy across closed and open-weight models. The reason is multi-party game theory.

In a two-player game, both labs rationally close. In a three-player game, one player can always choose to open-source specifically to deny profit to the other two — the prisoner’s dilemma outcome, where at least one actor defects. As long as one frontier lab holds the line open, the others cannot permanently close. So, even though under pressure to profit, it is believed that the Chinese AI ecosystem will not be closed forever. The read is that at least one frontier-grade open-source model will persist, which caps how high closed-source pricing can go.

It’s been widely reported, shared, and recognized that Chinese labs, in particular, lean toward openness for non-ideological reasons: they need an ecosystem, brand recognition, and community traction that a closed Chinese model cannot easily earn. The interesting take is that despite it sounds like a philosophical choice, whether to open or close, that debate is driven by where your economic interest lies: if you own compute, you want everything open; if you own cash, you want everything behind an API.

Competition intensifies, but no one is backing down

No consolidation is expected. This is where I realized I might be wrong, as I’ve been saying there must be consolidation happening later in this year at this burn rate. One of the lab people told me that researchers at each lab are strong, the frontier techniques are widely understood, and new entrants like Xiaomi are increasing competitive intensity rather than reducing it. But no one has an interest in dropping out of the race willingly.

The part where it did align with my recent analysis is that Moonshot (Kimi) is probably the best company and the best AI. The thing is, I was told that “good AI = good company, not marketing.” ;p

As we’ve written before, Moonshot’s founder, Yang Zhilin, was a student of Z.ai's founder, Tang Jie, so much of their vibe or aura can be characterized as technology-first and deliberately uncommercial. Maybe Yang a bit more cool with his rock band meeting rooms and what not. MiniMax is described very differently: SenseTime-lineage DNA, more commercially aggressive, more mature on capital markets mechanics, more willing to hire professional operators and play the fundraising game. That explanation, while not always investor-flattering, arguably explains the valuation trajectory better than benchmarks do. If the thesis is that the API layer is the rent layer, a lab unusually good at commercialization attracts exactly the kind of capital chasing that rent, independent of model leadership.

The endgame analogy

So I asked, " What is ‘to win’? The endgame, in the insider’s framing, will look like the auto industry rather than like a single winner-takes-most platform. Some will be Rolls-Royce with a low number of sales but extremely high margin, and some will be Mercedes, BMW, and Audi sit at the premium-meets-volume sweet spot. Mass-market players generate volume at thinner margins. The strongest model does not automatically win the commercial market — the winner is whoever maximizes price times quantity at their chosen tier. Developer, consumer, and enterprise positioning are each a bet on where that product peaks.

And for most Chinese labs, their simple goal right now is to win on speed over the big techs. They’re going to lean into lean organizational charts so they can iterate and ship faster.

1. Where is agentic AI actually taking us?

   We wrote about agentic shopping led by Alibaba, agentic work obvs with Claude Cowork and all, agents like Manus, Kimi Agents and the action loop and all. People rave about its capabilities, but I keep coming back to the same question: how much value is it truly driving? Is it not just the 1% using agents right now? How much of it is just hype at this point? It's hard to understand the real economic impact. I found online this kind of random report by the St. Louis Fed saying that AI-related spending contributed roughly 0.9 percentage points to U.S. real GDP growth in the first three quarters of 2025. Sounds meaningful until you adjust for imports — then it’s about 0.4 to 0.5 percentage points, or roughly 20-25% of real GDP growth as of Jan this year.

   And almost all of that is capex. Companies buying chips, building data centers, spinning up cloud. That’s spending on AI, not productivity from AI. The productivity flywheel everyone’s pricing in hasn’t started spinning. This is pretty symbolic for the whole industry across the U.S., but also in China.

2. So how do we justify these valuations? Price-to-earnings or price-to-dream?

   If the economic value hasn’t materialized, what exactly are we paying for when we price these AI labs? The IPOs of Minimax and Z.ai were initially said to be undervalued compared to the valuations of American labs at the time, but what profit have they brought in? What sustainable business model have we seen? Let me put some numbers on it.

   MiniMax generated $79 million in revenue in FY2025 and recorded $512 million in losses in just the first nine months, per its IPO prospectus. It currently trades at a market cap of roughly HKD 309 billion — about $40 billion according to their latest filings. There is no P/E ratio because there are no earnings. Zhipu did RMB 724 million (~$105 million) in 2025 revenue, up 132% year-on-year, but widened full-year losses to CNY 4.7 billion (~$683 million) according to its first post-IPO report. Its market cap sits around $44 billion. Shares are up nearly 600% since its January listing. Again, no earnings — so no P/E.

   For context: JD.com has a market cap of roughly US$39-42 billion with a trailing P/E of about 16x according to Yahoo Finance. Alibaba is valued at roughly US$297-372 billion, with a trailing P/E of 18-23x. Tencent, still everyone’s favorite (‘Tencent, is the winner, it has to be right?’ - quote from U.S. investor), is at US$579 billion with a P/E of about 18-20x. Look, I’m not a finance person, but these numbers don’t make sense to me.

   So MiniMax is valued in the same range as JD — a company doing over $150 billion in annual revenue with real earnings — while generating less than $80 million and losing (investing/spending) money at a rate of roughly RMB 2 billion a month. Zhipu is valued higher than JD on $105 million in revenue and nearly $700 million in annual losses.

   If you remember Alan, the hedge fund PM we interviewed a while back here, he coined a term - the ‘price to dream ratio’ when we were chatting about this valuation disconnect. There’s no price-to-earnings here. This is a price-to-dream ratio. And the dream has a burn rate that gives these companies maybe 12-18 months of runway from IPO proceeds before they need fresh capital or a lifeline.

   Am I not dreaming big enough, or is that saying in Chinese applicable to me right now, ‘我的（智商）贫穷限制了我的想象力’, my raw IQ is limiting my imagination.

3. What’s the clear strategy for any of these labs in China now?

   If open source can’t sustain a business and the pressure to monetize continues, the playbook defaults to API sales. But that’s a race to the bottom. DeepSeek’s input pricing sits at roughly $0.028 per million tokens — about 1/180th of equivalent GPT pricing. DeepSeek V4 launched in March at $0.30 input / $0.50 output per million tokens. Everyone else — Moonshot, Qwen, Doubao — is or let’s say was competing on “cheaper and almost as good.”

   The Chinese LLM space is the most competitive, with various labs carving a different angle: ByteDance on consumer scale, Alibaba on open weights, DeepSeek on pushing the frontier, Moonshot on agentic capabilities. Easily, the most competitive AI model market outside the U.S., with less investor money. But competitive doesn’t mean profitable. At these token prices, I don’t see how standalone API revenue covers compute costs, talent acquisition, or any of these expenditures, let alone how that can fund frontier research.

   However, on the other end of the spectrum are people saying that people complaining about Anthropic making OpenClaw more expensive are just a bit delulu right now. The argument that once VC subsidized infrastructure disappears, the massive supply crunch is coming, and users will either need to host on their own device or leverage more open source or switch to much cheaper paid options, such as Chinese models. And that will be a win-win case for Chinese labs.

   So this leads to the next question: Is this the end of China’s open-source AI era?

   Even Alibaba’s Qwen seems to be faced with this dilemma now. With an end product- cloud to sell, the team is also faced with pressure to justify the AI spend and lack of profit. If selling cloud isn’t enough to cover the investment into AI infrastructure, and open-source model development doesn’t convert to revenue, then the economics simply don’t work. I think we’re watching the open-source window close in real time. And the startup labs that survive will either find a hyperscaler patron or figure out a new business model that does not exist right now. Anyway, that’s my consolidation thesis. And here writes about the need for an open source consortium.

4. This is where the concerns around privacy and security sit.

   Open-source models can be jailbroken for malicious use — generating child exploitation material, assisting with bioweapons, you name it. There’s no self-regulatory mechanism in China’s open-source ecosystem, and that is likely now being flagged as regulations around the technology catch up with the technicalities.

   At the recent Hong Kong AI governance conference, academics from both nations suggested the realistic path for U.S.-China cooperation might be a joint response to a major security breach. But what would that actually look like?

   Anthropic just gave us a preview. They built Claude Mythos, a model so capable at discovering zero-day software vulnerabilities that they partially withheld it from public release and launched Project Glasswing to deploy it defensively on critical infrastructure. ’s latest article goes into this too. But if that ‘worst case scenario’ were to happen, then that is literally the kind of threat those leading scholars and academics from both China and the U.S. were discussing.

   For more context, writes about how this kind of general stroke of fear can be misguided. And a little self-promo from the event:

5. How much proper management is going through these labs? And how will they compete for compute?

   Now, as we know, many of the top Chinese LLM labs have spun out of university research, but without the right business guidance and mindset, and a much less vibrant VC fundraising runway, they’re rushed to go public. How ready are these companies actually prepared for public scrutiny, pressure to turn a profit, and does that pressure change the initial purity of tech advancement into something else?

   Beyond that is what we’ve written about, the obvious compute constraint and the fight for more. How does a lab fight against the hyperscalers? Compute is the binding constraint, and nobody’s seriously addressing how Chinese labs solve it.

   U.S. export controls, now as we move into the first generation of models trained natively on the Grace Blackwells, what happens now? That capability gap widens? Will Huawei Ascends be actually viable at scale? This leads to the next question…

6. If AI diffusion continues, where does that lead for jobs and employment?

   This is what keeps governments up at night. If the open-source era closes and AI concentrates in a few hyperscalers, the diffusion benefits to the broader economy narrow. But tbh, the worry isn’t about the top 1% — it’s sustaining the livelihood of the mass population. Should hyperscalers bear societal responsibility to create jobs, or at least not destroy them faster than the economy can absorb? If mass job losses happen, will there be a drop in fiscal income taxes? What happens to the governments?

   According to Microsoft’s AI Diffusion report, it wrote that in 2H 2025, “The United States shows that leadership in innovation and infrastructure, while critical, does not by itself lead to broad AI adoption. The U.S. leads in both AI infrastructure and frontier model development, but it fell from 23rd to 24th place in AI usage among the working-age population, with a 28.3% usage rate. It lags far behind smaller, more highly digitized and AI-focused economies.” South Korea is punching way above its weight (which we’ll have a leading VC walk us through this soon on Differentiated Understanding season 2). And of course, post DeepSeek, China’s diffusion accelerated at rocketship speed.

   But as diffusion happens, what happens to the job market, and what happens to the economy? Is AI really contributing to any GDP in a meaningful way, yet? And governments know that when the economy doesn’t work, other policies’ legitimacy sometimes finds itself with less support or legitimacy too. Especially in China, where political legitimacy is more directly tied to economic delivery, the stakes are sharper. This should be a discount factor on every AI investment thesis. Is ensuring AI does not disrupt the top knowledge workers more important, or is ensuring social stability through taking care of the masses more important? writes about that rising anxiety in China.

7. So, how much of the AI diffusion has really worked?

   I'm a little skeptical. The Claw frenzy felt fleeting in China. Where are the stories now? Tencent is still everyone's favorite because of its ecosystem story, but why are there five Claw products? That fragmentation tells me nobody has found product-market fit yet. I'm questioning some of my own analyses and obsession with distribution now. We'll have someone from QClaw joining us on Differentiated Understanding season 2, so I'll push on this.

8. The (capability) human gap

   The gap was first closed, then it was supposedly stretched again. According to CNBC's coverage, Anthropic's Mythos is soo advanced that it fears to release it to the public, worried that it could be maliciously used. How much of that is fear-mongering, how much of that is legit? Probably a bit of both. See Nathan’s write-up here.

   The company hit $30 billion in annualized revenue as of April 2026, up from $9 billion at the end of 2025 — a 3x jump in four months. They now sit above OpenAI 's~$25 billion ARR. Claude Code alone reached $2.5 billion in annualized revenue. The company closed a $30 billion funding round at a $380 billion valuation in February, and is reportedly targeting a $60 billion+ IPO in October at a $400-500 billion valuation based on Bloomberg's numbers.

   At ~$30 billion ARR and $380 billion valuation, Anthropic trades at roughly 13x revenue — expensive, but anchored to real, rapidly growing income. MiniMax would need to multiply its revenue by roughly 39x just to justify its current market cap at the same ratio. That's the distance between a price-to-earnings world and a price-to-dream world. And what about the capabilities gap? How will that be closed? I hear more and more - it's just the people, and some are ‘‘just the sun of the orbit" kind of thing, but of the AI-verse.

9. Why do some models do so much better than others? And where I'm placing my bet.

   Is it architecture? Data? Or is it the end of the day, just people? Someone at Microsoft working on their model, who has access to OpenAI's open weights, told me it comes down to talent. I kept asking, really though? With all those resources, the differentiating point is just people?

   Which brings me to Moonshot. Out of all the Chinese labs right now, I feel most strongly about Moonshot. There's something about their commitment to technology and not frivolity. They launched Kimi K2.5 in January and Kimi Claw in February. DeepSeek and Moonshot seem like the only Chinese labs that will lead in genuine research and are still committed to this. By default, big tech will be the winners in distribution and reach but... they're not leading the breakthroughs.

   So the endgame might look like the OpenAI x Microsoft deal (pre-blowup) replicated in China: closed research labs selling to or merging with big tech distributors. If my consolidation thesis is right, the question is who buys these labs and at what price.

   Then, a side tangent on Moonshot and Openclaw. During a roundtable discussion in Beijing two weeks ago, Moonshot CEO Yang Zhilin said that the industry is structurally shifting from a training-dominated era to an inference-dominated one. Now, the bottleneck is serving tokens at scale, cheaply and fast enough for agents that run for hours or days. (see original CN source here)

   And then on April 10, Yang Zhilin was one of the select business leaders invited to brief Premier Li Qiang at a State Council economic policy roundtable (an honor DeepSeek’s Liang Wenfeng received as well in January 2025). Beijing is effectively pre-endorsing the company as a national champion in the agent era, reducing perceived regulatory risk and signaling that the IPO pipeline for top-tier Chinese AI firms remains open and supported. A pretty special nod that Minimax and Z.ai didn’t get.

   There is something mystical about Moonshot. I want to learn more. Any intros would be appreciated.

10. So then what is winning?

    I keep circling this question. Is winning having the most raw compute and energy capacity? Then the U.S. hyperscalers win the first half, and China as a whole wins the second half of that equation. Is it building the best model? Then Anthropic and OpenAI are pulling away — Mythos proved that. Is it the fastest diffusion — getting AI into the most hands, the most apps, the most workflows? Then maybe China's open-source ecosystem/ Alibaba, and Tencent still have a claim, at least for now.

    But here's what I keep coming back to: diffusion doesn't matter if there's no economic progression underneath it. You can put AI in every app in China, and if it doesn't generate real productivity, real revenue, real jobs — then it's just adoption without value. And if there's no economic value, then even the most legitimate policies lose their footing. Governments can't defend AI investment to their populations if the returns aren't showing up in paychecks and GDP. And that's not a China-specific problem.

    So the thread through all ten of these points is really one question: can any of this translate into economic reality before the capital runs out and the political patience expires? Right now, the value of agentic AI is half proven/ half still hype. Based on the conversation between Luo Fuli from Xiaomi MiMo, Zhang Peng at Z.ai, and Yang Zhilin at Moonshot, the consensus is that the companies best positioned are those solving the token supply chain through efficient architectures, inference infrastructure, and agent orchestration that makes every token count. The next phase of AI value creation will be about making models work autonomously, efficiently, and for long enough to produce real output.

    As of now, China lab valuations are still running on dreams, not earnings. The only real proof of step-change capability is coming from U.S. labs with massive compute access. China labs can't replicate that under current (unfixable) constraints, thus consolidation feels inevitable. And the bet — my bet — is on which lab survives long enough to be the one worth acquiring.

    Someone help me out here…

    ---

    AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

But it’s been two months since that initial frenzy, now what? Why are Alibaba and Tencent throwing massive events to talk about their enterprise strategy of integrating this open-source software at their cloud summits?

Tong didn’t talk about the lobsters as a consumer product. He talked about them as the forcing function that finally made enterprises take agents seriously. How could this be an opportunity for Tencent, which has been a bit slow in AI?

“AI deployment is not just an algorithm problem. It’s an engineering problem.”

That line was the thesis of his entire keynote. And it turned out to be the thesis of both summits.

Tencent thinks it can win because it has better distribution and ecosystem than everyone else. Alibaba thinks it has a better developer ecosystem and cloud service than everyone else. Let’s rewind.

Two days earlier and a thousand kilometers away, about a hundred developers, cloud architects, and a handful of investors had gathered at the glossy space shuttle-looking office building that is supposedly the “most expensive business real estate in Hong Kong” – The Henderson, for Alibaba Cloud’s “Data+AI Forward” summit, co-hosted with AMD.

Although the target audience was much more client-facing and the discussions largely focused on data storage and add-on services offered by Alibaba, some themes were the same. It was equally about building the harness, context engineering, agent orchestration, skills marketplace, and enterprise governance.

Even MiniMax, the fast-rising model startup, appeared at both events in some capacity — at Alibaba’s stage as a case study in cross-stack data support, at Tencent as a guest speaker integrated into their TokenHub marketplace. The ecosystem is so intertwined yet so cutthroat at the same time.

What was clear was that the consumer battle was halfway done, or at least they realized that ROI is not going to be there anytime soon. As OpenClaw took off, a new wave of optimism swept through the management’s ranks. This could be an opportunity to court enterprise clients. Sell tokens? Full-stack service? Cloud, the foundation of it all.

The Consumer Party Is Over

For most of 2025, China’s AI story was a consumer story; we’ve written about it here and here. It was about capturing mindshare, super-app integrations, viral chatbots, the race for monthly active users between Doubao, Kimi, DeepSeek, and a dozen others.

Li Qiang (李强), Tencent’s Group VP and President of Government & Enterprise Business, put the scale in perspective. He started his keynote showcasing a big screen indicating that there are over six hundred million generative AI users in China, that is, every second netizen. Token prices collapsed by 99% over two years, from 50–100 RMB per million tokens to a few RMB or even cents. Daily token consumption hit 14 trillion by March 2026. These numbers cannot be ignored; they mean that, in some capacity, “AI for all” 全民AI has been achieved.

As we all know, since the start of 2026, the “AI for all” narrative has been taken to the next level, as we have seen a frenzy around claw installations. And “raising lobsters” (养虾) has become both a cultural meme and a product category. Millions of users traded deep system permissions for automated productivity. Every big tech under the sun scrambled to ship one. But the commercial lesson of that moment wasn’t about consumers. It was about their limits.

The consumer agent frenzy demonstrated that lots of people wanted agents. Nobody wanted to pay for them. Personal users will happily trade privacy for convenience, but they won’t pay enterprise prices. And open-source models have made the consumer-facing AI layer almost impossible to monetize directly — a reality that loomed over both events, even when it wasn’t spoken aloud. And the new battlefield, as Wu Yunsheng put it, was moving from AI that 'can answer' (能回答) to AI that 'can do' (能做事) — making ‘好用的AI’, AI that's actually useful."

Meanwhile, at the Alibaba summit, the recent leadership reshuffle within the Qwen team went entirely unmentioned. No one addressed what happened, why it happened, or what it means for the model roadmap. Qwen was talked about as the backdrop infrastructure support. The Qwen family was presented in its full regalia — Qwen3-Turbo, Qwen-Max, Qwen-VL, CosyVoice, and the Wan video generation suite. For a company staking its cloud growth narrative on open-source model leadership, this silence was, in its own way, a telling data point showing its strategy pivot.

The unspoken implication at both summits: open-source models are strategically essential but economically punishing in their own right. Maybe at this point, the pressure to see ROI in AI is becoming too heavy to bear.

What everyone knows is that enterprise clients are where the money has to come from. China’s internet era has always given excuses for a lack of willingness to pay. But when tokens become gas and water, and ‘AI is turned on like tap water’ as Tencent’s Li Qiang puts it, then maybe now, is when companies are going to pay, or must pay.

But my question is, why OpenCLaw now? Well, enterprise agents, governed agents, agents that live inside corporate workflows and generate cloud consumption, these are the ways Alibaba and Tencent intend to collect.

The WeChat question I have

When OpenClaw first became viral, WeChat did not integrate it. Even though it felt like Tencent led the strategic embrace and was the most hyped up about it, there was a lot of caution for WeChat, the main app at first. Allen Zhang Xiaolong (张小龙), WeChat’s legendary product manager — the person who has historically guarded the platform’s simplicity and user trust with near-religious discipline — resisted.

For him, the risks were that agents with deep system permissions operating inside WeChat’s social graph raise profound privacy, security, and compliance concerns. A misbehaving agent in a personal chat is embarrassing. A misbehaving agent in a WeChat group, one that can read messages, trigger workflows, and access contacts, is a liability at societal scale.

So what happened was that the first integration was through WeCom (企业微信), the enterprise version of WeChat. Now, this was definitely a conscious move. Enterprise use cases are more controlled — fixed user populations, IT governance, compliance frameworks, and defined workflows. An agent operating within WeCom has a bounded sandbox. The company sets permissions, the IT admin reviews skills, and the security layer audits every action. Wu Yunsheng (吴运声), speaking from the Tencent Cloud product side, laid out the enterprise security architecture in detail: four principles of agent governance — visibility (see how many agents exist), traceability (audit every action to who did what), controllability (hard limits on network and command access), and trustworthiness (security screening for every external skill). Enterprises can block agents from downloading external skills entirely, restricting them to an internal-only marketplace of vetted capabilities.

And why Allen Zhang initially withheld WeChat access.

But then Lark and DingDing integrated agents — both enterprise-first platforms with consumer-adjacent surfaces, both aggressively marketing agent capabilities. The competitive pressure became untenable, and the FOMO turned real.

If WeChat, THE platform that is Chinese digital life, didn’t offer an agent layer, it risked looking like the one platform that couldn’t do AI. But ‘doing AI’ has its risks.

So Zhang relented. Qclaw connects to WeChat via a QR code. WorkBuddy allows remote task orchestration from WeChat. But notice what is still not open: WeChat groups. (see here for more on Tencent’s claw embrace) You can invoke an agent from WeChat, but you cannot yet deploy one into a group conversation. The privacy risks and liability potential remain too high, and Zhang’s instinct and protective nature of his brainbaby WeChat, which has historically been correct, is that the downside of a social-graph agent failure far outweighs the upside of being first.

This tension that is between competitive urgency and product discipline is perhaps the most honest expression of where Chinese tech actually stands with agents. Everyone feels the pressure to push out the next ‘it’ thing, but not everyone is confident they should, or at least not test it within its products that work. If it ain’t broke, don’t fix it. right?

The Harness Thesis

Even two weeks ago, I wrote about how the biggest new buzzword in the industry feels like it’s “harness.” Tencent Dowson Tong argued at the cloud summit that as mainstream model capabilities converge — and they are converging fast, with open-source models closing the gap with proprietary ones on most benchmarks — the differentiator is no longer the model. It’s the harness, echoing what we’ve been writing about. It is the engineering scaffolding of tool calling, context management, long-term memory, and workflow design that wraps around a foundation model to make it useful in production.

This is the token refinery framework — models produce crude oil, customers pay for gasoline, and the harness is the refinery. The harness is the refinery and value migrates downstream to whoever can transform raw intelligence into something usable, trusted, and repeatable. Money moves downstream; it always does.

Tong claims that Tencent is building the best refinery. The four pillars of their Agent Core are Context Engineering, Agent Orchestration, Agent Memory, and Agent Ecosystem, which can connect to CodeBuddy on the developer side and WorkBuddy on the office-worker side, with their ADP development platform providing lifecycle management. Inside Tencent, 2,000 employees have been using WorkBuddy since January across document editing, data cleaning, and business analysis. Tasks that took hours now take twenty to thirty minutes. Token consumption keeps climbing. Tong joked that soon, employees won’t just ask their bosses for raises, but they’ll negotiate for higher monthly token budgets.

The thesis is intellectually provocative, but it raises a question that neither summit attempted to answer: if every cloud vendor is building a harness and trying to build the best refinery. The harnesses converge on the same architecture, then what exactly differentiates any of them?

Same same but different

This brings me to my final point today. I felt like everything was getting to a point where it was all a bit same-same.

For two companies with such different DNA, business models, and business units, the AI monetization logic at both events felt eerily similar.

Give enterprises agent-building tools. Those tools produce agents. Agents consume tokens. Token consumption drives cloud usage. Cloud usage drives revenue. Layer on security governance and IM distribution to make the agent sticky, and the flywheel spins. Heck, Alibaba has a new business unit helmed by its CEO Eddie Wu called the Token Hub. Meanwhile, you see below on the big screen - those big words Tencent Token Hub.

Is the thinking almost too simplistic? Integrate claw = agent tools = more agents =more token consumption = more cloud revenue.

Li Qiang was unusually candid about the fragility of this logic. He compared tokens to fuel, and that highlights the risk. As switching costs for token buyers approach zero, users switch to whoever is cheaper. I don’t care where I get my gas. I check which franchise offers me the most coupons. And he said, plainly, that Tencent would not chase token volume as a top KPI - a smart message to staff, but how are you going to make money then?

The candor laid bare the tension at the center of both companies’ strategies. If tokens aren’t the margin and models aren’t the margin, the entire enterprise bet depends on the harness layer generating enough proprietary value to justify billions in R&D. Tencent disclosed RMB 18 billion in AI spending for 2025, with further growth in 2026. Alibaba’s capex trajectory tells a similar story. But neither has yet demonstrated that agent orchestration produces sustainably differentiated revenue. For now, they are building refineries on faith.

However, if you must, underneath the convergence, there are two real strategic differences worth separating from the noise.

Alibaba proudly claimed that it is now the number one cloud choice for AI usage in China, and as the company expected, public cloud demand has grown as AI applications continue to find new use cases and penetrate deeper into vertical cases. Their most distinctive thesis: embed AI inference directly within the database. Enterprise data already lives in databases — a lifecycle from cold archival to warm training sets to hot production. By running the model inside the database — zero latency, data sovereignty, data gravity — the database becomes the gravitational center of the enterprise AI stack. And gravity, once established, is extremely difficult to escape. Migrating requires moving not just data, but the trained agent behaviors, skill libraries, and memory states entangled with it.

Their new ApsaraDB Claw products give this thesis a commercial form. Tbh, part of it was too technical for me. But at a high level, DataClaw, PolarClaw, and ADBClaw are agents that live where the data lives, which covers DBA, development, analytics, and operations around the clock.

In comparison, Tencent is betting on workflow embeddedness. Where Alibaba locks you in through your data, Tencent intends to lock you in through your habits. It’s probably not an exaggeration to say that WeChat is the default platform for any social interaction in China, whether enterprise use or personal use; thus, the vision, as Wu Yunsheng put it, is that AI will become like “water and electricity” as utilities, functioning in the background and turned on with a faucet, not consciously counted.

[Translation for below: Tencent strategic pyramid: Energy —> Compute —> Models —> Agents —> Applications]

So where does value accrue now?

Schumpeter observed nearly a century ago that the economic surplus from a new technology rarely accrues to its inventors. It accrues to the people who figure out how to use it. The cloud vendors are playing the role of the inventor here, building general-purpose agent infrastructure at enormous expense. But the most interesting things I saw at either summit weren’t platforms. They were the customers.

Bilibili built an AI-powered video commercialization pipeline: Qwen-Audio and Qwen-VL analyze KOL product reviews, extract brand mentions, and match them to advertiser catalogs. Bilibili's service can be very valuable because of its content taxonomy and its relationships with advertisers, which no cloud vendor can replicate.

On the Tencent side, the case study that was shared was about the Chinese Marriott, Huazhu Hotels. The hotel chain manager deployed ‘华小二’, ‘Little Hua', an AI concierge covering 152 scenarios across 38 workflow modules. For example, when a guest says, “Send me some snacks.” Within five seconds, an agent will alert the manager to dispatch the nearest staff member robot to the room. It’s now already live in over 5,000 hotels, with 94% daily usage in pilots, and is scaling to all 12,000 properties. This is again because Huazhu knows how customers behave and how to manage hotel workflows.

The model is merely an intelligence input at this point, but the industry context is the moat. If Schumpeter is right again, the billions flowing into harness layers may end up submerging the margins of the infra providers.

The bigger theme I think we can recognize is that in the cloud era, raw infrastructure has been commoditized, and the most durable value has been captured by software companies that have turned it into systems of work. In mobile, the platforms mattered enormously, but so did the companies that transformed distribution into entirely new behaviors.

In the agent era, both Alibaba and Tencent are maybe realizing that $ will not be in infra, but in the markup lives in the harness layer — the orchestration, memory, governance, and distribution scaffolding around the model. They may be right. But if the harnesses all look the same, then wouldn’t the differentiator still be in the foundation models and capabilities?

Anyway, the bet is that the value is moving one layer further out, to the companies that weave those harnesses into specific industries, specific workflows, and specific institutional contexts, no platform vendor can replicate, which, incidentally, is why vibecoding won't replace industry-specific software — context isn't something you can prompt-engineer.

I left Shanghai thinking about something Li Qiang said almost in passing. He called AI “the greatest force of certainty in an uncertain era.” That’s a good line for a keynote. But certainty about what, exactly? Will those agents exist? Sure. That those tokens be consumed? Obviously. Will the cloud vendors building all this infrastructure be the ones who profit? That part is less certain than anyone on those stages wanted to admit.

The WeChat question might be the best lens for the whole thing. Everyone knows agents need to be everywhere. Not everyone knows if letting them loose is wise. And the companies that figure out where to draw that line — in product design, in guardrails, in knowing when not to ship — might matter more than the ones that built the fastest refinery.

But just yesterday, ByteDance seemed to have stealthily rolled out the function across its video editing tools (in non-US markets). And making it feel more and more like a reality where AI-generated video will no longer just be slop, but become a vertical of its own, potentially replacing some aspects of acting as we know it and moving toward industrial use. People raved about the capabilities here and here.

The thing is, I think Hollywood is still looking at the wrong battlefield, at least for now. The first meaningful commercial win for Chinese AI video is unlikely to be the two-hour blockbuster that demands sophisticated acting, screenwriting, costumes, lighting, motions, character arc, and cinematography. It’s the 90-second microdrama that is made for the phone, vertical, bingeable, optimized not for Emmys but for thumb-stopping cliffhangers.

While I'm not a fan of these microdramas myself, I did give them a try to better understand the product. I have to say, the business model to integrate AI is quite smart.

Think of it as Content Factory vs. Content Studio. A conventional studio will push forward and hire a massive crew to create a masterpiece that often incites intellectual debate, is conscious of social and political context, and evokes emotions and fosters universal connectivity; that’s what usually makes these big-screen movies a hit. An AI-film factory, on the other hand, tests, iterates, acquires traffic, and monetizes retention. Hollywood is worried about the studio model being disrupted, but the real disruption is the factory, which, in many ways, looks more like streaming services and influencer content, and now AI is what makes the factory run.

As I have written before, the interesting part of AI is increasingly not who has the flashiest frontier model, but who turns that capability into an indispensable application with a repeatable business model. In creative AI specifically, the question is not “how many people tried the app,” but whether the product solves a commercial workflow with enough ROI to justify the cost of inference and enough urgency for people to pay. Microdrama is a near-perfect case study of that logic.

And we’ve seen that over the last year, there’s been growing interest in microdramas from China. MIT Tech Review wrote about it last year, WSJ wrote about it recently, and more people are starting to see the connection to AI-generated video. So let me break it down.

Share

What Microdramas Actually Are

To be clear, Microdramas are not just shorter TV shows. They are smartphone-native, retention-optimized, and monetized more like gaming than like traditional streaming. Episodes run one to a few minutes. They often end each episode on betrayal, revenge, fake marriages, secret heirs, werewolves, billionaires, and miraculous recoveries.

Just as you’re getting super worked up, the app will pause and ask you to do what is more similar to what mobile gaming has trained people to do for years: pay to unlock the next episode, wait, or watch an ad. To end on that hook, to make people yearn to find out what’s next (even though usually it’s quite obvious) is exactly the point, and people are willing to pay to find out what happens next.

The product is not pretending to be prestige storytelling, nor is it even attempting to be jaw-dropping cinematography; it is shamelessly positioned as narrative dopamine.

Reuters reported that one of their reporters visited a set where an entire season was shot in six days. A shabby room or studio that can be staged for essentially all the scenes, and some (no offense) D-listers or amateur or student actors. That’s all it takes. For different markets, studios may hire different ethnic actors to play out the same screenplay. Budgets typically range from about US$28,000 to $280,000, which is like a rounding error next to a Netflix original.

The content is cheap enough to be disposable, but sticky enough that one hit can pay for many misses. At its core, it is not pretending to be high-brow art, but rather visual dopamine hits created with SEO in mind and performance marketing attached.

Subscribe now

The Scale Is No Longer Niche

Micro-drama as a sector is not so micro anymore, it is already a real consumer-spending category and proven to no longer be an experiment. And to my surprise, the North American market is actually the main profit pool for most of these companies.

Sensor Tower says micro-drama apps ReelShort and DramaBox generated about $130 million and $120 million of in-app revenue in Q1 2025 alone, bringing their cumulative global in-app revenue to roughly $490 million and $450 million, respectively, by March 2025. Of which, the U.S. market now contributes 60–70% of Chinese micro drama’s overseas short-drama revenue, and the average payment per download for ReelShort and DramaBox in the U.S. is about six times that of other markets.

Why AI Fits This Format First

Then there is the obvious shift in embracing AI tools for creating microdramas. (plus investors in this space have told me about this trend even a few months ago)

AI video won’t replace Christopher Nolan or James Cameron; it also won’t replace Jennifer Lawrence or Leonardo DiCaprio. In many ways, what it will do is actually move value and influence toward the cream of the crop even further (yes, than it already is).

The argument I’ve heard from multiple sources is that the impact has the potential to be both good and bad, depending on whose perspective you’re getting. For actors who have not made their name yet, it might mean they no longer even get booked for these lower-budget films, the ones that they used to build up their portfolio. For directors, they may think they no longer need actors and can actually produce a film with minimal investment. For screenwriters, they may be empowered to try to make their vision a reality with fewer hurdles.

For micro-drama studios, the desire to reduce the cost and time required to produce something that is already “good enough” for a one-minute vertical episode watched half-distracted in bed will drive their rapid adoption of AI tools. Even the spokespeople at Kuaishou told me that, for AI-native film studios, microdrama does not require cinematic perfection. This vertical requires speed, emotional pull, and just enough visual coherence to keep the thumb/eyeball moving. A continuity and consistency error that would ruin a TV (mid-screen?) drama may barely matter here.

That is why AI finds product-market fit in the content factory before it reaches the content studio.

Kuaishou says Kling now serves more than 60 million creators worldwide, has generated over 600 million videos, and has partnered with more than 30,000 enterprise clients. In January, Kuaishou also said Kling’s monthly revenue had exceeded $20 million in December 2025, implying a $240 million annualized run rate. Now, Kuaishou is a publicly traded company with every incentive to showcase AI monetization to investors, so treat these numbers as directional rather than audited — but the trajectory is unmistakable.

The company spokesperson told me that even as early as last year, Kling has already been used by professionals across marketing, e-commerce, film and television, short plays, animation, and gaming. And recent partnerships with Tokyo, Hong Kong, and more film festivals, and even collaborations with film schools to empower indie or independent filmmakers, are a niche they continue to try to cultivate.

The Value Chain: IP, Tools, and the Funnel

This is also why the companies behind AI microdramas matter more than the shows themselves. The value chain has three layers, and each tells you something about where the moat actually sits.

A. The IP layer is China’s least appreciated advantage. Reuters reported that China had 575 million online literature readers in 2024 and 352 million overseas readers across more than 200 countries. China’s robust digital literary space was especially vibrant during the 2000s, where online-native writers turned web posts into bestselling books and then into blockbuster movies. That era enabled talented young writers like Hanhan, who is the face of that era of openness.

But that online ecosystem is not as simple as a database of essays; it is also a training ground for writers who have spent years optimizing for cliffhangers, reversals, binge behavior, and emotional hooks. Which is exactly the skills microdrama demands. COL Group, the Chinese digital content company that has a large stake in the studio that owns the app ReelShort and its ecosystem. COL itself claims to have more than 4.5 million authors and 5.5 million digital works. The point is not just that China has “a lot of content.” That is a hugeeee upstream reservoir of serialized IP and retention-trained writers ready to write microdrama scripts essentially. Thus, there have been rumors that companies like COL continue to buy out vast amounts of these IPs, in a way, playing the VC game. If they ship out enough series, one will become big, and that one will be enough to justify the production costs of the others.

B. The tool layer is where ByteDance and Kuaishou are competing the most fiercely. ByteDance’s edge has long been recommendation infrastructure, product iteration speed, and the ability to funnel users across a wider ecosystem. In many ways, ByteDance’s release of Dreamina Seedance 2.0 is probably going to give Kling’s give a run for its money. Both have the advantage of being able to train on vast amounts of video data, but ByteDance’s may be higher in quality and diversity. The company’s editing tools have also taught its AI taste, preference for subtle and nuanced ones.

For ByteDance, microdrama is not merely a customer vertical; it is a proof point for why multimodal AI belongs inside an existing distribution and monetization machine. Kuaishou’s strategy, although similar, has slight differences. Kling is already monetized through a combination of consumer subscriptions, API services, and customized solutions. and focused on relationship building directly with indie studios and institutions.

C. The app layer is where the funnel lives — and tbh, where I think the real moat forms. Operators like ReelShort, DramaBox, GoodShort, and ShortMax are not “studios.” They are mobile consumer businesses. Their real assets are distribution, recommendation, paywall design, localization, and user-acquisition know-how. In this world, AI is not just a production tool; it is a margin lever and a velocity lever. If AI lowers the cost of making scenes, dubbing material in any language of demand, generating trailers, or testing thumbnail variants, then these apps can push more content into the same funnel and learn faster which stories convert.

The Funnel Wins

So here is the strategic core, and this is the part I want to highlight: AI lowers production costs, but it does not automatically make content businesses suddenly lucrative. In fact, once production becomes cheaper for everyone, production itself becomes less scarce, and in many ways, competition becomes fiercer. The moat shifts.

The moat is in distribution, and that is what ReelShorts and DramaBox have been investing heavily in. Even though the microdrama business is not lucrative or even loss-making right now. The thinking is to first get as many users to download their apps as possible, and if they own the distribution, then once the cost of production goes down drastically, the opportunity to monetize at scale is theirs to take.

So, looking at the forces, I’d rank distribution and IP ownership as the two that matter most. Distribution because it determines who captures the economics of attention — and in a world of infinite content, attention is the only scarce thing left. IP because it determines who has a defensible supply of stories that are pre-optimized for the format, and that supply is not easily replicated. Payment conversion, localization, and recommendation matter too, but they’re features of the funnel, not independent moats. And that also proves my point that it is going for the streaming platforms’ lunch first, not Hollywood, nor even YouTube/creator content first.

The companies that own both the stories and the funnel — like the COL Group / Crazy Maple / ReelShort nexus — are structurally advantaged. You see what I mean? It’s less about who has the best AI model and more about who controls the full loop.

This is why I’d frame AI microdrama less as a “film” story and more as an application story. It sits at the intersection of content, gaming, and performance marketing. The model provider supplies the tool. The IP company supplies the story. The app operator supplies the funnel. The app store takes its cut. The ad platforms sell the traffic.

The highbrow studios and the traditional blockbuster model will survive for prestige, for cultural events, for the kind of storytelling that requires genuine human nuance and craft. But the content factory is where volume, velocity, and AI economics converge. And that factory is already running.

Risks and Quiet Winners

Disney and Paramount have already challenged ByteDance over Seedance. Domestically, China’s regulator is also tightening control over the category. Reuters reported that the country now requires licenses to broadcast microdramas, part of a broader regulatory push after earlier campaigns removed tens of thousands of shows and roughly 1.4 million episodes for lowbrow, violent, or vulgar content, or material that promotes ‘unhealthy behavior’. Regulation will probably further consolidate the market, which favors incumbents with the resources to comply and disadvantages smaller operators, thus making the first movers like ReelShorts and DramaBox in a much better position.

And there is a structural irony worth flagging: if AI sharply lowers production costs for everyone, the market may be flooded with even more supply, which could compress producers' margins and shift even more of the economics toward traffic owners and app stores. Apple and Google, who extract 15–30% of in-app purchases, are quiet winners in the microdrama economy. They will be collecting rent on hundreds of millions of dollars in transactions without producing a single frame of content.

As of now, it looks like these Chinese micro-drama studios are building something closer to an AI-native content factory: web-novel IP upstream, retention-trained, serialized writers, short-video recommendation systems, mobile paywalls, performance marketing, and, increasingly, capable video models that power production.

It is actually a consumer-application story in which production costs collapse and the moat shifts from creation to distribution. It happened in music, where Spotify and playlists matter more than studios. It happened in publishing, where Amazon’s algorithm became the gatekeeper. And it is now happening in short films.

Relevant reads:

• 大公报：国产微短剧海外受捧 下载量超YouTube

• AI Proem: Kuaishou’s Strategy: Riding the Rise of Short Mini Dramas + Industry Expertise Over AGI Chase

• AI Proem: The Ghibli Hype: Next Frontier is in Video & Introducing China’s Leading AI Video App, Kling

• AI Proem: ByteDance’s Second Act in AI + OEM Partnership + DeepSeek V3.2

• 人民网：竖屏新势力，中国微短剧如何“踏浪出海”