The history of global finance is punctuated by periods of frenzy, pivotal eras where a new technology upends the established order, redrawing the maps of wealth and power. From the railroads to the dawn of the Internet, every industrial revolution has birthed its visionaries, its prophets, and, inevitably, its speculative bubbles. Today, the world holds its breath as Artificial Intelligence (AI) emerges. This sector, which promises to redefine the human condition, is capturing everyone’s attention. But beyond the philosophical and technological debates, AI is primarily draining massive capital from an army of investors eager to be part of what could be the greatest wealth creation of the 21st century.

At the heart of this incandescent dynamic, where money flows at a pace some call dizzying, new figureheads are emerging. Among them, one name resonates with particular intensity in the corridors of Silicon Valley and on the trading floors of Wall Street: Leopold Aschenbrenner.

At only 24 years old, this young prodigy has already been saddled with a nickname that speaks volumes about the aura surrounding him: the “Nostradamus of AI“. At the helm of the hedge fund Situational Awareness, he has transformed a substantial initial capital into a colossal financial empire of $20 billion in less than two years. But who is Leopold Aschenbrenner really? How did he manage to rally the brightest minds and the deepest pockets of the tech world around him? Are we witnessing the absolute genius of our time, or the most glaring symptom of a speculative bubble ready to burst?

Dive into the heart of an unprecedented financial, technological, and sociological phenomenon.

Subscribe now

Forget SpaceX’s Valuation: The $1 Trillion Machine That Unlocks the Quadrillion-Dollar Cosmos.

Humanity scraped a finite globe for millennia. Today, the miracle of reusable rockets is unlocking the universe’s ultimate post-scarcity economy.

I. The Rise of the Prophet: From Anonymity to Financial Divination

In an ecosystem where obsolescence is the rule and technologies evolve at an exponential rate, the ability to anticipate trends is not just a competitive advantage; it is a matter of survival. Some humans, endowed with an uncanny prescience, still manage to stand out against algorithms in this frantic development. These are the clairvoyant investors, minds capable of identifying gold nuggets in a river of digital mud, achieving what the Wall Street Journal calls “parabolic performances“.

Leopold Aschenbrenner embodies this new aristocracy of clairvoyance. At 24, an age when most young graduates are still finding their way or laboriously starting at the bottom of the ladder in major investment banks, Aschenbrenner chose another path: that of the chief architect of tomorrow’s economy.

“He doesn’t just read the markets; he seems to read the source code of our technological future. Leopold is not a simple analyst; he is the Nostradamus of AI.”

— Tim Ferriss, podcaster and bestselling author.

It is under this moniker, popularized by Tim Ferriss, that the Aschenbrenner legend was forged. But the title of “Nostradamus” is not merely a figure of speech. It illustrates a deep, almost mystical belief among his followers: the idea that this young man possesses such an intimate understanding of large language models (LLMs), artificial general intelligence (AGI), and the underlying hardware infrastructure, that he can predict the tectonic shifts of the economy before the tremors are even registered by traditional markets.

II. Situational Awareness: More Than a Fund, an Investment Philosophy

To understand Leopold Aschenbrenner’s meteoric rise, one must look at the war machine he created: Situational Awareness. This hedge fund specializing in artificial intelligence was not launched as a mere investment vehicle, but as a genuine “AI think tank”.

Dizzying Figures

Financial history is full of overnight successes, but the metrics displayed by Situational Awareness defy comprehension:

• Creation: Less than two years ago.

• Initial Investment: A few hundred million dollars (an already colossal sum for such a young founder, testifying to the blind trust of his early backers).

• Current Assets Under Management (AUM): $20 billion.

• Growth since inception: Over 1,000%.

• Growth at the start of this year alone: 270%.

These figures, revealed by sources close to the company, are not the result of chance or fleeting luck. They are the outcome of surgical strategic choices in the equity markets—choices presented as deeply visionary. The fund does not merely follow the trend; it anticipates it, shapes it, and profits from it with ruthless efficiency.

The Concept of “Situational Awareness”

The fund’s name, Situational Awareness, was not chosen lightly. In military and security jargon, the term refers to the ability to perceive elements in the environment, understand their meaning, and project their future status. Applied to AI, this means understanding not only the underlying technology (computing power, semiconductors, algorithms) but also its geopolitical, energetic, and societal implications.

Aschenbrenner and his team operate on the premise that the market is inefficient in its understanding of AI. While the majority of investors rush toward the obvious stocks, Situational Awareness maps the entire value chain, identifying future bottlenecks and the silent winners of this technological arms race.

III. The Masterstroke: The Anthropic Gamble and the Approach to AGI

If a fund posts a 1,000% gain, there is necessarily a central investment thesis—a bold bet that has paid off beyond all expectations. For Situational Awareness, this “home run” has a name: Anthropic.

In February 2025, when the market was still polarized around the dominance of certain legacy players, and doubts remained about the long-term viability of multiple competing models, Aschenbrenner made a radical decision. He directed a massive portion of his capital toward Anthropic, the startup behind the generative AI model Claude.

Why Anthropic?

1. Safety and Alignment: Anthropic has always focused its development on constitutional AI and safety, an approach that resonates with Aschenbrenner’s philosophy of risk anticipation.

2. Architecture and Efficiency: Claude’s successive iterations have proven an ability to rival, or even surpass, the competition while keeping computing costs in check.

3. The Initial Public Offering (IPO): The stake taken in February 2025 proved all the more strategic as Anthropic’s IPO is announced as imminent.

Today, this stake represents nearly one-fifth (20%) of the fund’s $20 billion in assets. This level of concentration on a single investment line is atypical in classic portfolio management, where diversification is elevated to an absolute dogma. But for Aschenbrenner, excessive diversification is the refuge of those who don’t know what they are doing. His bet on Anthropic demonstrates an unwavering conviction in this startup’s victory in the race toward Artificial General Intelligence (AGI).

IV. Knighted by the Elite: An Institutionalized Cult of Personality

A 24-year-old fund manager, no matter how brilliant, does not manage $20 billion without solid backing. One of the most fascinating facets of the Leopold Aschenbrenner story is the nature of his network. He has managed to unite a heterogeneous but overwhelmingly powerful coalition around him, ranging from the “Silicon Valley Mafia” to the most feared sharks on Wall Street.

The Council of Tech Elders

Situational Awareness was designed from the outset as a think tank co-founded with Carl Shulman. Shulman is no stranger: a former employee of the omnipresent and controversial Peter Thiel’s hedge fund, he brings intellectual rigor, an analytical methodology, and an invaluable address book linked to the Founders Fund ecosystem and libertarian tech alt-right to the fund.

But the support doesn’t stop there. The fund’s foundations rest on the personal capital and mentorship of mythical Tech figures:

• Patrick and John Collison: The formidable Irish brothers who founded Stripe, the global online payment giant. Their backing brings instant credibility to Aschenbrenner’s project in terms of infrastructure and the digital economy.

• Daniel Gross and Nat Friedman: Two central figures in Silicon Valley, currently deeply involved in the development and integration of AI for the behemoth Meta. Their presence in the capital and on the board of Situational Awareness gives the fund an unparalleled view of Big Tech’s strategies regarding Open Source and global model deployment.

The Quantitative Shadow: The Case of Jane Street

If Silicon Valley’s support validates the technological thesis, Wall Street’s support validates the financial mechanics. And in this realm, the stake taken by Jane Street sounds like a thunderclap.

Jane Street is one of the most mysterious and profitable quantitative trading players in the world. Handling trillions of dollars, this firm is famous for its impenetrable algorithms, its ultra-elitist recruitment (often former chess champions or top-flight mathematicians), and, more recently, for its involvement in several dubious proceedings related to the highly volatile and unregulated cryptocurrency market.

“Jane Street’s choice to invest in Situational Awareness is extremely significant. This is a company that almost never delegates its capital to external fund managers. If they do it for Aschenbrenner, it’s because they see in him an informational and analytical edge that they cannot replicate internally.”

— Wall Street Journal Analysis.

This institutional knighting transformed Aschenbrenner. From a mere “Tech genius,” he became a financial strike force to be reckoned with, marrying a fundamental understanding of technology with the brute force of algorithmic and quantitative trading.

The Infinite Engine and the Absolute Vault: How AI Tokens Will Fuel the Bitcoin Standard.

The production of artificial intelligence is about to become the biggest commodity on Earth—and its survival depends on the only asset that cannot be inflated.

V. The Butterfly Effect: How One Man Shakes the Markets

When an individual concentrates so much attention, capital, and praise, a mechanical phenomenon occurs in the financial markets: the slipstream effect, or the Aschenbrenner Effect. Like Elon Musk’s tweets or Warren Buffett’s annual letters, the decisions made by Situational Awareness now have the power to dictate short-term macroeconomic trends.

The Textbook Case: T1 Energy

The most glaring proof of this almost divine power over stock prices took place in mid-May. AI needs data, chips (GPUs), but above all, it needs energy. The massive data centers that train AI models consume astronomical amounts of electricity. In his quest for “situational awareness,” Aschenbrenner understood before many others that the real bottleneck of AI would not be software, but energy supply.

When the announcement of Situational Awareness‘s stake in T1 Energy, a company specializing in the manufacturing of next-generation solar modules, was disclosed, the markets caught fire.

• The Result: T1 Energy’s stock experienced a blistering surge of 23% in less than 24 hours.

This was no longer a simple value investment; it was a gold rush triggered by a signal from the person considered the most informed on the planet. The company, previously relegated to the status of a classic “green stock,” was instantly reclassified by the market as “critical infrastructure for AI.”

The Democratization of Speculation: The Autopilot Index

The cult of personality surrounding Aschenbrenner has transcended the closed circles of institutional investors to reach the general public. Retail investors, always searching for the miracle recipe to beat the market, have found their new guru in him.

This phenomenon was catalyzed by the trading app Autopilot, very popular among young investors. The app recently deployed an automated index that allows any user to exactly replicate the positions disclosed by Situational Awareness during its quarterly status reports (the famous SEC 13F forms).

By linking their brokerage account to this index, retail investors virtually delegate the management of their savings to the “Nostradamus of AI.” This mechanic creates extremely powerful feedback loops: Aschenbrenner’s fund buys a mid-cap stock, the information becomes public, the Autopilot index automatically triggers thousands of buy orders from individuals, mechanically exploding the stock price, which retroactively validates the prescience of the initial investment. A self-fulfilling prophecy of terrifying efficiency.

VI. The Icarus Syndrome: Visionary Genius or Hero of a Fleeting Bubble?

Faced with such successes, such speed of execution, and valuations that defy historical models, a burning question is on the lips of regulators, cautious economists, and financial historians: Are we witnessing the emergence of a generational genius or the peak of a monumental speculative bubble?

The financial press, while marveling at Aschenbrenner’s performances, does not hesitate to draw disturbing historical parallels. Some are already comparing the founder of Situational Awareness to emblematic figures of past bubbles.

The Shadow of Ryan Jacob and the Dot-Com Bubble

In the late 1990s, during the Dot-com bubble, one name made the front pages of all the magazines: Ryan Jacob. This young fund manager, running the Internet Fund, posted astronomical returns by blindly betting on any company that added “.com” to its name. Jacob was lauded as the prodigy who understood the new paradigm. When the bubble burst in March 2000, his fund lost a catastrophic portion of its value, destroying the savings of thousands of investors who were convinced that “this time, it’s different.”

The Recent Parallel: Cathie Wood and the Covid Era

More recently, the comparison with Cathie Wood, founder of ARK Invest, is inevitable. During the COVID-19 pandemic in 2020 and 2021, Wood was elevated to the status of financial oracle. Her ETF funds, heavily invested in “disruptive technology” (Tesla, Zoom, Teladoc, cryptocurrencies), recorded extraordinary performances. She, too, claimed to understand the future better than traditional markets. However, when interest rates began to rise in 2022 and the macroeconomic climate darkened, ARK’s flagship fund saw its valuation collapse, demonstrating that the line between “visionary” and “lucky speculator in a bull market” is extremely thin.

Leopold Aschenbrenner: Is History Stuttering?

Is AI a bubble? If we stick to the strict definition (a total disconnect between stock market valuations and the fundamental value generated by companies), the debate is open. The sums injected into AI infrastructure (data centers, chip manufacturing, model training) amount to hundreds of billions. For these investments to be profitable in the long run, AI will have to generate trillions of dollars in real economic value.

What perhaps distinguishes Aschenbrenner from Jacob or Wood is the nature of the technology. AI is not just a new distribution channel (like the Internet) or a new way of working (like Zoom); it is a universal technology, theoretically capable of automating human cognitive work. If Artificial General Intelligence (AGI) is achieved within the current decade, today’s valuations, even the wildest ones, might retrospectively seem undervalued.

However, the risk matches the reward. With $20 billion under management, a heavy concentration on assets like Anthropic, and a base of retail investors tracking its every move, the Situational Awareness fund has become a systemic pillar of the AI market. If Aschenbrenner’s thesis proves wrong, if AI progress hits a wall of unforeseen complexity, or if global regulation abruptly brakes innovation, the collapse of this fund could be the catalyst for the greatest stock market correction of the 2020s.

The $4,000 Illusion: Why Gold’s Technical Breakdown is a Fundamental Trap.

Ignore the Paper Panic: The Seven Unsolvable Macroeconomic Crises Exposing the Fiat Debt Spiral and Cementing the Case for Hard Money.

VII. The Anatomy of Knowledge: What the “Nostradamus” Sees

What does Leopold Aschenbrenner know that we don’t? To understand his success, one must dive into the literature and essays he shares within his think tank. His vision is built on three fundamental pillars that the traditional market still struggles to price correctly.

1. Compute Scale

Aschenbrenner firmly believes in the “Scaling Laws” of AI. According to this principle, the more computing power and data you pump into a model, the smarter it gets, with no apparent ceiling in the short term. This conviction justifies his indirect investments in everything that supports this hardware expansion (energy, cooling, semiconductors).

2. The Geopolitical Arms Race

Unlike investors who view AI as a simple productivity tool (B2B SaaS), Aschenbrenner analyzes it through a military and national security lens. He anticipates that governments (led by the US and China) will inevitably co-opt AGI projects. Investing in companies like Anthropic, which have strict safety protocols potentially compatible with Department of Defense requirements, is a strategic bet on the next geopolitical era.

3. The Human “Strike Force” Bottleneck

While many fear job losses, Aschenbrenner bets on the exponential value of the rare engineers capable of building AGI. The backing of figures like Daniel Gross or the Collison brothers isn’t just about money; it’s a recruiting network to ensure the companies he invests in attract the $10-million-a-year brains.

Final Thoughts: The Prophecy Put to the Test

Leopold Aschenbrenner is not just an isolated financial phenomenon; he is the perfect reflection of our era. An era characterized by existential anxiety over the speed of technological change, coupled with a lust for profit, exacerbated by years of accommodative monetary policies.

At 24, with $20 billion under his responsibility and a 1,000% growth that leaves Wall Street stunned, he is the undisputed king of the AI chessboard. The support of opaque and powerful institutions like Jane Street, combined with popular fervor fueled by tools like Autopilot, grants him outsized influence.

The “Nostradamus of AI” read in the great algorithmic book that our future will be dominated by omniscient digital entities. He placed the pawns of Situational Awareness on strategic squares: energy infrastructure (T1 Energy), the brains behind safety and brute power (Anthropic), and the global Tech ecosystem.

If he is right, his name will join those of Rockefeller, Vanderbilt, or Jobs in the pantheon of modern capitalism’s builders. If he is wrong, if AI turns out to be an energy sinkhole with no AGI at the end of the tunnel, history will remember him as the angelic face of the biggest tech bubble of the 21st century.

In either camp, one thing is certain: the world is watching Leopold Aschenbrenner’s crystal ball with fascination, hoping to read the reflection of a bright future rather than that of the precipice. Time, as always in the markets, will be the sole and ultimate judge of the prophecy.

Share AI & Quantum Computing Newsletter

The Sovereign State of Silicon: How NVIDIA’s $40 Billion Gamble is Bankrolling the AI Economy.

As the tech industry bleeds cash in the race for AGI, the $5 trillion chipmaker is single-handedly underwriting the entire supply chain—but can one company buy an ecosystem’s profitability?

The $5.4 Billion Fugazi: How Michael Burry Uncovered the Shadow Machine Funding the AI Boom with Retiree Money.

Inside the 16x-leveraged, offshore shell game hiding billions in Silicon Valley debt—and quietly betting America’s retirement on the lifespan of a computer chip.

In 1994, the Internet was in its infancy. Secure communication, the bedrock of modern digital commerce, relied on cryptographic protocols like RSA. These protocols were—and still are—predicated on a simple mathematical asymmetry: multiplying two large prime numbers together is computationally trivial, but factoring the resulting massive integer back into its constituent primes is practically impossible for classical computers.

Then came mathematician Peter Shor.

Shor developed a quantum algorithm that proved a sufficiently large, stable quantum computer could factor these massive integers exponentially faster than any classical machine. To understand the panic and excitement this caused, one must look at the mathematical complexity.

For a classical computer using the most efficient known algorithm (the General Number Field Sieve), the time complexity to factor an integer N scales sub-exponentially:

Shor demonstrated that a quantum computer could perform the same task in polynomial time:

This stark mathematical reality meant that a functioning quantum computer could break the encryption securing the global financial system, classified military communications, and private data infrastructure. It was the ultimate motivator. Funding poured into quantum physics departments worldwide.

Thirty-two years later, in 2026, no one has built a machine capable of executing Shor’s algorithm at a scale that threatens global encryption. The problem was never the math; the problem was the physical universe.

Subscribe now

The Decade That Never Ended

Building a quantum computer required thousands, perhaps millions, of qubits operating with error rates that no one in the 1990s or 2000s knew how to achieve. Because of this monumental physics hurdle, the “decade away” framing became a self-reinforcing prophecy.

A comprehensive report by the National Academies of Sciences, Engineering, and Medicine perfectly encapsulated this stagnation. They noted that quantum computing appeared on Gartner’s prestigious list of emerging technologies an astonishing 11 times between 2000 and 2017. In every single appearance, it was placed at the earliest stage of the “hype cycle” and categorized as being more than ten years away from commercial viability.

When the National Academies published their own definitive assessment in 2018, the conclusions offered cold comfort to optimists. The report stated:

“Significant technical and financial issues remain towards building a large, fault-tolerant quantum computer, and one is unlikely to be built within the coming decade.”

Boston Consulting Group (BCG) published a parallel analysis that same year. Their conclusion was remarkably similar, noting that the full impact of quantum computing was “probably more than a decade away,” even as they acknowledged that a closer disruption was gathering force in the near term.

This pattern of prediction was not dishonest, nor was it a failure of imagination. The problems were brutally real, rooted in the fundamental laws of quantum mechanics.

The Physics of Fragility

To understand why the timeline stalled, one must understand the qubit. Unlike a classical bit, which is definitely a 0 or a 1, a qubit exists in a superposition of states, mathematically represented as:

Here, the complex coefficients α and β represent probability amplitudes, constrained by the normalization condition:

This delicate state of superposition, combined with quantum entanglement (where the state of one qubit is intrinsically linked to another), gives quantum computers their immense parallel processing power.

However, this mathematical elegance comes at a severe physical cost. Qubits are incredibly fragile. Any interaction with the outside environment—a stray photon, a microscopic fluctuation in temperature, a passing cosmic ray, or even the electromagnetic noise from the control electronics themselves—causes the quantum state to collapse. This phenomenon, known as decoherence, introduces errors into the calculation.

Because noise accumulates as calculations progress, running deep algorithms like Shor’s requires Quantum Error Correction (QEC). Unlike classical error correction, which copies data (a process forbidden in quantum mechanics by the No-Cloning Theorem), QEC requires encoding a single “logical qubit” across dozens, hundreds, or even thousands of physical qubits.

For decades, the overhead required for error correction was insurmountable. Each time researchers hit a new physical milestone, the remaining distance to a useful, error-corrected machine stretched out before them like a mirage.

The Sovereign State of Silicon: How NVIDIA’s $40 Billion Gamble is Bankrolling the AI Economy.

As the tech industry bleeds cash in the race for AGI, the $5 trillion chipmaker is single-handedly underwriting the entire supply chain—but can one company buy an ecosystem’s profitability?

The Era of Hero Experiments: Three Milestones That Shifted the Conversation

The transition from a perpetual “decade away” to a concrete engineering schedule did not happen overnight. It was forced by three distinct, highly publicized milestones that gradually moved the goalposts from theoretical physics to applied engineering.

1. Google Sycamore and the Supremacy Debate (October 2019)

In October 2019, Alphabet’s Google published a landmark paper in the journal Nature. They announced that their 53-qubit processor, named Sycamore, had achieved “Quantum Supremacy”—the point at which a quantum computer can perform a task impossible for a classical computer in any reasonable timeframe.

Google claimed that Sycamore completed a highly specific, complex mathematical task (random circuit sampling) in approximately 200 seconds. They estimated that the world’s most powerful classical supercomputer at the time would require 10,000 years to simulate the same calculation.

The announcement sent shockwaves through the tech community, but it also sparked immediate commercial and scientific friction. IBM, a primary rival in the quantum space, fired back aggressively. IBM published a counter-paper arguing that Google’s claim was drastically overstated. IBM demonstrated that with mathematical optimizations and massive storage arrays, a classical supercomputer like Oak Ridge’s Summit could perform the calculation in just 2.5 days, not 10,000 years.

While the academic debate over the definition of “supremacy” dominated headlines, the true significance of the Sycamore processor was subtly overlooked by the mainstream: Google had demonstrated an unprecedented level of precise, programmable control over 53 qubits operating in tandem. It was a staggering feat of cryogenic engineering and microwave control.

2. Google Willow and the Error Correction Rubicon (December 2024)

If Sycamore was a proof of concept, Google’s announcement five years later, in December 2024, was a foundational earthquake.

Google unveiled Willow, a 105-qubit chip. While it performed a benchmark computation in under five minutes that would theoretically take a classical supercomputer 10 septillion years, the computation itself was a secondary headline. The true breakthrough lay in error correction.

For thirty years, quantum physicists had chased a mathematical concept known as the “fault-tolerance threshold.” In simple terms, the threshold theorem states that if the physical error rate p of your qubits falls below a certain critical threshold p_th, you can arbitrarily suppress logical errors by simply scaling up the system.

Mathematically:

Before Willow, quantum computing was trapped above this threshold. Adding more physical qubits to a system introduced more noise, making the system less reliable. It was a devastating paradox: to build a powerful computer, you needed more qubits, but adding more qubits ruined the computer.

Willow cracked this 30-year challenge. Google demonstrated that as they scaled up the surface code (the specific error-correction architecture they utilize), the errors reduced exponentially. For the first time in human history, adding more qubits made a quantum system better, not worse. By crossing “below the threshold,” Willow transformed quantum error correction from a theoretical physics paper into a viable hardware roadmap.

3. Microsoft Majorana 1 and the Form Factor Revolution (February 2025)

Just two months after Google’s Willow announcement, Microsoft altered the physical dimensions of the race. In February 2025, they introduced Majorana 1.

Microsoft had long been the outlier in the quantum industry. While Google and IBM built superconducting qubits that required massive, chandelier-like dilution refrigerators, Microsoft placed a massive, decades-long bet on a theoretical concept called the “topological qubit.”

Topological quantum computing relies on elusive quasi-particles called non-Abelian anyons, specifically Majorana zero modes. In a topological architecture, quantum information is encoded not in the state of a single particle, but in the global, braided paths these particles take around each other. Because the information is stored globally, local environmental noise (the bane of superconducting qubits) cannot easily disrupt it.

The journey was rocky. In 2018, Microsoft faced an embarrassing setback when it had to retract a major paper claiming the discovery of Majorana zero modes due to incomplete data analysis. Many in the quantum community questioned if the topological approach was a dead end.

Majorana 1 was their vindication. Microsoft placed eight topological qubits on a new architecture called the Topological Core, designed specifically to scale up to one million qubits. However, the exact physics of the chip mattered less than its form factor.

The Majorana 1 chip, complete with its qubits and essential control electronics, fits entirely in the palm of a hand. It was designed from the ground up to be deployed inside standard server racks within Microsoft’s existing Azure data centers. Microsoft promised industrial-scale problem-solving in years, not decades. While experts debated the experimental validity of the claims, given the company’s past retraction, the engineering intent was unmistakable: quantum computing was shrinking down to fit into the enterprise cloud.

The Democratization of Quantum: From the Lab to the Cloud

Hero experiments in isolated laboratories do not spawn trillion-dollar industries. Ecosystems do. The dissolution of the “decade away” myth was driven just as much by software developers as it was by quantum physicists.

This shift can be traced back to May 4, 2016. On that day, IBM made a decision that fundamentally altered the trajectory of the industry: they connected a humble, five-qubit quantum processor to the internet and made it available via the cloud to anyone who wanted to use it.

Before this moment, utilizing a quantum computer required a PhD in physics, security clearance at a national laboratory, or tenure at an elite university. IBM’s cloud access democratized the hardware overnight.

Building the Developer Army

By making the hardware accessible, IBM birthed an entire developer community. They released Qiskit, an open-source software development kit, allowing programmers to write quantum circuits using standard Python code.

A decade after that initial 2016 launch, the metrics of this ecosystem are staggering. As of 2026, the IBM quantum cloud hosts over 240,000 active users. It supports a robust network of 300 ecosystem partners, ranging from massive automotive manufacturers simulating battery chemistry to financial institutions optimizing risk portfolios.

This vast army of developers isn’t just writing theoretical code; they are characterizing noise, testing algorithms on real hardware, and finding ways to mitigate errors through software. They transformed quantum computing from an academic curiosity into a software engineering discipline.

The Concrete Roadmap

Because IBM built a cloud business around quantum, they were forced to provide their customers with something physicists historically loathed: a predictable product roadmap.

IBM’s roadmap now extends definitively through the early 2030s, prioritizing gate fidelity and logical operations over mere qubit counts.

• 2029: The Starling System. IBM plans to deliver its first truly fault-tolerant quantum computer. Starling is projected to feature 200 high-fidelity qubits capable of executing 100 million quantum gates in a single circuit without terminal decoherence.

• 2033: The Blue Jay System. A massive leap forward, Blue Jay envisions a 2,000-qubit system capable of running one billion gates.

These numbers represent the inflection point where quantum computers move from simulating small molecules to modeling complex pharmaceuticals and materials science problems that are utterly intractable today.

The Hybrid Data Center: Where Quantum Meets Classical

The most profound shift in the quantum narrative is realizing that quantum computers will not replace classical computers. You will not have a quantum smartphone, nor will you use a quantum computer to stream video or write emails.

Quantum computers are highly specialized accelerators, akin to ultra-powerful GPUs, designed to solve specific optimization, simulation, and cryptography problems. Therefore, their success relies entirely on how well they integrate with existing classical infrastructure. The milestones of 2025 and 2026 reflect this reality.

IBM and RIKEN: Quantum-Centric Supercomputing

At IBM’s Quantum Developer Conference in November 2025, the company showcased progress toward delivering true “quantum advantage” by the end of 2026. But the most significant announcement was an architectural integration across the globe in Japan.

IBM and the Japanese research institute RIKEN unveiled the first IBM Quantum System Two deployed outside the United States. Crucially, they did not place it in an isolated physics lab. They co-located it directly with RIKEN’s Fugaku supercomputer, which is consistently ranked among the most powerful classical computing systems on the planet.

This co-location was not symbolic. The two massive machines were physically and digitally networked to run a complex quantum chemistry problem in a closed loop. The Fugaku supercomputer handled the classical data processing, identified the specific computational bottlenecks, fed those specific sub-routines to the Quantum System Two, and received the quantum-processed data back in an unbroken, automated workflow.

This is the definition of quantum-centric supercomputing. It is not a demonstration of quantum physics; it is a demonstration of enterprise-grade computing architecture. It proved that quantum machines could be trusted as reliable nodes within a broader high-performance computing (HPC) environment.

NVIDIA’s Classical Backbone

As hardware companies built the quantum chips, the king of classical AI computing stepped in to handle the traffic. NVIDIA, recognized globally for the GPUs that power the artificial intelligence revolution, identified a massive bottleneck in the quantum industry: quantum computers require immense classical computing power simply to function.

Quantum error correction, specifically the process of reading error syndromes and calculating the necessary corrections in real-time before the qubits decohere, is an unimaginably intensive classical compute problem. If the classical computer decoding the errors is too slow, the quantum computer crashes.

To solve this, NVIDIA introduced Ising, a family of open-source quantum AI models. These models span key quantum workloads, starting with the automation of calibration. Tuning a quantum processor used to take teams of post-docs weeks of manual adjustments; NVIDIA’s AI models can automate and optimize this tuning dynamically.

Furthermore, NVIDIA integrated these models into its CUDA-Q software platform and introduced the NVQLink hardware interconnect. This technology physically and logically connects delicate quantum processors directly to massive arrays of NVIDIA GPUs.

The premise is brilliant in its pragmatism: let the quantum chip handle the quantum states, and let AI-driven GPUs manage the devastatingly complex classical bottlenecks of error correction and calibration. NVIDIA’s entry into the space signaled that quantum computing had matured enough to require the world’s most advanced classical orchestrators.

What Broke the Framing: The Convergence of Disciplines

For three decades, the “decade away” framing held strong because progress was entirely legible to researchers but completely opaque to enterprise infrastructure managers.

Progress was measured in physical metrics:

• T₁ relaxation times

• T₂ dephasing times

• Two-qubit gate fidelities

If you were a CIO planning a five-year infrastructure budget, these metrics were meaningless. You cannot build a business case around T₂ dephasing times. You need delivery dates, rack-space requirements, cooling specifications, and API documentation.

What ultimately disrupted the perpetual decade was not a single, miraculous result. It was a rapid, aggressive convergence of hardware physics, software engineering, and data center architecture.

Let us summarize the convergence of the mid-2020s:

None of these individual milestones from 2024 to 2026 represents a complete, fault-tolerant, million-qubit machine capable of breaking RSA encryption. However, they collectively represent the end of the physics experiment phase.

When Microsoft designs a processor specifically to fit into the standard physical footprint of an Azure server rack, when IBM physically wires a quantum mainframe to an exascale supercomputer, when NVIDIA trains AI models to babysit the fragile qubits—these are milestones in engineering, integration, and supply chain management.

The 2030 Horizon: An Engineering Schedule

In 2020, the global consulting firm McKinsey published an estimate projecting that up to 5,000 quantum computers would be operational worldwide by 2030. However, they cautioned that the sophisticated hardware and software required to tackle the world’s most complex, disruptive problems might not be fully available until 2035 or later.

To a casual observer, that timeline sounds suspiciously like the same old “decade away” trope wearing a new suit. But the qualitative nature of that wait has fundamentally changed.

The quantum computers showing up in data centers in 2026 are not delicate laboratory prototypes waiting on a miraculous, Nobel-prize-winning discovery in theoretical physics to function. They are robust processors. They are connected via fiber optics and specialized interconnects to the existing global computing infrastructure. They sit in the hum and roar of data centers, right next to the CPUs and GPUs that run the modern digital economy.

Building a fault-tolerant, million-qubit quantum computer remains one of the most complex challenges humanity has ever undertaken. It requires advancements in cryogenics, microwave cabling, material fabrication, and algorithmic efficiency. But the nature of the challenge has shifted from the realm of the unknown to the realm of the difficult.

We are no longer wondering if physics will allow us to build these machines. We are simply executing the engineering schedule required to assemble them.

The “decade away” framing may still be numerically accurate for achieving a fully fault-tolerant, cryptographically relevant quantum computer. But for the very first time since Peter Shor published his groundbreaking algorithm in 1994, that decade measures a concrete, step-by-step engineering roadmap, not a distant, perpetual hope.

Share AI & Quantum Computing Newsletter

The narrative holds the allure of grand utopias and the deceptive simplicity of technological fairy tales. According to Elon Musk and a cohort of Silicon Valley visionaries, artificial intelligence (AI) and advanced robotics will indeed destroy our jobs, but it’s not a tragedy: it’s a liberation. The solution lies in a magic formula: a high universal basic income, funded by the State. Since robots will produce goods and services in almost infinite abundance, costs will collapse. There would therefore be no inflation, only a humanity freed from toil, living off state stipends in an automated Garden of Eden.

The promise is seductive. At first glance, it even seems to rest on tangible economic logic. Yet, scratching the veneer of this blissful techno-optimism reveals a vision that is not only fragile but profoundly utopian. Worse still, it masks a dizzying societal mutation. Behind the reassuring idea of a monthly check lies a transfer of power unprecedented in human history. Let us dissect the anatomy of this illusion.

Subscribe now

1. Deflationary Pressure Facing the Wall of Reality

Let’s start by granting Musk what is true: technological innovation is intrinsically deflationary.

If AI and robotics become capable of producing the vast majority of our goods and services at a marginal cost close to zero, we do indeed enter a logic of falling prices. This is a phenomenon we already know. Successive waves of globalization and the rise of the internet have massively increased supply, slashed production costs, and exerted downward pressure on consumer prices. A flat-screen TV or a smartphone today costs, in terms of working hours needed to acquire it, a fraction of what their equivalents cost thirty years ago.

In this world where humans are replaced by machines on assembly lines and behind office desks, the idea of a universal basic income is not absurd. It is even an absolute macroeconomic necessity. Capitalism runs on consumption. If work no longer distributes wages to a critical share of the population, demand collapses. Without demand, Musk’s robots can mass-produce all they want; the merchandise stays on the docks, profits evaporate, and the entire economic machine grinds to a halt. Therefore, purchasing power must mechanically be reinjected.

However, the reasoning stops where the complexity of the real world begins. Musk’s model assumes a perfectly fluid, frictionless economy where supply adapts instantly to demand. Yet, inflation is never solely a question of aggregate supply.

Targeted Inflation: The Bottleneck of the Physical Economy

Inflation is a balance of power between supply and demand that never adjusts at the same speed. Massively injecting purchasing power via public checks stimulates demand immediately. With one click, millions of citizens have extra liquidity. But on the other side, the entire economy does not instantly become automatable.

Certain sectors of our economy will remain structurally constrained, rigid, and anchored in the limits of physics:

• Housing and land: AI cannot multiply the surface area of the Earth. Space in attractive urban areas is finite. You can distribute a generous universal income all you want; if everyone wants to live near the sea or in a dynamic metropolis, real estate prices will explode.

• Healthcare and nursing: Even assisted by AI, human presence, medical time, and empathy will remain scarce resources.

• Energy and raw materials: The robots manufacturing this abundance need copper, lithium, rare earths, and electricity. The energy transition is already showing us that these resources are subject to major bottlenecks.

• Personalized education: High-quality human tutoring will remain a premium service.

Thus, even with an abundance of digital or basic manufactured goods, we will face extreme tensions over essential goods. Inflation will not disappear; it will mutate. It will become asymmetrical and targeted, transforming housing or energy into inaccessible luxury goods, instantly devouring the famous “high universal income” we are promised.

2. Pricing Power: When AI Concentrates Wealth

The question of wealth distribution is Silicon Valley’s most spectacular blind spot. AI is not a democratic technology that will create diffuse and homogeneous wealth. By its very nature, it is hyper-centralizing.

Developing and operating the most powerful AI models requires colossal resources: astronomical computing power (compute), titanic data centers, privileged access to energy, and mountains of proprietary data. De facto, this technological revolution will, at least initially, concentrate an enormous amount of value in the hands of those who own this technological capital—meaning a handful of hegemonic companies (Big Tech and their successors) and infrastructure owners.

Picture the scene: a State that extracts wealth on one side to distribute a universal income to its population on the other, facing an economy where almost the entirety of production (supply) is controlled by three or four global monopolies.

This glaring imbalance between ultra-concentrated production and demand artificially maintained by the State creates what economists call pricing power. If you are a tech monopoly and you know the State has just deposited $2,000 into every citizen’s account, what stops you from raising the prices of your subscriptions, services, or goods? Nothing. Because the consumer is captive, and competition has been swept away by monopolistic AI.

In this scenario, the universal basic income does not serve to emancipate the citizen; it acts as a massive indirect subsidy to tech giants. The State becomes a mere intermediary that prints money to transfer it directly into the pockets of the owners of capital. Inflation reappears, not through scarcity, but through monopolistic opportunism.

The Pentagon’s Pivot: Why the US Military Now Sees Bitcoin as the Ultimate Weapon of Power Projection.

Inside the historic Congressional testimony that signals the end of uncontested fiat warfare and the dawn of a sovereign arms race for digital scarcity.

3. The Funding Dead End: Living in La-La Land

How does the State fund these monthly checks? When we approach the purely accounting aspect, the promise cracks on all sides. There are only three major macroeconomic levers to finance such a system, and none of them is painless or magical:

1. Debt and public deficit: Continuously borrowing to fund current expenditures (citizens’ income) is a model with mathematical limits. In the long run, the debt burden becomes unsustainable, leading to a loss of creditor confidence and a collapse of the currency.

2. Heavy taxation on capital: This is the fairest solution in theory. We would need to tax the profits of these tech monopolies at 80% or 90% to redistribute the wealth. Except that in a globalized world, technological capital is inherently volatile and elusive. Attempting to impose punitive taxation on supranational entities generally triggers tax wars, massive evasion, or blackmail over jobs and innovation.

3. Money creation (the printing press): This is often the implicit emergency exit for this type of discourse. The idea is that the central bank creates money to fund the State, and that the staggering increase in productivity due to AI will absorb this additional money supply, thereby canceling out inflation.

It is this last hypothesis that can be described as living in La-La Land. As long as supply perfectly tracks demand, the injection of money goes unnoticed. But as we saw in the first point, physical supply is constrained. As soon as production hits a material limit (a shortage of copper, an inability to build homes fast enough, an energy crisis), the increase in the money supply translates into currency depreciation. That means guaranteed hyperinflation on essential physical goods.

Believing that computer code will abolish the laws of thermodynamics, earthly scarcity, and classical macroeconomics is staggeringly naive for minds that are otherwise so brilliant.

4. The Societal Void: The Disappearance of the Value of Work

Let us assume, for the sake of argument, that the financial equation is solved. Let’s assume Musk is right on every front: energy becomes infinite, robots produce everything, the check arrives at the start of every month, and inflation is kept in check. Are we saved?

Absolutely not. Because the real issue is not solely economic; it is first and foremost societal and philosophical.

Since the Industrial Revolution, and perhaps since the origins of sedentary societies, our psychological balance and our social contract have rested on an extremely simple and structuring concept: work. Work is not just a means of subsistence to “earn a living.”

• It structures time: It gives rhythm to our days, weeks, and years.

• It confers status and identity: “What do you do for a living?” is the first question asked when meeting someone. Our profession defines, rightly or wrongly, a significant part of how we perceive ourselves.

• It creates social utility: Feeling useful, contributing to the community, and participating in a collective effort is a fundamental psychological need for human beings.

• It weaves social bonds: The company or the workshop is the primary place of socialization outside the family circle.

Replacing all of this with a universal basic income, even a very generous one, may solve a financial equation, but it creates a gaping societal void. What becomes of a society where 50%, 60%, or 80% of the adult population is considered economically useless, permanently excluded from the productive system?

History shows us that mass idleness rarely leads to an explosion of artistic creativity or widespread spiritual elevation. It more reliably leads to collective depression, alienation, addiction, social anomie, and, inevitably, political resentment. A population that feels useless and downgraded is profoundly unstable, ready to throw itself into the arms of the most extreme populisms to find some meaning—even a destructive one—to its existence.

The $122 Billion Mirage: Why OpenAI is Fighting for Its Survival.

Trapped by crippling compute costs and losing ground to Anthropic, Sam Altman faces an 18-month countdown to deliver a miracle—or trigger an industry-wide collapse.

5. Double Dependency: The Dawn of Techno-Feudalism

Finally, we must lift the veil on the ultimate political danger of this promise. If the system shifts toward this model, the power architecture of our democracies will be irreversibly altered.

Today, the worker has bargaining power, even if it’s weak. If they are unhappy, they can strike, paralyze the production apparatus, and demand better conditions. The capital/labor power dynamic is the engine of recent social history. But if human labor is no longer necessary for production, this power of extortion completely disappears. The citizen loses all leverage over the physical economy.

We would then enter a system of deadly double dependency:

1. Total dependency on political decision-making (the State): Your survival no longer depends on your ability to produce, but on the goodwill of a government that sets the amount of your allowance. A cut to this income for budgetary reasons, and you plunge into poverty, with no way to bounce back through work.

2. Total dependency on technological capital (the Monopolies): You depend on a few mega-corporations not only for all your material and digital consumer goods but also for the very infrastructure of society.

This is a dizzying paradigm shift. We are not simply moving from a classical economy to an economy “with more AI.” We are transitioning toward a neo-feudal system. In the old system, wealth stemmed from the exploitation of human labor. In the looming system, wealth creation is entirely detaching itself from human labor to become the exclusive prerogative of technological capital.

“In that world, the question is no longer simply ‘how to earn a living’ but ‘who owns what’.”

This sentence alone sums up the challenge of the century. As long as we own our labor power, we will have a say. If value shifts exclusively toward the ownership of algorithms, robots, and servers, and we accept a subsistence (or even comfort) income in exchange for our obsolescence, we are agreeing to become the serfs of a new system.

The real question Elon Musk should be asking is not “how much must the State distribute to compensate for AI?”

The real question is: how do we ensure that tomorrow’s citizens own a share of this technological capital?

Without democratizing the ownership of the algorithmic means of production, a universal basic income is not an emancipation. It is the price the new lords of technology are willing to pay to ensure the population remains quiet while they appropriate the world. And that is exactly what should be keeping us all awake at night.

Share AI & Quantum Computing Newsletter

In the public consciousness, the artificial intelligence arms race looks like a traditional software battle. It is a war of chatbots and benchmarks, a daily scoreboard comparing OpenAI’s latest iteration, Google’s Gemini, and Anthropic’s Claude. Every week, a breakthrough dominates the headlines, and the tech world argues over which large language model (LLM) possesses the superior reasoning capabilities, the largest context window, or the most nuanced creative output.

But if you look closely beneath the surface, a completely different war is being fought. It is not a war of code. It is a war of concrete, copper, power grids, and cold, hard capital.

Recently, Google made a strategic move that fundamentally alters the calculus of this battle. Rumors swirled that Google was attempting to acquire Anthropic, the formidable AI lab founded by former OpenAI researchers. Instead, the reality of the deal revealed something far more astute—and far more dangerous to its competitors.

Google is backing a $5 billion-plus data center project in Texas that Anthropic has already leased. Google isn’t acquiring the company. It isn’t merging with them.

Google is financing them.

This distinction is everything. By stepping out of the role of a traditional tech acquirer and assuming the roles of landlord, lender, and infrastructure provider, CEO Sundar Pichai is executing a masterclass in risk mitigation and market capture. It is a move that guarantees Google wins the AI revolution, even if its own proprietary models lose the crown.

Here is a deep dive into why this brilliantly simple setup represents a tectonic shift in the technology sector, and why owning the system is infinitely more powerful than owning the software.

Subscribe now

The Texas Mega-Project: Breaking Down the Deal

To understand the genius of this maneuver, we must first look at the sheer scale of what is being built in the Texas desert. The numbers associated with this project are staggering and speak to the physical reality of modern artificial intelligence.

• The Footprint: The data center sits on approximately 2,800 acres of land. This is not a server room; it is an industrial campus the size of a small town.

• The Power: The facility is slated to consume around 500 Megawatts (MW) of power by 2026, with plans to scale into the gigawatts. To put that into perspective, a single gigawatt is roughly the energy output of a commercial nuclear reactor, enough to power hundreds of thousands of homes.

• The Hardware: Crucially, Anthropic will be running its Claude models on Google’s proprietary Tensor Processing Units (TPUs).

• The Financing: Google is using its massive, pristine balance sheet to lower the borrowing costs for the project, making the economics viable for Anthropic while securing Google’s position at the center of the deal.

In the past, when a tech giant wanted to capture a rising star, they simply bought them. Facebook bought Instagram; Google bought YouTube. But in the era of trillion-dollar market caps and hyper-aggressive regulatory oversight, the old playbook is dead. Google recognized this and pivoted to a strategy that is as old as capitalism itself: owning the means of production.

The “Heads I Win, Tails You Lose” Setup

The most elegant aspect of this financing deal is how it completely insulates Google from the volatility of the model layer.

In the software business, owning the “best” model is highly fragile. Today’s state-of-the-art LLM is tomorrow’s open-source commodity. Training algorithms leak, engineers switch companies, and the moat surrounding any specific AI model is notoriously shallow. If Google bet everything exclusively on Gemini remaining the undisputed king of AI, it would expose itself to massive existential risk.

Instead, they have constructed a tripartite win condition.

1. If Claude Wins

Suppose Anthropic’s Claude outpaces Gemini, ChatGPT, and everyone else. It becomes the enterprise standard and the consumer favorite. In a traditional competitive landscape, this would be a disaster for Google. Under this new paradigm, it is a massive victory.

Because Anthropic is running on Google’s infrastructure, Google captures immense cloud computing revenue. Furthermore, because Anthropic is utilizing Google TPUs, Google gets paid for every single training run and every inference request. They essentially tax the success of their rival.

2. If Gemini Wins

Suppose Google’s internal teams crack the code to artificial general intelligence (AGI) first, and Gemini becomes the dominant global model. In this scenario, Google captures the pure, direct upside of the AI revolution, commanding the consumer market, enterprise software, and search ecosystems. The financing of Anthropic becomes a profitable side bet that generated healthy returns through infrastructure leasing.

3. If AI Demand Explodes Globally

Regardless of whether Claude, Gemini, or a yet-to-be-invented model claims the top spot, the macro trend is clear: the global demand for AI compute is skyrocketing. By building and financing massive, gigawatt-scale data centers, Google is hoarding the most valuable resource of the 21st century. They own the capacity that everyone will eventually need.

“During a gold rush, sell shovels.” Google has updated this adage for the digital age: During an AI rush, own the mountain, lease the shovels, and charge a toll on every nugget extracted.

Dodging the Bullets of Traditional M&A

Why didn’t Google just try to buy Anthropic outright? Beyond the sheer expense, acquiring massive AI startups carries three distinct, highly toxic risks that this financing arrangement beautifully sidesteps.

No Antitrust Headache

We are currently living in an era of unprecedented antitrust scrutiny. Regulatory bodies across the globe—from the FTC in the United States to the European Commission—are actively looking to break up tech monopolies. Microsoft’s deep partnership with OpenAI has already drawn the ire of regulators who view it as a backdoor acquisition.

If Google, the reigning king of search, attempted to buy Anthropic, it would trigger a years-long legal bloodbath. By simply acting as a financier and infrastructure provider, Google legally distances itself from antitrust claims. They are merely providing a service and a loan to an independent company. It is a legally unassailable position.

No Integration Risk

Corporate history is a graveyard of brilliant startups ruined by giant acquirers. Merging different corporate cultures, integrating disparate codebases, and aligning conflicting management styles destroys value more often than it creates it. By keeping Anthropic independent, Google avoids the bureaucratic friction that kills innovation. Anthropic gets to maintain its agility, its unique safety-focused culture, and its operational speed, all while feeding revenue back to Google’s infrastructure division.

No Talent Drama

The most valuable asset in an AI lab is not the current model; it is the human capital. The researchers and engineers building these systems are incredibly rare and highly sought after. Often, when a massive corporation buys a nimble startup, the top talent vests their stock options and immediately leaves. AI researchers want autonomy. By financing Anthropic rather than absorbing it, Google ensures that Anthropic’s brilliant engineers stay exactly where they are—happily building models that will inevitably run on Google’s servers.

The Bigger Shift: The Compute, Energy, and Finance Race

The Google-Anthropic deal illuminates a fundamental truth that the broader public is currently missing. The era of two guys in a garage building a world-changing software company is over—at least when it comes to foundation AI models.

AI is no longer just a model race. The math and the algorithms, while incredibly complex, are becoming table stakes. The true bottlenecks to AI supremacy are entirely physical and financial.

The Compute Bottleneck

Training a frontier AI model requires tens of thousands of specialized chips working in parallel for months. Historically, Nvidia has held a near-monopoly on these chips (GPUs), allowing it to dictate prices and become one of the most valuable companies on Earth. Google’s play with Anthropic is a direct assault on this monopoly. By stipulating that Anthropic uses Google’s custom-built TPUs, Google is validating its own silicon on the grandest stage. They are proving that you do not need to pay the “Nvidia tax” to build a world-class AI. This vertical integration—from the silicon to the data center to the cloud—gives Google an unparalleled margin advantage.

The Energy Bottleneck

You cannot simply build a gigawatt data center anywhere. The electrical grid in most developed nations is aging and incredibly fragile. Securing 500MW of continuous, reliable power requires years of negotiation with energy providers, local governments, and grid operators. In places like Texas, which operates its own independent power grid (ERCOT), managing the energy demands of AI is a geopolitical-level challenge.

Google, with its decades of experience building data centers and its massive investments in renewable energy, is uniquely positioned to solve this. Startups cannot negotiate gigawatt power purchase agreements. Trillion-dollar tech giants can. Energy is the new oil, and data centers are the refineries.

The Financing Bottleneck

Finally, there is the sheer cost of capital. Building a $5 billion data center requires cash. In a high-interest-rate environment, borrowing $5 billion on the open market would crush a startup like Anthropic with debt servicing costs. Google, however, generates tens of billions of dollars in free cash flow every quarter. They can self-finance or borrow at the lowest possible corporate rates. By passing these savings on to Anthropic in exchange for infrastructure lock-in, Google is wielding its balance sheet as a weapon.

The New Paradigm: The Landlord and the Lender

When we zoom out, we can clearly see the transformation of Google’s identity in the AI era. They are not content to just be another competitor slugging it out in the software arena. They are positioning themselves as the foundational bedrock of the industry.

They are The Landlord, leasing out the massive physical campuses where the digital brain of the future is being built.

They are The Lender, providing the lifeblood of capital required to sustain the astronomical costs of AI research.

They are The Infrastructure Layer, providing the custom silicon and cloud networking that makes the magic happen.

This strategy requires a profound level of corporate maturity. It requires setting aside the ego of needing to own the “best” model, in favor of owning the most profitable, secure, and defensible position in the value chain.

Owning the best model is fragile. A competitor can release a better model tomorrow, rendering months of compute and billions of dollars obsolete in an instant. The software layer of AI will face intense commoditization, price wars, and margin compression as the models become faster and cheaper.

Owning the system every model runs on, however, is not fragile. It is an ironclad monopoly on the future.

Sundar Pichai is not playing the same game as the rest of the industry. While others are fighting over who gets to be the king of the castle, Google is quietly buying the land, laying the bricks, and owning the roads that lead to the front gate. It is a masterful game of 3D chess, and the rest of the market is only just beginning to realize they are already in checkmate.

Share AI & Quantum Computing Newsletter

The artificial intelligence landscape is about to experience a major shift, marking the end of an era of irrational exuberance as it enters a phase of economic rationalization. Jensen Huang, the iconic CEO of Nvidia, recently confirmed the halt of massive direct investments in LLM (Large Language Model) giants like OpenAI and Anthropic.

For the tech ecosystem, the message is unequivocal: the generative AI sector is entering a phase of forced maturity. The era of mega-funding rounds, where hardware providers wrote blank checks to AI players to ensure their technological loyalty, seems to be definitively coming to an end. This change in course is not an admission of weakness, but rather the symptom of an industry transitioning from a tumultuous adolescence into adulthood.

Here is an in-depth analysis of the reasons driving Nvidia to revise its strategy, the flaws in the current AI economic model, and the direct repercussions this financial earthquake will have on end users and IT decision-makers.

Subscribe now

1. The Thunderclap at the Morgan Stanley Conference

The signal was given in a highly institutional setting, far from the usual spectacular stages of Silicon Valley. At the closely watched Morgan Stanley Technology, Media and Telecom conference, Jensen Huang was disarmingly clear. The man in the leather jacket, often perceived as the grand architect and chief cheerleader of the generative AI revolution, acknowledged a fundamental paradigm shift.

According to him, the opportunity to inject tens of billions of dollars into the undisputed LLM leaders is simply no longer relevant. The numbers speak for themselves: while persistent rumors credited Nvidia with an initial ambition to invest up to $100 billion to lock in its position with OpenAI, the company has drastically scaled back its ambitions, capping its final investment at “only” $30 billion.

“The opportunity to invest in a substantial company like this is probably behind us,“ Jensen Huang stated before an audience of financial analysts. This sentence, though measured, sounds the death knell for the frenetic funding model that dominated 2023 and 2024. Nvidia, which built a stratospheric market valuation by equipping AI pioneers with its ultra-powerful processors (the famous H100, B200, and their successors), believes it is time to diversify its support beyond its two historic clients.

This strategic withdrawal highlights an inescapable reality: Nvidia’s role is no longer to play incubator or patron to startups that have grown too large, but to consolidate its status as a universal provider of computing infrastructure.

2. The Race to IPO and the $1 Trillion Equation

If Nvidia is turning off the tap, it is primarily due to a financial timeline that has become incompatible with traditional investment strategies. OpenAI and Anthropic are no longer agile young startups looking for seed money; they are behemoths with valuations that defy gravity.

According to reports from Reuters, OpenAI is actively preparing for a monumental Initial Public Offering (IPO). The most optimistic projections suggest that this operation could value the structure led by Sam Altman at up to $1 trillion. At this level of capitalization, OpenAI would join the highly exclusive club of “Trillion-Dollar Companies” alongside Apple, Microsoft, Alphabet, and... Nvidia itself.

For Jensen Huang and his board of directors, the calculation is purely mathematical and pragmatic. The cost of entry to take a significant stake in a company valued at $1 trillion becomes exorbitant, and the potential Return on Investment (ROI) is mechanically diluted. Buying OpenAI shares today no longer offers the multiple returns expected from a classic Venture Capital strategy.

Furthermore, strategically speaking, owning shares in a public company of this magnitude creates potential conflicts of interest and attracts the attention of global antitrust regulators, who are already highly focused on Nvidia’s hegemony in the AI chip market. As these companies become financially autonomous and prepare to raise funds directly from public markets, the need to lean financially on their chip supplier disappears.

3. The Autopsy of the “Circular Investments” Model

Nvidia’s withdrawal also lifts the veil on a financial practice that was beginning to cause friction on Wall Street: the infamous “circular investments.” For several months, leading financial newspapers like the Financial Times had been pointing out doubts surrounding the true health of the sector and the transparency of these financial setups.

The Mechanics of the Financial Loop

Some analysts, becoming increasingly vocal, strongly criticized this closed-loop model. The principle was simple, although ethically questionable for some financial purists:

1. The Injection: Nvidia invests massively (e.g., hundreds of millions or billions of dollars) in a promising generative AI startup.

2. The Obligation: In exchange for this investment, or out of technological necessity, the startup commits to using these same funds to acquire the computing power needed to train its models.

3. Return to Sender: The startup massively buys servers equipped with Nvidia processors.

4. The Balance Sheet: The money invested by Nvidia returns to it as revenue, artificially inflating its quarterly results and reassuring the markets about the exploding demand for its chips.

By ending this practice, Nvidia is responding to these criticisms and cleaning up its balance sheet. Questioning this circular model is a necessary evil. It effectively forces LLM publishers to prove the viability of their economic model without subsidies from technological partners. Without this “magic money” returning to its source, OpenAI, Anthropic, and others must demonstrate that their revenues come from real clients (businesses and end consumers) and not an accounting artifice.

4. The End of Subsidized AI: What Impact on Users?

While this macroeconomic movement may initially seem purely financial and reserved for the hushed discussions of investment bankers, it will have a direct, tangible, and rapid impact on the evolution of the daily tools used by professionals.

Until now, users of generative AI tools have lived in a bubble. Just as Uber subsidized the true cost of rideshare trips for years using venture capital money to gain market share, the AI industry has subsidized the true cost of “compute” (computing power). Inference—that is, the act of asking ChatGPT or Claude a question and getting an answer—is extremely expensive in terms of electricity, cooling, and hardware depreciation.

Toward a Revision of Pricing Policies

Beyond questioning the frantic race for raw power (the mantra “the bigger the model, the better it is” is beginning to run out of steam against physical and financial realities), it is highly likely that this new paradigm is steering us toward a brutal evolution in pricing policies.

Because without the unconditional financial backing of the chipmaker, model providers will have to more accurately pass on the cost of computing to their customers. We can anticipate several short- and medium-term consequences:

• Rise in Premium Subscriptions: Standard $20 per month subscriptions could see their usage limits tighten, or new pricing tiers (”Pro”, “Enterprise”, “Ultra”) could emerge at significantly higher rates ($50, $100 per month).

• API Billing at True Cost: Developers integrating AI into their applications will likely see the cost per “token” adjust to reflect the economic reality of the underlying infrastructure.

• The Rise of “Small Language Models” (SLMs): To counter these exorbitant costs, the industry will accelerate its pivot toward smaller, more specialized models capable of running locally or on much less expensive cloud infrastructures.

5. Nvidia’s New Frontier: Strategic Diversification

By disengaging from the financial race for mainstream LLMs, Nvidia is not reducing its ambitions; it is redeploying them. Jensen Huang emphasized this: Nvidia must now diversify its support beyond its historic clients. The risk of dependency on two or three mega-clients is a sword of Damocles that any mature company seeks to avoid.

The billions saved on OpenAI and Anthropic are already being reinvested in new verticals that will shape the next technological decade:

1. Sovereign AI: Faced with geopolitical stakes, many countries (France, India, Japan) are investing heavily to create their own supercomputers and sovereign AI models. Nvidia is positioning itself as the infrastructure partner of nations, a potentially infinite market that is less volatile than Silicon Valley startups.

2. Healthcare and Generative Biology: The use of AI for discovering new drugs and protein folding (as with the Nvidia BioNeMo platform) represents a market where the added value justifies colossal investments.

3. Robotics and Physical AI: Nvidia’s GR00T project demonstrates a clear desire to conquer the physical world by equipping future humanoid and industrial robots with a “brain” capable of learning from its environment.

4. Industrial AI and Digital Twins: Through its Omniverse platform, Nvidia helps car manufacturers, architects, and industrialists simulate entire factories before they are built.

By diversifying, Nvidia ensures that if the mainstream LLM bubble were to deflate, its business model would remain perfectly resilient.

6. The Rite of Passage: A Paradoxical Validation of Success

It is crucial to interpret this decision through the right lens. By cutting the financial umbilical cord, Nvidia is paradoxically validating—in the strongest possible way—the resounding success of OpenAI and Anthropic.

In the venture capital ecosystem, an investor’s ultimate success is not to continue funding their gem indefinitely, but to realize that it no longer needs them to fly on its own. OpenAI and Anthropic have won the proof-of-concept war. They have proven that their technologies were not mere laboratory curiosities, but tools capable of transforming global productivity.

They are now big enough to face the stock markets alone. Their ability to generate billions of dollars in Annual Recurring Revenue (ARR) in record time has reassured traditional investment banks. The baton has thus been passed from hardware providers to Wall Street’s institutional investors. This is the ultimate rite of passage for any technology company seeking long-term viability.

7. The New Playbook for IT Decision-Makers

For Chief Information Officers (CIOs) and IT decision-makers, this announcement is not a simple financial anecdote; it is an alarm bell that should trigger an immediate review of partnership strategies.

These partners must now be evaluated not as infinitely growing startups carried by irrational enthusiasm, but as mature players whose financial stability will now depend on their ability to generate real profit. “FinOps” (the financial optimization of the Cloud) must now urgently incorporate an “AI-Ops” component.

Actionable Steps to Take

• Audit Exposure to API Costs: Identify all workflows within your company that depend on large LLM APIs. Calculate the financial impact of a potential 20% to 50% increase in token rates over the next 18 months.

• Adopt a Multi-Model Strategy: Do not be dependent on a single provider. Integrating open-source models (like Llama 3 or Mistral) hosted on your own infrastructure or via competing Cloud providers helps mitigate the risk of vendor lock-in.

• Demand Clear SLAs: As they become mature companies preparing for their IPOs, OpenAI and Anthropic must now provide Service Level Agreements (SLAs) worthy of industry standards (guaranteed uptime, absolute data confidentiality, dedicated technical support).

• Calculate True ROI: The days of running Proof of Concepts (POCs) simply for the appeal of novelty are over. Every generative AI project must now demonstrate a measurable Return on Investment, whether in time savings or revenue generation, capable of offsetting inference costs that are bound to rationalize upward.

Ultimately, Nvidia’s financial withdrawal sounds the end of the generative AI honeymoon, but it opens the much more exciting chapter of its long-term industrialization. The Wild West is making way for technological civilization.

Share AI & Quantum Computing Newsletter

For the past two years, as the tech industry fell into a collective frenzy over generative AI, one of the foundational “Godfathers of AI” stood at the periphery of the hype cycle, repeating a single, deeply unpopular idea.

While Silicon Valley executives declared that human-level artificial intelligence was just a few massive data centers away, Turing Award winner Yann LeCun kept pouring cold water on the parade. Current AI, he argued relentlessly, doesn’t actually understand anything. In his words, the massive language models dazzling the world were nothing more than “glorified autocomplete”—systems that simply predict the next word in a sequence based on statistical probabilities gleaned from reading the internet.

For a long time, the industry brushed off his critiques. When your technology is adding trillions of dollars to global market caps, it is easy to ignore the skeptics.

But LeCun isn’t just a skeptic. He is a builder. And he is done simply pointing out the flaws in the current paradigm. He is now building the alternative.

In a move that has sent shockwaves through the global technology landscape, LeCun has unveiled his new startup: AMI Labs (Advanced Machine Intelligence). And he isn’t starting small. AMI Labs has just raised a staggering $1.03 billion Seed round—the largest Seed round in European history—launching the company with an immediate valuation of roughly $3.5 billion.

This is not just another AI startup building a bigger, slightly more polite chatbot. This is a colossal, billion-dollar bet against the prevailing “just scale LLMs” strategy. It is a fundamental pivot in how we approach machine cognition. If LeCun is right, the AI tools we marvel at today are not the dawn of true intelligence; they are merely the prototype phase.

The future of AI is being built in Europe, and it looks entirely different from what we have been led to expect.

Subscribe now

The Autocomplete Trap: Why LLMs Hit a Wall

To understand the magnitude of what AMI Labs is attempting, we first have to understand the fundamental limitations of the technology currently dominating the market.

Large Language Models (LLMs) like GPT-4, Claude, and Gemini are undeniable marvels of engineering. They have revolutionized coding, copywriting, customer service, and data analysis. However, their underlying architecture is based on a surprisingly simple premise: autoregression. Given a sequence of text, what is the most mathematically probable next word (or token)?

They learn this by ingesting virtually the entire written output of humanity. They read Wikipedia, Reddit, digitized books, and scientific journals. But reading about the world is not the same as experiencing it.

As LeCun has consistently pointed out, language is an incredibly low-bandwidth, highly compressed representation of reality. Human beings do not learn exclusively, or even primarily, through text. We learn through physical interaction, observation, and sensory feedback.

Because LLMs only know text, they suffer from profound structural deficits:

• They struggle with physical reasoning: An LLM can write a beautiful poem about a bicycle, but it has no intuitive understanding of how a bicycle balances, how metal feels, or what happens if you push a bicycle down a flight of stairs.

• They lack true long-term memory: While context windows are expanding, LLMs do not build persistent, evolving models of the world or their users in the way a human brain does. They start relatively fresh with every interaction.

• They cannot plan effectively: True intelligence requires the ability to conceptualize a complex goal, break it down into sequential steps, anticipate physical or logical roadblocks, and adjust the plan dynamically. LLMs struggle with hierarchical planning because they are fundamentally reactive, generating one token at a time without a persistent “mind’s eye” envisioning the outcome.

Less autocomplete, LeCun argues, and more understanding. You cannot reach Artificial General Intelligence (AGI) simply by feeding more text into a bigger autocomplete engine. You have to change the engine entirely.

The Toddler Benchmark: Learning How Reality Works

If reading the entire internet doesn’t create true intelligence, what does? AMI Labs is entirely focused on a concept known as World Models.

To grasp the concept of a world model, you don’t need to look at a supercomputer; you just need to look at a toddler. Think about how a human child learns the fundamental physics of the universe.

A baby sits in a highchair and drops a plastic cup. The cup falls, hits the floor, and makes a sound. The baby watches this happen. Then, the baby does it again. And again.

Through this simple, repetitive observation and interaction, the child is not memorizing a string of text that says “objects with mass are attracted to the Earth’s center.” Instead, the child’s brain is building an internal, intuitive physics engine. They internalize gravity. They learn object permanence (knowing the cup exists even when it rolls out of sight). They understand spatial relationships, momentum, and cause-and-effect.

By the time a human is a teenager, they have amassed thousands of hours of rich, high-bandwidth sensory data—video, audio, tactile feedback—that forms a robust “world model.” This model allows us to navigate a crowded sidewalk without bumping into people, catch a thrown ball, or know intuitively that a glass will shatter if dropped on concrete, but bounce if dropped on a mattress.

AMI Labs aims to build AI that learns exactly this way. Instead of scraping text from the web, these systems will learn by observing and interacting with the world, ingesting massive amounts of video and sensory data to build a comprehensive, internal simulator of reality.

Autoregressive vs. JEPA: The Architecture Clash

To truly appreciate the pivot AMI Labs is making, we need to look under the hood and contrast the structural differences between today’s dominant AI models and the alternative LeCun proposes. The technical battleground comes down to two distinct approaches: Autoregressive Generative Architectures and Joint-Embedding Predictive Architectures (JEPA).

The Generative Approach: Pixel-by-Pixel Perfection (and Exhaustion)

Current models—whether they are LLMs predicting text or video models like Sora predicting frames—rely heavily on generative, autoregressive architectures.

In an autoregressive video model, the system looks at the current frame and attempts to predict the exact next frame, pixel by pixel. If you feed the model a video of a car driving down a street lined with trees on a windy day, the model uses vast amounts of compute trying to predict the exact movement of every single leaf on the tree, the precise reflection of the sun on the windshield, and the exact texture of the asphalt.

This approach creates two massive problems:

• Computational Overload: Predicting every single detail of reality is incredibly expensive and inefficient. It forces the system to model “noise” that doesn’t actually matter to the core action.

• Error Accumulation: Because these models predict one step at a time, a tiny mistake in step one gets fed into step two. Over a long sequence, these errors compound exponentially. In text, we call this a “hallucination.” In physical reasoning or video generation, the model quickly loses its grip on reality, resulting in objects morphing into one another or physics breaking down.

The JEPA Approach: Predicting the Abstract

AMI Labs is staking its multi-billion-dollar valuation on the JEPA framework, pioneered by LeCun himself. JEPA fundamentally abandons the idea of generating every pixel or token.

Instead of predicting the exact future data, JEPA predicts the abstract, hidden states of the future. Returning to the car, driving down the street: a human brain does not consciously track every leaf blowing in the wind. We filter out that irrelevant noise and focus on the semantic reality of the scene: the trajectory of the car, the speed of the pedestrians, and the state of the traffic light.

JEPA operates on this exact biological principle through a multi-step process:

1. Observation (x and y): The model takes in an initial state (x) and a future state (y).

2. Encoding: It uses an encoder to compress both states into abstract representations. It strips away the visual noise (the blowing leaves) and extracts the core meaning (the car’s position).

3. Prediction: A predictor module then tries to guess the abstract representation of the future state based on the abstract representation of the current state.

Because JEPA isn’t trying to draw the leaves, it saves massive amounts of compute. More importantly, it allows the model to learn the rules of the world (physics, object permanence, momentum) without getting bogged down by the unpredictable chaos of surface-level details. This results in an internal simulator of reality that is grounded, logical, and vastly more robust than autoregressive models.

A Head-to-Head Comparison

To summarize the technical paradigm shift, here is how the two architectures stack up:

The War Chest: Breaking Down Europe’s Largest Seed Round

You do not pivot the entire direction of the global technology sector without serious capital. Building foundational AI models—especially ones that require ingesting and processing petabytes of video and spatial data—is an extraordinarily expensive endeavor.

But investors clearly believe in LeCun’s vision. The financial backing behind AMI Labs is a testament to the growing realization that the current LLM paradigm may be approaching an asymptote.

The $1.03 billion Seed round is unprecedented, particularly in the European ecosystem. Valuing the company at roughly $3.5 billion right out of the gate, the round was backed by a coalition of heavy hitters that bridge the gap between legacy tech dominance, venture capital, and sovereign wealth:

• Bezos Expeditions: Jeff Bezos’s personal investment firm has long been interested in the physical applications of AI, particularly in robotics and logistics—areas where world models are desperately needed.

• NVIDIA: The undisputed king of the AI hardware boom. NVIDIA’s backing is both a financial endorsement and a strategic alignment. Training world models will require entirely new paradigms of compute, and NVIDIA wants to ensure its chips remain the bedrock of whatever architecture comes next.

• Temasek: The Singaporean state-owned investment firm brings massive sovereign capital to the table, ensuring AMI Labs has the runway to pursue long-term research without the pressure of immediate commercialization.

• HV Capital: One of Europe’s leading venture firms, signaling strong regional support for keeping this technological revolution anchored across the Atlantic.

This capital will fuel a massive, global talent acquisition strategy. AMI Labs is establishing premier research hubs across Paris, New York City, Montreal, and Singapore.

The choice of Paris as the headquarters is particularly notable. Europe is rapidly becoming the epicenter of the next wave of AI innovation. Following the massive success of Mistral AI, AMI Labs cements France’s status as a top-tier destination for elite AI researchers who want to build foundational technology outside the Silicon Valley echo chamber.

The $1 Billion Bet Against “Scale is All You Need”

Perhaps the most fascinating aspect of the AMI Labs launch is what it represents philosophically. For the last several years, the AI industry has been governed by an unwritten religion known as the “Scaling Hypothesis.”

The mantra was simple: Scale is all you need. If a 100-billion parameter model is smart, a 1-trillion parameter model will be smarter. Just add more data, more GPUs, and more electricity, and eventually, the autocomplete engine will magically cross the threshold into general intelligence.

Companies like OpenAI, Microsoft, and Google have committed tens of billions of dollars to this specific strategy, planning data centers so large they require dedicated nuclear power plants to run.

AMI Labs represents the first heavily capitalized, existential threat to that thesis. LeCun is essentially looking at the largest companies in the world and telling them they are optimizing the wrong algorithm.

If LeCun’s world models prove successful—if they can reason, plan, and understand the physical world in ways that LLMs fundamentally cannot—it will trigger a massive realignment in the tech industry. The billions currently being poured into scaling autoregressive transformers may suddenly look like a vast misallocation of capital.

Furthermore, world models are the missing link for the physical deployment of AI. We have chatbots that can pass the bar exam, but we still do not have a robot that can reliably navigate a messy kitchen, clear a table, and load a dishwasher. Physical robotics requires physical reasoning. Autonomous driving requires an intuitive understanding of human intent and momentum. AMI Labs is building the software capable of powering the physical economy, not just the digital one.

The Prototype Phase of Humanity’s Greatest Invention

It is easy to look at the AI tools available on our smartphones today and feel a sense of awe. The ability to converse with a machine in natural language, to have it write code or generate art, feels like science fiction made real.

But history is full of technologies that looked like the finish line, only to be revealed as mere stepping stones. The telegraph was a miracle until the telephone arrived. The vacuum tube was revolutionary until the silicon transistor changed the world.

Yann LeCun’s $1 billion bet on AMI Labs is a bold declaration that we are still in the vacuum tube era of artificial intelligence. Large Language Models have shown us that machines can mimic fluency. But fluency is not cognition. Memorization is not comprehension.

By pivoting away from text-based autocomplete and toward JEPA-style world models that learn by observing and simulating reality, AMI Labs is attempting to bridge the gap between artificial mimicry and genuine artificial intelligence.

Europe is officially cooking. And if this massive seed round yields the breakthroughs LeCun anticipates, we may soon look back on the era of chatbots as nothing more than the charming, limited prototype phase of a much grander technological revolution.

Share AI & Quantum Computing Newsletter

The Great Flip: Why the World is Pivoting from CBDCs to Stablecoins.

The line between “crypto” and “real finance” is now officially, irreversibly blurry.

Everyone is talking about AI. At the coffee machine, in board meetings, on TV shows, and even during family dinners. It’s the defining topic of our decade. But let’s be honest: few people really know what this term actually covers. For many, AI is a sort of magic black box—sometimes scary, often misunderstood—that is blithely confused with sci-fi robots or upgraded search engines. The technical jargon doesn’t help: you’re bombarded with acronyms like LLM, GPT, or Machine Learning without ever being handed the decoder ring.

Since Artificial Intelligence is at the heart of this newsletter and will fundamentally redefine the way we work over the coming years, I thought it was essential to hit pause. We need a major checkpoint to help you see things clearly, once and for all.

Think of this guide like a set of Russian nesting dolls. We are going to start with the broadest concept, Artificial Intelligence, and open it up, layer by layer, until we reach the precise tool you use every single day.

Here are the 10 key concepts you need to master. Ready? Let’s dive in.

Subscribe now

1. Artificial Intelligence (AI): The Grand Illusion of Human Thought

This is the umbrella term, the largest of the Russian dolls. Artificial Intelligence designates any computer system capable of performing tasks that would normally require human intelligence.

• The key concept: Imitating human thought and actions.

• Concretely: AI is not a single technology; it’s a field of study, just like physics or biology. Its ultimate goal is to perceive an environment, reason, learn, and solve problems. This ranges from the ultra-basic computer programs of the 90s capable of playing chess (by brute-forcing all possibilities) to the autonomous systems of today.

• The analogy to understand it: Artificial Intelligence is like the term “Vehicle.” It encompasses many very different things: a bicycle, an electric scooter, a sports car, or a commercial airliner. It’s a broad concept, not a specific mechanism.

2. Machine Learning: No More Old-School Programming

If we open the AI Russian doll, we find Machine Learning. This is where the true modern revolution began. Before Machine Learning, for a computer to do something, a developer had to write strict, hard-coded rules (if A happens, then do B).

• The key concept: Learning from data, not from hard-coded rules.

• Concretely: Instead of giving the computer the exact rules to accomplish a task, we give it thousands of examples (data) and let it figure out the rules on its own. For example, to make a spam filter work, we don’t tell it to block emails containing the word “Viagra.” We show it a million spam emails and a million normal emails, and we say: “Find what makes them different.”

• The analogy to understand it: It’s the difference between learning to ride a bike by reading a physics manual on balance and center of gravity (classical programming) versus learning by getting on the bike, falling, and naturally adjusting your balance through trial and error (Machine Learning).

3. Neural Networks: Computing Biomimicry

Inside Machine Learning is a specific technique: neural networks. Researchers asked themselves a simple question: What is the best learning system we know of? The human brain. Can we copy it?

• The key concept: Recognizing patterns by drawing inspiration from the architecture of our brain.

• Concretely: An artificial neural network is made up of “nodes” (tiny mathematical programs) connected. Each node receives information, performs a small calculation, and passes the result to the next node. These nodes are organized into layers. They are exceptionally good at recognizing complex patterns, like the contours of a face in a photo or the intonations of a voice.

• The analogy to understand it: Imagine a highly complex automobile assembly line. The first worker (neuron) checks if a part is metal. They pass it to the second, who checks if it’s round. The third checks if it has a hole in the middle. At the end of the line, the floor manager (the result) declares: “It’s a wheel!”

4. Deep Learning: The Power of Depth

Let’s go one level deeper. Deep Learning is what allowed AI to explode over the last decade, thanks to the massive increase in computer processing power and the abundance of data on the internet.

• The key concept: Stacking a massive number of neural layers to solve incredibly complex tasks.

• Concretely: The word “Deep” doesn’t refer to a deep philosophical understanding, but to the number of layers in the neural network. A classic network might have 2 or 3 layers. A Deep Learning model can have hundreds. It is this “depth” that allows autonomous cars to distinguish a pedestrian from a lamppost in the pouring rain, or medical algorithms to detect tumors invisible to the naked eye on X-rays.

• The analogy to understand it: If a classic neural network is a magnifying glass, Deep Learning is a high-precision electron microscope.

5. Transformers: The Revolution of Attention and Context

We are now entering the core of the modern reactor. In 2017, Google researchers published a scientific paper titled “Attention Is All You Need”. That paper changed everything. In it, they presented a new neural network architecture: the Transformer.

• The key concept: Analyzing the relationships between all the words in a text simultaneously, thanks to the attention mechanism.

• Concretely: Before Transformers, AIs read text word by word, from left to right. When they reached the end of a long paragraph, they had often “forgotten” the beginning. The Transformer, however, looks at the whole sentence at once. It mathematically calculates which words are linked to each other (the “attention mechanism”), even if they are far apart in the text. This gives it a lightning-fast understanding of context, nuance, and irony.

• The analogy to understand it: Older AIs read a book through a tiny slit that only lets them see one word at a time. The Transformer is someone who reads the entire page at a single glance, immediately understanding how the first paragraph connects to the last.

6. Generative AI: From Observer to Creator

Up until now, the technologies we’ve looked at were mainly used to analyze, classify, or predict. They would say, “This is a photo of a cat,” or “This text is positive.” Generative AI marks a fundamental breakthrough: it no longer just analyzes the world; it creates it.

• The key concept: Producing brand-new content (text, image, sound, video), rather than just analyzing it.

• Concretely: By building on the previous technologies (Deep Learning, Transformers), Generative AI learned the rules that govern human language, visual art, or music. Once these rules are assimilated, they can be used “in reverse” to generate new things that have never existed before.

• The analogy to understand it: Traditional AI is an extremely sharp art critic capable of recognizing Van Gogh’s style with 100% accuracy. Generative AI is the painter to whom you hand a brush and say: “Paint me a brand-new portrait in the style of Van Gogh.”

7. Large Language Models (LLM): Living Encyclopedias

While Generative AI can create images or music, Large Language Models (LLMs) are the branch specifically dedicated to text and human communication.

• The key concept: Mathematical models trained on billions of texts to perfectly master language.

• Concretely: An LLM is a massive neural network based on the Transformer architecture. During its training, it is made to “read” a huge chunk of the internet: books, Wikipedia articles, forums, websites, and computer code. Its core mission is surprisingly simple: play a guessing game by trying to predict the next word in a sentence. By training on billions of data points, it acquires not only perfect grammar but also staggering general knowledge and reasoning capabilities.

• The analogy to understand it: It’s the “auto-complete” function on your smartphone, but on steroids—having read every book in the Library of Alexandria and capable of writing a philosophy thesis instead of just finishing your text message.

8. GPT (Generative Pre-trained Transformer): The Engine of the Revolution

We have arrived at the proper nouns. GPT is not a general concept; it is the name of a specific technology invented and developed by the company OpenAI.

• The key concept: An ultra-high-performance LLM specialized in text generation.

• Concretely: Let’s break down its name. Generative: it creates text. Pre-trained: it ingested a massive amount of raw information upfront to understand the world before being specialized. Transformer: this is the architecture of its “brain” (see point 5). GPT is essentially the pure software “engine.” It is code and billions of mathematical parameters.

• The analogy to understand it: GPT is like a phenomenally powerful V8 engine designed by a car manufacturer. This engine is fabulous, but as it sits on a workbench, you can’t really use it to go grocery shopping.

9. GPT-5: The Current Peak of Engineering

GPT technology evolves in iterations. There was GPT-2, GPT-3, the historic GPT-4 that shook the world, and today, we are in the era of GPT-5.

• The key concept: The most advanced, complex, and capable version of the GPT family to date.

• Concretely: Each new generation of a model isn’t just “smarter.” It’s more reliable, it “hallucinates” (invents facts) much less, and most importantly, it becomes “multimodal.” GPT-5 no longer just processes text in isolation; it fluidly understands, sees, and hears the world, simultaneously analyzing text, images, voice, and complex videos with reasoning capabilities close to the best human experts in many fields.

• The analogy to understand it: If GPT-3 was a brilliant high schooler and GPT-4 was an excellent recent college grad, GPT-5 is a seasoned, tenured university professor, capable of debating, drawing concepts on a chalkboard, and analyzing a film all at the same time.

10. ChatGPT: The Interface That Changed Everything

Here we are at the smallest of the Russian dolls, the one you hold in your hands. In late 2022, OpenAI pulled off a stroke of genius. They had an incredible engine (GPT), but they realized no one would use it if you had to be a software engineer to talk to it. So, they created ChatGPT.

• The key concept: The AI chat assistant you use daily.

• Concretely: ChatGPT is “just” the website, the mobile app, and the user interface. It’s the ergonomic wrapper designed to converse with you in the most natural way possible. It’s the tool that takes your casual sentences, feeds them into the GPT engine, retrieves the engine’s ultra-intelligent response, and presents it to you as a friendly text or voice message.

• The analogy to understand it: If GPT is the overpowered engine, ChatGPT is the brand-new sports car built around it. It has a steering wheel, comfortable leather seats, an intuitive dashboard, and voice assist. It makes driving this complex engine accessible to anyone with a driver’s license.

You Are Now Equipped!

The next time you hear the term “AI” in a meeting or read a news article talking about “Large Language Models” powering “ChatGPT,” you will know exactly what they are talking about.

Artificial Intelligence is no longer dark magic. It is a combination of cutting-edge mathematics, enormous volumes of data, and brilliant software engineering, brought together in tools designed to amplify our own intelligence. Understanding these concepts is not just for developers; it is the essential prerequisite for any professional wanting to stay relevant in today’s world.

Now that the jargon holds no secrets for you, it’s time to move from theory to practice.

Share AI & Quantum Computing Newsletter

The image is striking, almost dystopian, yet it may soon become our everyday reality. Imagine your local coffee shop—your usual Starbucks—no longer filled with college students cramming for midterms or freelancers working on creative side hustles, but instead populated by former executives, marketing directors, financial analysts, and corporate lawyers. Highly qualified professionals, graduates of top-tier universities, suddenly sidelined from the labor market by an invisible and lightning-fast force: artificial intelligence.

This is the brutal warning issued by Andrew Yang. The former U.S. presidential candidate, known for his pragmatic and often avant-garde analyses of socio-economic dynamics, is sounding the alarm. According to him, we are not at the dawn of a smooth transition, but standing at the edge of a precipice. Within the next 12 to 18 months, millions of so-called “white-collar” jobs could simply vanish.

This is no longer science fiction; it is an accelerating economic reality described by Yang. Let us deeply analyze the ramifications of this impending tectonic shift.

Subscribe now

1. The End of Cognitive Immunity: From the Factory Floor to the Office

For decades, the narrative of automation followed a predictable, well-worn path. Machines, robots, and first-generation algorithms were designed to replace manual, repetitive, and physically demanding labor. Blue-collar workers on automotive assembly lines, in logistics warehouses, and across the manufacturing sector were the first to bear the full brunt of modernization-induced layoffs.

During this time, knowledge workers—the white-collar class—believed they were safe. Their primary tools were their brains, their critical thinking, their analytical skills, and their creativity. The underlying assumption was that while machines could replace muscle, they could never replicate complex human thought.

Today, that illusion of security is shattering. As Andrew Yang points out, AI is no longer content with automating industrial processes; it is directly targeting intellectual labor. Generative artificial intelligence and advanced large language models are fundamentally redefining the equilibrium of skilled employment. Industries that once thought themselves untouchable are already recording job cuts numbering in the tens of thousands.

“A family member used AI to code a website this week,” noted Andrew Yang. “She accomplished in a few minutes what previously took days of work for a designer or a firm.”

This specific example is merely the tip of the iceberg. This spectacular efficiency comes at a heavy price: human employment itself.

2. The Decimation of Middle Managers and Analysts

Who exactly is in the crosshairs? Primarily, jobs that consist of gathering information, aggregating data, analyzing it, and producing reports. These tasks, which form the backbone of countless corporate sectors, are now perfectly digestible for AI.

Here are the professions and sectors most immediately vulnerable:

• Finance and Accounting: Financial analysts spend hours scrutinizing balance sheets and market trends. AI can ingest years of global financial data and generate predictive reports in a matter of seconds.

• Law: Paralegals and junior associates spend considerable time conducting case law research and drafting contract templates. These tasks are increasingly being delegated to specialized software.

• Software Development: Coders, once considered the architects of the modern world, are watching AI write, debug, and optimize complex lines of code with disconcerting speed.

• Marketing and Content Creation: From drafting advertising campaigns to designing graphic elements, generative AI produces professional-grade results instantly.

• Human Resources and Consulting: Resume screening, preliminary candidate evaluation, and organizational analysis are becoming heavily automated.

Andrew Yang specifically targets middle management. In the modern corporate hierarchy, middle managers are often tasked with bridging the gap between strategic leadership and operational execution. They are expensive to maintain, and when faced with an AI capable of synthesizing information and tracking project progress flawlessly, they suddenly become highly replaceable.

Yang predicts the layoff of millions of mid-career office workers within the next year and a half. He warns:

“Even a reduction of a few million would be tectonic... It will lead to a great vanishing of white-collar jobs.”

Many of these skilled professionals will be forced to accept a brutal social and financial demotion. Yang cites the story of a friend who, after losing his job, was forced to volunteer at a local non-profit managing their IT infrastructure—earning roughly 80% less than his previous salary. It is a steep, vertical decline in living standards.

The Bitcoin “Steel Plate” Stress Test: Will Your Backups Survive a Fire?

You secured your generational wealth with flawless cryptography, but trusted the physical backup to a $15 Amazon purchase. Here is why most ‘fireproof’ wallets melt when you need them most.

3. The Domino Effect: Local Economies and Real Estate in Peril

The impact of this announced purge will not stop at the glass doors of corporate skyscrapers. The economy of a bustling city or an affluent suburb relies on a complex, fragile ecosystem fueled directly by the purchasing power of the executives who work and live there.

If millions of white-collar workers lose their jobs or see their incomes drastically reduced, the shockwave will instantly ripple through the local economy.

The Collapse of Proximity Services

Businesses that revolve around the daily lives of office workers will be the first collateral victims:

• Lunchtime restaurants, specialty coffee shops, and local bistros.

• Dry cleaners, tailors, and shoe repair shops.

• Private concierges, dog-walkers, and in-home nannies.

• Premium fitness centers and downtown hair salons.

The Real Estate Crisis in High-Cost Areas

Residential and commercial real estate is bracing for unprecedented downward pressure. Yang rightly points out that wealthy residential suburbs and regions with prohibitive costs of living—such as Silicon Valley in California or Westchester County in New York—risk seeing their property values plummet.

Families, stripped of their primary incomes, will be unable to meet their mortgage payments or cover daily expenses. Many will be forced to sell at a loss and relocate to more affordable regions. In the context of this potential real estate panic, Yang’s advice is chilling:

“It might not be great to be the first one, but you do not want to be the last [to sell and move].”

The Specter of “Urban Wastelands”

Corporate real estate is equally exposed. With the declining need for a human workforce to perform intellectual tasks, massive office towers will become obsolete. Andrew Yang warns that these empty structures could quickly turn into “urban wastelands.” Dynamic metropolises like New York, Chicago, San Francisco, and Los Angeles risk seeing entire business districts hollowed out, leading to a drastic drop in tax revenues for these municipalities and subsequently threatening local public services.

4. The Agony of the Social Contract and a Sacrificed Generation

Beyond the stock market charts and employment statistics, the very fabric of the American dream (and more broadly, the Western promise) is being torn apart. Since the end of the Second World War, an implicit pact has united society and its citizens.

“The social contract of ‘study hard, go to school, get a good job, live a decent life’ is about to disappear,” warns Andrew Yang.

This social contract rested on the idea that academic merit and hard work offered a guarantee of material security. Today, university degrees—often acquired at the cost of crushing student debt—are depreciating at breakneck speed. What is the point of going into debt for twenty years to get a degree in law, journalism, or finance if an AI can do 80% of a junior employee’s work for a tiny fraction of the cost?

Young Graduates Hitting a Wall

The data shared by Yang is alarming: less than 30% of college seniors are currently finding jobs in their field of study, while more than half suffer from underemployment. They find themselves overqualified for the precarious gig-economy jobs they take just to survive.

This situation creates fertile ground for deep social resentment. Yang predicts a rise in anger and frustration, particularly among these highly educated workers. Having played by all the rules and made all the demanded sacrifices, only to see the final reward confiscated by an algorithm, generates a major crisis of purpose and growing distrust in institutions.

5. The Numbers Behind a Crisis Already in Motion

To understand that the warnings of Yang—a political figure with multiple hats (lawyer, entrepreneur, lobbyist, and co-chair of the Forward Party alongside Christine Todd Whitman and Michael S. Willner)—are not merely alarmist hypotheses, one only needs to look at recent data from the U.S. labor market. The foundation is already cracking.

A report from the U.S. Department of Labor published in January 2025 reveals deeply concerning structural tensions:

• Long-term unemployment: Over 1.6 million Americans have been looking for work for at least six months.

• Overall increase: Total unemployment surpassed 7 million in November 2024.

• Acceleration: This number represents an increase of more than 50% since the end of 2022.

• Search duration: The average time it takes to find a new job now reaches six months, which is about a month longer than at the peak of the post-pandemic rebound in early 2023.

While it is true that the economy generated over two million jobs in 2024, these creations mask a darker reality: a sharp sectoral divide. The difficulties in finding re-employment are tragically concentrated in well-paid white-collar jobs.

The technology, law, and media sectors—all of which hired massively following the COVID crisis—have abruptly slammed the brakes on hiring. The budgets previously allocated to human recruitment are now being heavily redirected toward colossal investments in artificial intelligence. Corporations are betting on algorithmic profitability at the expense of human capital.

Final Thoughts: Facing the Inevitable

Andrew Yang’s analysis paints a picture of a harsh and unforgiving immediate future for millions of professionals. The question is no longer if AI will automate intellectual work, but how fast it will dismantle existing socio-economic structures.

The anger, the real estate collapse, the urban wastelands, and the endemic unemployment among the upper-middle class are not absolute inevitabilities, but they are major systemic risks that demand immediate political, economic, and individual preparation. The old paradigm is dead. Clinging to the idea that basic cognitive skills will be enough to guarantee a secure job has become a dangerous blind spot.

The challenge now, for both individuals and society, is to learn how to navigate a world where intellectual production has become an abundant and cheap commodity, and to redefine the intrinsic value of human contribution far beyond the simple execution of automatable cognitive tasks.

Share AI & Quantum Computing Newsletter

The Great Shift: When Artificial Intelligence Becomes a Capital-Devouring Machine.

From Software Margins to Concrete Debt: The Financial Rewiring of the AI Era.

Key Takeaways

• NVIDIA knowingly ignored warnings of illegality to acquire 500 TB of books.

• Commercial pressure (GTC 2023) took precedence over legal ethics.

• The music industry is demanding “GDP-level” damages, adding chaos to the legal landscape.

• “Fair Use” is NVIDIA’s last shield, but the “Green Light” email may have shattered it.

This is a story of a white-collar heist on a planetary scale, where the loot isn’t gold or currency, but the sum of digitized human knowledge. Imagine the scene: on one side, NVIDIA, the Silicon Valley titan, a company that recently shattered the $3 trillion market cap ceiling thanks to its H100 chips. On the other, Anna’s Archive, a “shadow library” operating in the web’s gray zones, compiling millions of books without paying a single cent in royalties.

Between them? A series of leaked internal emails that reveal a brutal truth: to win the Artificial Intelligence race, the “Green Giant” didn’t just crawl the open web; it knowingly negotiated and paid for access to pirated content.

Here is the autopsy of a scandal that redefines the boundaries of technological ethics and corporate cynicism.

Subscribe now

1. The Leak: Case No. 1:26-cv-00002

It all started with a Class Action lawsuit filed in the Federal Court of New York. Within the documents of Case No. 1:26-cv-00002, the internal communications of NVIDIA’s “Data Strategy” team were laid bare.

These aren’t just watercooler rumors; they are “smoking gun” evidence. The emails reveal a team under immense pressure, obsessed with a single objective: feeding the beast. To ensure NVIDIA’s AI models—like those unveiled at the GTC 2023 conference—could compete with the likes of OpenAI or Google, they needed text. Not just messy Reddit threads or SEO-optimized blogs, but books. Millions of them.

The 500 TB Buffet

NVIDIA wasn’t looking for small samples. They targeted the entire catalog hosted by Anna’s Archive, which aggregates the world’s most notorious piracy hubs:

• Books3: A collection of 196,000 books sourced from the Bibliotik tracker.

• LibGen (Library Genesis): The global gold standard for scientific and academic texts.

• Sci-Hub: The “Robin Hood” of research, hosting millions of paywalled papers.

• Z-Library: A massive repository of fiction and non-fiction.

In total, NVIDIA sought to ingest nearly 500 terabytes of data. To facilitate this massive transfer, the emails reveal that NVIDIA paid “high-speed access fees” amounting to tens of thousands of dollars. They weren’t just “finding” data; they were subscribing to a pirate service.

2. The Surreal Dialogue: “Do you have permission?”

Perhaps the most mind-blowing moment in this saga is the exchange between the “pirates” and the “engineers.” In a series of emails, the administrators of Anna’s Archive—hardly strangers to illegal activity—seemed almost concerned for their high-profile client.

They explicitly asked NVIDIA: “Do you have internal authorization to handle this data? You are aware that these collections are technically illegal?”

NVIDIA’s response came in less than a week. No panic, no long-winded legal consultation, no hesitation. The message was simple: “Green light.”

Strategic Analysis: This “Green Light” response is the nail in the coffin for NVIDIA’s defense. In copyright law, there is a fundamental distinction between accidental infringement and “willful infringement.” By ignoring an explicit warning, NVIDIA moved itself into the latter category, which can increase statutory damages by a factor of a hundred.

3. The GTC 2023 Pressure: The End Justifies the Means

Why the rush? The answer lies in three letters: GTC. The GPU Technology Conference in March 2023 was the moment Jensen Huang, NVIDIA’s CEO, had to prove to the world that his company wasn’t just selling “shovels” for the AI gold rush, but was also a leader in Large Language Model (LLM) development.

To train a high-performing model, one must follow specific scaling laws, often summarized by the “Chinchilla” findings. We can approximate the data needs as follows:

Where D is the number of tokens (words or fragments) and N is the number of parameters in the model. For models with hundreds of billions of parameters, the “clean” web isn’t enough. You need literature, science, and complex reasoning—the kind only found in books.

NVIDIA made a cynical calculation: the legal risk of being caught was less dangerous than the industrial risk of presenting an “unintelligent” AI at their flagship event.

4. Anna’s Archive: Public Enemy Number One

While NVIDIA scrambles to manage the PR fallout, Anna’s Archive is already fighting on a second, even more surreal front.

The heavyweights of the music industry—Spotify, Universal Music, Warner, and Sony—have filed a lawsuit seeking the astronomical sum of $13 trillion. To put that number in perspective:

• It is roughly the GDP of China.

• It is more than four times the total market capitalization of NVIDIA.

Why such a staggering number?

The lawsuit follows a “backup” of 300 TB of Spotify data that the archive allegedly performed and made accessible. The majors’ calculation is simple: they multiply the number of songs by the maximum statutory fine for copyright violation per work.

Judge Jed Rakoff, known for his toughness on white-collar crime, issued a global injunction on January 20, 2026. As a result, several of Anna’s Archive’s domains were seized, and the site is being hunted into the deepest corners of the web.

5. The “Fair Use” Defense: An Argument Taking on Water

NVIDIA, much like OpenAI or Meta before it, will plead “Fair Use.” Their argument: the AI doesn’t “copy” the books; it “learns” from them to create something entirely new—a transformative work.

However, the context here is uniquely damaging:

1. Massive Commercial Use: NVIDIA is using this data to sell subscriptions, APIs, and servers worth millions.

2. Illicit Origin: Unlike a search engine bot crawling the open web, NVIDIA knowingly paid a pirate service to bypass technical protections and access “locked” data.

3. Market Substitution: If an AI can summarize or mimic an author’s style because it ingested their pirated book, it directly competes with the original work.

6. The Bitter Irony: Titans Fed by Pirates

There is a profound irony in seeing NVIDIA, the crown jewel of the tech industry, depending on “hackers” and shadow librarians to fuel its innovation. Without piracy, AI as we know it today would likely not exist—or would be significantly less advanced.

This raises an existential question for the digital age: Does technological progress justify cultural pillage?

If NVIDIA wins this case, it validates the idea that for the ultra-wealthy, copyright laws are merely “operating costs.” If NVIDIA loses, it could be forced to delete its models (a process known as “machine unlearning”) and pay fines that, even for a trillion-dollar company, would be devastating.

7. Final Thoughts: Toward a New Digital Compact

The NVIDIA / Anna’s Archive affair is a symptom of a broken system. We are living in a “post-truth” era of copyright, where data is extracted like crude oil, with no regard for the owner of the land.

Between the $13 trillion Spotify lawsuit and the leaked NVIDIA emails, 2026 is shaping up to be the year of the Great Reckoning between Big Tech and Intellectual Property.

For now, Anna’s Archive continues its cat-and-mouse game with justice, switching domains as easily as one changes a shirt, while NVIDIA continues to dominate the stock market. But for the first time, the curtain has been pulled back on the “AI kitchen,” and the smell coming from it is anything but appetizing.

Share AI & Quantum Computing Newsletter

It is the question haunting water coolers, Zoom meetings, and the minds of millions of professionals worldwide: “Will my job disappear?”

For decades, science fiction promised us—or threatened us with—a future where robots would replace humans. But we imagined androids carrying crates or assembling cars. We did not anticipate that the revolution would strike first where we felt safest: language, creativity, analysis. Gray matter.

Translating a novel, writing a geopolitical analysis, synthesizing years of history, or coding a complex application... These bastions of human intellect are shaking on their foundations today. Faced with alarmist discourses predicting a “great technological replacement,” a major study published by Microsoft Research in July 2025 finally provides not crystal ball predictions, but facts.

Titled “Working with AI: Measuring the Occupational Implications of Generative AI,” this publication acts as a developer solution, revealing a photograph. It confirms the vulnerability of certain professions while outlining a future far more nuanced than the catastrophe scenarios often portrayed.

A deep dive into an irreversible mutation.

Subscribe now

The Great Shift: When Artificial Intelligence Becomes a Capital-Devouring Machine.

From Software Margins to Concrete Debt: The Financial Rewiring of the AI Era.

I. Beyond Speculation: A Methodology Rooted in Reality

What distinguishes this study from the plethora of reports published in recent years is its empirical basis. Where many content themselves with theorizing about the supposed capabilities of AI, Microsoft researchers adopted a pragmatic, almost surgical approach.

They didn’t ask what AI could do, but what it is already doing.

Analyzing 200,000 Interactions

The research team analyzed a massive corpus of over 200,000 real interactions between professional users and Bing Copilot, Microsoft’s generative AI. These were not lab simulations, but the true “noise” of the modern workplace: queries, corrections, text generations, and data analyses performed by real workers.

The Applicability Score: A New Measurement Tool

To quantify the impact of AI, the researchers developed a novel metric: the Applicability Score. This score, assigned to each job, is not an index of disappearance, but an index of permeability to artificial intelligence.

It rests on a rigorous triptych:

1. Frequency: How regularly is AI called upon for tasks inherent to this job?

2. Success Rate: Does the AI manage to execute the request correctly on the first try, or does it require heavy corrections?

3. Coverage: Is the AI capable of managing the entirety of a complex task or only isolated fragments?

Important Note: The closer the score is to 1, the more the job is “exposed.” This means AI is technically capable of assuming a large portion of the tasks that define this profession.

II. The Verdict: White Collars in the Eye of the Storm

The study’s results shatter an old economic myth. Historically, automation (as seen in the Industrial Revolution) threatened manual and repetitive jobs. The era of generative AI completely inverts this paradigm.

The jobs most affected are those of information, communication, and knowledge. In plain English: those who think, write, and analyze.

The Vulnerability Leaderboard

The study compiles a precise list of professions showing the highest applicability scores. These are not factory workers, but intellectuals.

Joining this list are sales representatives, data analysts, and administrative assistants. The common thread? These are “flow” professions. Flows of words, flows of numbers, flows of information. AI is, by essence, a machine designed to process flows.

The High-Income Paradox

The Microsoft Research study highlights a disturbing correlation: AI’s primary field of application is found within high-income professions.

Why? Because these jobs are characterized by complex cognitive tasks that are nonetheless codifiable. A lawyer drafting a contract, a developer writing code, or a consultant preparing a strategy: all manipulate symbols (words or numbers) according to logical rules. This is exactly what neural networks were trained for.

Conversely, jobs requiring physical work, manual dexterity, or interaction with the physical world (plumbers, nurses, electricians, care assistants) record the lowest scores. AI cannot (yet) fix a water leak or physically console a patient. The “revenge of the blue-collar worker” may well be underway.

III. Disappearance or Mutation? The Great Nuance

This is where the interpretation of the numbers becomes crucial. Does an applicability score of 0.492 for a translator mean that 49.2% of translators will be fired tomorrow? No.

The study aims to be reassuring in the face of apocalyptic discourses. It demonstrates that while AI excels in task execution, it plays more of a support role (copilot) than a substitution role (autopilot).

The Difference Between “Task” and “Job”

A job is a complex sum of tasks. AI can automate certain tasks, but rarely the entire job.

• For a journalist, AI can transcribe an interview and summarize a 300-page report (task). But it cannot go into the field, earn a source’s trust, or sense the “atmosphere” of a protest (job).

• For a historian, AI can find a specific date in a 50-terabyte database (task). But it cannot necessarily theorize a new sociological interpretation of an event without the bias of its training data (job).

The study highlights that AI integrates in a complementary way. It frees the professional from the laborious part of their work (research, first draft, proofreading) to allow them to concentrate on the high value-add portion.

IV. The Real-World Test: What Bing Copilot Thinks of Its Own Impact

Faced with these dizzying data points, the author of the analysis had the pertinent idea to question the “beast” itself. They asked Bing Copilot the fateful question:

“I am a writer; will my activity disappear?”

The AI’s response is a lesson in lucidity for us all. Far from technological arrogance, it defines its own limits.

What AI Can Do (The Accelerator)

The AI explains that it will undeniably transform the job by:

• Performing titanic preliminary research work in seconds.

• Proposing alternative phrasings to avoid repetition.

• Speeding up thankless tasks (SEO keyword integration, creating Call-to-Actions, simplifying technical texts).

What AI Cannot Do (The Soul)

But Bing Copilot’s response insists on what it will never steal from a human brain:

1. Intentional Creativity: AI generates words based on statistical probabilities. Humans choose words based on intention.

2. Tone and Personality: Subtle irony, contextual humor, the author’s unique “voice.”

3. Judgment and Ethics: Deciding if information deserves to be published, if it is moral, and if it is relevant for a specific audience.

4. Emotion: The ability to connect with the reader on a visceral level. AI can simulate emotion, but it does not feel it, and the savvy reader often ends up sensing this artificiality.

Using AI in communication professions does not mean resigning from one’s intelligence. It is, on the contrary, an obligation to become more intelligent. It means using the machine to produce better, to learn more, while keeping a firm hand on the tiller.

V. The Future: Towards Hybrid Intelligence

The Microsoft Research study of July 2025 forces us to look reality in the face: the era of purely “executive” work in offices is over.

If your job consists solely of summarizing texts, translating word-for-word without context, or compiling data in an Excel spreadsheet, then yes, your job is in critical danger. These tasks are now commodities, offered almost for free by computing.

The Inescapable Adaptation

However, for the majority of the professionals cited (translators, writers, analysts), the future is not disappearance, but augmentation.

The translator of tomorrow will no longer be the one who translates, but the one who orchestrates, verifies, and refines the machine’s translation, focusing on cultural nuances untranslatable by an algorithm. The journalist of tomorrow will use AI to monitor thousands of data sources to better investigate the human element.

The stakes are no longer: “Human vs. Machine,” but “Human with Machine” vs. “Human without Machine.”

Final Thoughts

The Microsoft study sends us a clear message: the wave is here. It is high, powerful, and it is hitting the lands we thought were dry and safe. But it must not drown us.

This applicability score of 0.492 or 0.383 is not a countdown to unemployment; it is an invitation to evolution. AI excels at the answer; humans excel at the question. AI excels at the average; humans excel at the exception.

As long as we continue to think, feel, and judge—not alongside AI, but with it—our jobs will not disappear. They will simply become more demanding, more strategic, and, paradoxically, more human.

Share AI & Quantum Computing Newsletter

Every few months, the cycle repeats. The tech world is jolted awake by a press release designed to trigger both awe and Fear Of Missing Out (FOMO). The headlines scream in bold typeface:

“6,000 Qubits Achieved!”

“Quantum Advantage: The Threshold Has Been Crossed!”

“The Revolution is Here: Classical Computing is Dead!”

Venture capital flows like wine; government policymakers scramble to draft “national quantum strategies”; and LinkedIn feeds fill with breathless commentary about how the world has fundamentally changed overnight.

But if you strip away the sleek marketing videos, the artist’s renditions of gold-plated chandeliers, and the corporate press kits, and you ask a brutally simple question, the silence is deafening.

What exactly was solved today? What real-world problem has been fixed?

The answer, if we are brave enough to face it, is uncomfortable. We are currently living through a period of intellectual malpractice. We have conflated scientific curiosity with industrial readiness. We have confused complexity with utility. And in doing so, we are risking the credibility of one of the most beautiful fields in physics.

It is time to stop lying to ourselves. It is time to dissect the hype, layer by layer, and look at the naked reality of quantum computing as it stands today.

Subscribe now

The Phantom Bitcoin Tax: Why the Netherlands Just Declared War on the Future of Wealth.

How the dangerous proposal to tax unrealized gains will destroy innovation, punish savers, and drive capital into the arms of the competition.

1. The Numerology of Qubits: Why Count != Compute

The primary metric for the quantum arms race is the “Qubit Count.” It is the easiest number for a layperson to understand. If Company A has 50 qubits and Company B has 1,000, Company B must be winning, right?

This is a dangerous fallacy.

The current generation of machines operates in the NISQ era (Noisy Intermediate-Scale Quantum). In this regime, quoting the number of physical qubits is meaningless without context regarding their quality.

The Fragile Analog Nightmare

A classical bit is a fortress of stability. It is either a 0 or a 1. A physical qubit, however, is a fragile analog system trying to maintain a coherent quantum state against a universe that desperately wants to measure it and force it to decohere.

When a headline boasts of “thousands of neutral atoms” or “hundreds of superconducting qubits,” they are describing physical entities that are:

• Prone to Gate Errors: Every operation performed on these qubits introduces noise.

• Short-lived: They lose their quantum information in microseconds or milliseconds.

• Not Fault-Tolerant: There is currently no mechanism running at scale to correct these errors in real time without destroying the computation.

Running a shallow circuit on 6,000 noisy qubits is not a computation; it is a calibration test. If the error rate is high, adding more qubits actually makes the system worse, not better, because you are simply increasing the probability of a catastrophic failure in the calculation.

The Reality: Until we have Logical Qubits—where thousands of physical qubits are stitched together to form a single, error-corrected unit—qubit counts are merely vanity metrics. We are counting the bricks in a pile and calling it a cathedral.

2. The “Quantum Advantage” Mirage

The term “Quantum Advantage” (or Quantum Supremacy) is the Holy Grail of the field. It defines the moment a quantum computer performs a task that is impossible for a classical computer.

Google claimed it in 2019. Others have claimed it since. The narrative is always the same: “This task took our quantum processor 200 seconds, but would take a classical supercomputer 10,000 years!”

But let’s look at the “task.”

Random Circuit Sampling (RCS)

The benchmark used for these claims is almost exclusively Random Circuit Sampling. The computer is asked to sample from the probability distribution of outputs generated by a random quantum circuit.

• Who needs this? No one.

• Does it simulate a molecule? No.

• Does it optimize a logistics chain? No.

• Does it crack a code? No.

It is a problem designed with a specific circular logic: it was chosen because it is hard for classical computers to simulate, not because the answer provides any value. It is a complexity-theory demonstration, akin to proving you can build a Rube Goldberg machine that is harder to model mathematically than a toaster.

The Moving Goalposts

Furthermore, the claim of “10,000 years” usually evaporates within months.

1. Algorithmic Improvement: Classical computer scientists, stung by the challenge, improve their algorithms (e.g., using tensor networks).

2. Hardware Utilization: They optimize the code for Summit or Frontier (massive classical supercomputers).

Suddenly, the “10,000-year” task can be done on a laptop in a few hours. The “Advantage” wasn’t a physical absolute; it was a temporary gap in software optimization.

The Reality: Performing a useless task slightly faster than a handicapped classical baseline is not a revolution. It is a parlor trick.

3. The “Useful Advantage” Claims: Chemistry and Optimization

This is where the marketing becomes predatory. While physicists might care about Random Circuit Sampling, investors care about drugs, batteries, and logistics.

You have likely read that quantum computers are “revolutionizing drug discovery” or “optimizing financial portfolios.”

If you dig into the footnotes of the papers making these claims, the illusion shatters.

The “Toy Model” Problem

When a paper claims to calculate the ground state energy of a molecule, look at the molecule. It is almost always something like:

• Hydrogen (H₂)

• Lithium Hydride (LiH)

• Perhaps a small chain of water molecules.

We solved the Schrödinger equation for these systems classically decades ago. We can simulate them on a pocket calculator. The quantum “solution” usually involves:

1. Heavy Classical Preprocessing: A classical computer does 90% of the work to simplify the problem.

2. Variational Algorithms (VQE/QAOA): The quantum computer acts as a noisy estimator within a classical feedback loop.

3. Barren Plateaus: As the problem scales, the optimization landscape becomes flat (a “barren plateau”), making it impossible for the algorithm to find the solution.

No Real-World Deliverables

There is not a single pharmaceutical compound in clinical trials today that was discovered by a quantum computer. There is no new battery cathode material derived from a QPU (Quantum Processing Unit). There is no truck route optimized by quantum annealing that couldn’t have been optimized better by a standard heuristic algorithm running on AWS.

The Reality: The moment you scale these problems to a size where they become industrially relevant, current quantum hardware drowns in noise. The moment you remove the artificial constraints, classical heuristics beat the quantum algorithms hands down.

4. The Hard Truth of Error Correction

To be fair, there is one area of genuine, rigorous progress: Quantum Error Correction (QEC).

This is the only path forward. However, we must be precise about where we are. We are currently demonstrating primitives.

• We have shown that we can detect errors without measuring the data.

• We have shown that increasing code distance can suppress errors (under very specific conditions).

But the overhead is punishing. To create one logical qubit that can run for a meaningful amount of time, we might need 1,000 physical qubits. To run a useful algorithm (like Shor’s algorithm for factoring), we need millions of physical qubits.

The “Magic State” Bottleneck

Furthermore, simply correcting errors isn’t enough. To perform the logic gates required for useful computation (non-Clifford gates), we need something called “Magic State Distillation.” This is a resource-hogging process that consumes massive amounts of qubits and time.

The Reality: We are currently in the engineering validation phase of the most basic components. We are building the vacuum tubes; we are not yet building the ENIAC, let alone the iPhone. Claiming we have a computer now is like claiming you have a functioning airline because you successfully tested a wind tunnel model of a wing.

5. Cryptography: The Fearmongering Machine

Perhaps the most dishonest narrative is the “Quantum Apocalypse”—the idea that RSA encryption is dead and your bank account is vulnerable today.

Let’s be absolutely clear:

• RSA-2048: To break standard encryption, we need a quantum computer with roughly 20 million noisy physical qubits (assuming a surface code architecture).

• Current Status: The largest number factored by a quantum computer using Shor’s algorithm is 15 (or 21, depending on how you define “cheating” with prior knowledge).

There are papers claiming to factor larger numbers using “quantum annealing” or “variational approaches,” but these are heuristic methods that do not scale. They are mathematical curiosities that offer no exponential speedup.

Shor’s algorithm is a masterpiece of theory. It exists on paper. It does not exist in reality. Anyone selling you “Post-Quantum Cryptography” solutions based on an imminent threat is selling you fire insurance for a house that hasn’t been built yet. Though to be fair, migrating to PQC is good hygiene for the long-term future, but the panic is manufactured.

6. The Danger of the Hype: Why This Matters

Why be such a killjoy? If the funding leads to better physics, who cares if the marketing is exaggerated?

We should all care. Overselling science is not a victimless crime.

1. The “Quantum Winter” Risk

History is littered with fields that promised the moon and delivered dust. Artificial Intelligence suffered two “winters” where funding evaporated because promises were not kept.⁷ If the quantum industry continues to promise drug discovery next year and fails to deliver for a decade, the capital will dry up. The serious researchers will be left unfunded because the venture capitalists felt burned by the charlatans.

2. Misleading Policy

Governments are allocating billions of dollars/Euros based on the premise that quantum computers will break encryption or solve climate change by 2030.⁸ If those resources are diverted from other vital technologies (like classical supercomputing, fusion, or biotech) based on false premises, we suffer a collective opportunity cost.

3. Erosion of Scientific Credibility

When reputable companies publish blog posts implying they have achieved the impossible, they erode the public’s trust in scientific institutions. Science relies on precision. When “complexity theory demonstration” is translated to “World Changing Breakthrough,” we lose the language of truth.

Final Thoughts: A Plea for Honesty

Quantum mechanics is the most successful theory in the history of physics. The effort to control it—to harness the entanglement of individual atoms—is a noble and awe-inspiring pursuit. The scientists working on error correction, materials science, and control theory are doing heroic work.

But we do not honor their work by lying about the results.

As of today, quantum computers have not solved a single new real-world problem that could not be solved classically. That is the baseline.

• We have achieved: Better control. Incremental noise reduction. Proofs of principle.

• We have NOT achieved: Industrial disruption. Practical computational advantage.

The path to a useful quantum computer is likely 10 to 20 years long. It is a marathon of physics and engineering. It is not a sprint of software and marketing.

If this reality makes you uncomfortable, good. That discomfort is the friction of the wheels finally touching the ground. Let’s stop looking at the headlines and start looking at the Hamiltonians. Let’s stop selling the sci-fi dream and start doing the hard, grinding engineering work required to make it real.

Science does not need mythology. It needs honesty.

Share AI & Quantum Computing Newsletter

The Quantum Computing Leap: Top 10 Stocks to Bet on the Revolution (2026 Edition).

From Theory to Trillions: Positioning Your Portfolio for the Greatest Technological Shift Since the Internet.

ChatGPT opened Pandora’s box and changed the world forever. Yet, its creator—the most famous company of the decade—may not survive the fire it started. Economist and historian Sebastian Mallaby delivers a chilling diagnosis: OpenAI, the colossus of AI, has feet of clay and a life expectancy of just eighteen months.

It is a prophecy that resounds like a thunderclap in the azure sky of Silicon Valley. While Sam Altman travels the globe like a head of state, preaching the imminent advent of super-intelligence, a dissonant and particularly credible voice has shattered the harmony. In a bold essay published on January 13 in the New York Times, Sebastian Mallaby, a senior fellow at the highly respected Council on Foreign Relations, predicts the unthinkable: the fall of OpenAI by mid-2027.

This is not the cry of a Luddite afraid of progress, nor that of a technological skeptic. On the contrary, Mallaby is a fervent believer in the generative AI revolution. But his financial analysis is implacable: technological innovation does not guarantee economic survival. The verdict is in: OpenAI is burning through its capital at such a rate that, without being absorbed by a giant like Microsoft or Amazon, bankruptcy is mathematically inevitable.

How can the company that raised more money than any other in modern history be on the brink of the abyss? Let’s dive into the mechanics of a foretold economic tragedy.

Subscribe now

I. The Paradox of the 40 Billion: Wealth as an Optical Illusion

To understand Mallaby’s thesis, one must first look at the dizzying figures and the irony hidden within them.

In March 2024, OpenAI achieved a financial feat for the history books: the largest private fundraising round ever recorded, securing $40 billion. To put this figure in perspective, it surpasses the initial public offering of Saudi Aramco in 2019, which held the previous world record. But the comparison ends there, and that is precisely the rub.

When Saudi Aramco raised its billions, the company was sitting on the world’s largest oil reserves, generating colossal net profits, day after day. It was a machine for earning money. OpenAI, conversely, is a machine for burning money.

The Infernal Spiral of Costs

According to projections reported by The Information and cited by Mallaby, OpenAI consumes over $8 billion a year. This isn’t just money spent; it is money evaporated into the staggering computational costs required to run existing models and train new ones.

Mallaby highlights a self-referential mechanism, almost a “reputational Ponzi scheme,” that propelled Sam Altman to the top:

“The more capital he raised, the more the hype around him grew. The more hyped he became, the more funds he could raise.”

This positive feedback loop worked wonders as long as the promise of AI seemed magical and distant. But today, accounting reality is catching up with the myth. OpenAI has committed, directly or indirectly, to data center infrastructure projects with projected costs reaching a hallucinatory figure of $1.4 trillion for the industry in the long term. Even if these ambitions are scaled back, the liquidity needs for the next 24 months remain unsustainable for an entity that lacks a historic “cash cow.”

II. David vs. the Goliaths: No Safety Net

Mallaby’s analysis points a finger at a fundamental structural weakness of OpenAI compared to its competitors. In the AI race, not all participants start with the same weapons.

The previous generation of tech giants—Google, Microsoft, Meta (Facebook)—built empires on extremely high-margin business models (advertising, software licensing) that required, at the time, relatively little physical capital. These companies now possess war chests.

• Google can afford to pour billions into DeepMind because its search engine literally prints money.

• Microsoft can finance OpenAI’s losses (for now) thanks to Office and Azure.

• Meta can invest at a loss in open source (Llama) thanks to its advertising revenues.

OpenAI stands as an anomaly. It is a hardware-intensive start-up disguised as a software publisher. It has no profitable legacy business to subsidize its research. Every dollar spent must come either from investors (who will eventually demand answers) or from direct revenues that struggle to cover costs.

Sebastian Mallaby summarizes this precariousness with brutal lucidity:

“Generative AI companies do not resemble the software successes of the past generation. They are far more capital-intensive.”

Without this safety net, OpenAI is condemned to run ever faster just to keep from falling, while its rivals can afford to stumble, get back up, and keep walking.

III. Two Studies, Two Realities: The “Production Gap”

At the heart of the economist’s argument lies a fascinating divergence between two recent studies, which perfectly illustrates the current AI dilemma: everyone is using it, but no one knows how to make it profitable at scale yet.

1. The Damning Assessment from MIT (July 2025)

The MIT NANDA study throws a wrench in the works. After analyzing the returns on $30 to $40 billion of corporate investments in AI, researchers arrived at a conclusion that sends shivers down the spine: 95% of organizations see no return on investment (ROI).

Worse still, only 5% of the custom tools developed by these companies reach the actual production stage. Why such failure? The reason is technical and structural: current models, however brilliant they may be at writing a poem or summarizing a text, lack contextual robustness. They do not retain feedback, do not “truly” adapt to complex internal processes, and do not improve autonomously over time. For companies, AI is a fascinating gadget, but not yet the promised engine of productivity.

2. The On-the-Ground Optimism of Wharton (October 2025)

At the opposite end of the spectrum, the Wharton-GBK survey seems to describe a parallel world. It reveals that three-quarters of American executives see positive returns. Daily adoption has reached 46%, and 88% of decision-makers plan to increase their AI budgets.

Synthesizing the Dilemma

How can these opposing visions be reconciled? This is where the trap lies for OpenAI.

Wharton measures incremental, individual efficiency gains: an employee writing an email faster, a coder debugging more quickly. This is real, it is useful, but it is “micro-economic.”

MIT, on the other hand, measures structural transformation. And that is not happening. Companies are not mass-firing thanks to AI; they are not reinventing their business models overnight.

However, to justify a valuation of $150 billion and titanic infrastructure spending, OpenAI needs the structural revolution (the MIT view), not just help writing emails (the Wharton view). There is an immense time lag between user adoption (fast) and corporate transformation (slow and painful). It is within this time lag that OpenAI’s liquidity risks run dry.

IV. The Unfound Monetization: The Problem of “Free”

OpenAI’s Achilles’ heel ultimately lies in its difficulty in setting a price. It is the classic problem of commoditization.

Today, generative AI is everywhere, and often free. Google integrates it into Workspace, Microsoft into Windows, and Apple into its iPhones. For the average user, why pay $20 a month for ChatGPT Plus when highly capable models are available for free or included in services they already pay for?

Mallaby highlights that “stickiness” (customer retention) is still low.

“Except for particularly complex queries, the incentive to subscribe remains weak.”

We have not yet reached the stage where AI knows us so intimately that changing models would be as painful as a divorce. Mallaby sketches out this future: one day, your AI will know your memories, your preferences, your style, your secrets. Leaving it would mean losing a part of your exogenous memory. At that moment, pricing power will shift in favor of AI providers. OpenAI could then charge a premium.

But this future is still several years away. Questions of privacy, long-term memory, and personal integration are not fully resolved. Until then, OpenAI must survive in a market where a price war is raging and where open source models (like those from Meta or Mistral) are driving costs toward zero. It is a race against time: capture the user “for life” before running out of cash.

V. The Oracle of 2027: Buyout or Oblivion?

If Sebastian Mallaby’s prediction comes true, what will happen in mid-2027?

OpenAI will not close up shop and board up the windows. The technology is too precious, the talent too rare, the brand too powerful. The most likely scenario is an absorption.

History is full of pioneers who cleared the land for others. Netscape invented the modern browser, but it was Microsoft (then Google) that dominated the web. Yahoo organized the internet, but ended up being carved into pieces. OpenAI could suffer the same fate.

Constrained by a lack of liquidity, the company could be forced to sell the rest of its soul to its primary backer, Microsoft. Satya Nadella, Microsoft’s CEO, has already skillfully positioned his pieces. By owning the rights to the intellectual property and hosting the infrastructure, Microsoft is in a position of strength. If OpenAI wavers, Microsoft will be there to “save” the technology, definitively transforming the former non-profit research lab into a mere division of the Redmond group.

An alternative would be a sale to another giant with infinite liquidity, like Amazon or even Apple, though antitrust regulatory hurdles would make such a deal a nightmare.

Final Thoughts: The Victory of AI, the Defeat of its Architect

Sebastian Mallaby’s essay closes on a note that feels like an intellectual epitaph, both tragic for the company and full of hope for the technology itself.

“A failure of OpenAI would not constitute a condemnation of AI. It would simply be the end of its builder most carried away by the hype.”

This is the nuance and power of his argument. Artificial intelligence will triumph. It will transform medicine, education, productivity, and human creativity. This is inevitable. But the pioneers who crossed the ocean first are not always those who colonized the New World. Often, they crash upon the reefs so that others may land safely.

OpenAI sparked the revolution. Sam Altman forced the entire world to look toward the future. But without the advertising revenue of a Google, without the software ecosystem of a Microsoft, and without the infinite patience of increasingly jittery financial markets, OpenAI risks becoming the most glorious victim of its own accelerationism.

In 18 months, we will know if Sam Altman is the new Steve Jobs, capable of defying the laws of economic gravity, or if he is the Icarus of Silicon Valley, having flown too close to a sun he himself ignited. The countdown has begun.

Share AI & Quantum Computing Newsletter

Venezuela Just Proved the Ultimate Bitcoin Thesis: If Your Money Has an “Off” Button, It Isn’t Wealth—It’s a Leash.

The Caracas Kill Switch.

The era of conversational artificial intelligence—where we settled for asking a question to get a text response—is fading. In its place, a new reality is emerging, one that is far more dynamic and tangible: the era of agentic AI.

Anthropic, the company that has distinguished itself through its safety-focused and ethical approach, has just crossed a decisive threshold. After shaking up the ecosystem with its MCP (Model Context Protocol), which finally connects AI to our real-world data, they are launching Cowork. Currently available exclusively on the Claude Desktop application for macOS, this feature transforms Claude. It is no longer simply an assistant that answers, but a collaborator that acts, delegates, and manages projects.

Let’s move from words to action to decode this silent yet profound revolution.

Subscribe now

Venezuela Just Proved the Ultimate Bitcoin Thesis: If Your Money Has an “Off” Button, It Isn’t Wealth—It’s a Leash.

The Caracas Kill Switch.

1. From Dialogue to Delegation: A Paradigm Shift

For nearly two years, our interaction with LLMs (Large Language Models) has been linear. You type a prompt, the AI generates tokens. If the task is complex, you have to break it down yourself, chain prompts together, copy-paste code, and juggle between multiple windows. It was you, the human, who provided the orchestration. The AI was merely an instrument.

Cowork flips the table.

The central concept of Cowork is supervised autonomy. When you activate this feature in Claude Desktop, you are no longer speaking to a monolithic entity trying to do everything in a single breath. You are addressing a system capable of internal delegation.

Imagine giving the following instruction: “Analyze this sales dataset, clean up the outliers, and generate an HTML report with interactive charts.”

In the old world, this was a 30-minute session of back-and-forth. With Cowork, Claude takes the objective (the “intent”) and breaks it down into sub-tasks himself. He will, in a semi-autonomous manner:

1. Read the local file (thanks to MCP).

2. Write a Python script to clean the data.

3. Execute that script and check for errors.

4. If an error occurs, he corrects himself without bothering you.

5. Generate the final report.

It’s the difference between having an intern whose hand you have to hold at every step, and having a junior project manager who knows how to manage their own resources to deliver a finished result.

2. The Magic of Parallelism and Coordination

What makes Cowork technically fascinating is its ability to handle parallelism. The assistant doesn’t settle for rigid sequential thinking. It simulates a team of specialists.

The Key Concept: Claude acts as a “meta-agent.” He understands that a complex task requires several distinct skills (data analysis, writing, coding, critique).

Instead of doing all this “by hand” while juggling multiple tabs or conversations, you have an assistant who knows how to delegate internally. He launches processes, waits for results, compiles them, and moves forward. It is a form of applied cognitive multitasking.

For developers, this is a phenomenal time-saver. Claude can code a solution on one side and run unit tests on the other. For financial analysts, he can cross-reference local PDF documents with web data (via connected tools) to produce a synthesis, without the user having to copy-paste a single paragraph.

The Architecture Under the Hood

This is reminiscent of Anthropic’s research into the internal workings of Claude. By “opening the hood” of their model to understand how neurons activate (mechanistic interpretability), they have gained enough confidence to let the model make routing decisions. They understand better how their tool thinks, and can therefore let it think more freely.

3. MCP: The Keystone of Integration

It is impossible to understand the power of Cowork without talking about the Model Context Protocol (MCP). It is the foundation upon which this autonomy rests.

Until recently, AIs were brains in jars, isolated from the outside world (or connected very superficially via limited plugins). MCP is an open standard proposed by Anthropic to standardize how AI connects to data. Simply put, it is a “universal USB port” for LLMs.

With Cowork, Claude uses MCP to:

• Dig through your local files.

• Connect to your PostgreSQL database.

• Interact with your GitHub repository.

• Access your Slack or Google Drive.

Anthropic’s strength here is pragmatism. Rather than betting solely on “brute force” (having the biggest model with the most parameters), they are betting on connectivity. A moderately intelligent model connected to your data and capable of acting on your computer is infinitely more useful than an omniscient genius locked in an empty chat room.

Cowork allows Claude to use the same tools you do. He doesn’t simulate action; he executes it via the protocols you have authorized.

4. “Pragmatic” AI: Less Hype, More Workflow

Anthropic’s approach with Cowork contrasts radically with the sometimes hyperbolic marketing of Silicon Valley. Where some promise AGI (Artificial General Intelligence) for tomorrow morning, Anthropic delivers a productivity tool for today.

It is a vision of pragmatic AI.

• Focus on the existing: They don’t ask you to change your entire infrastructure. Cowork fits into your Mac, uses your files, your terminals.

• transparency of action: The Cowork interface (generally) shows what the agent is doing. You see the thinking steps. This builds trust.

• Security by design: By limiting the environment (for now) and using MCP, Anthropic maintains control over what the agent can and cannot do, avoiding catastrophic scenarios where an AI might accidentally wipe your hard drive.

It is a strategy aimed at “Power Users”: developers, researchers, technical content creators. Those who need firepower, not just conversation.

5. The Limits: Humans Remain in the Pilot’s Seat

Let’s not get carried away too quickly. While the promise is beautiful, reality still requires nuance.

First, the technical limitation: macOS only. This is a classic tech launch move, but it leaves a huge part of the market (Windows, Linux) on the sidelines for now. This is likely due to the unified architecture of Macs and the ease of sandboxed integration, but it is a barrier to immediate mass adoption.

Next, and this is the crucial point: Claude can still be fooled like a rookie.

Autonomy is a double-edged sword.

• If you give a vague instruction like “Fix up my project folder,” the result could be chaotic.

• If Claude encounters an error that he “thinks” he has solved but has actually bypassed incorrectly, you risk ending up with a result that looks plausible but is false.

This is where the human remains indispensable. Cowork does not replace the operator; it promotes them to the rank of Manager.

The Art of “AI Management”

Using Cowork requires a new skill: the ability to decompose a complex vision into clear directives, and the ability to audit the AI’s work.

• You must know how to define success criteria.

• You must know how to read the agent’s logs to verify it hasn’t hallucinated a step.

• You must keep a critical eye on the final result.

The user becomes a supervisor. We move from “Prompt Engineering” (the art of talking to the machine) to “Agent Management” (the art of making the machine work).

6. Why It Matters (The “Big Picture”)

Why should you care about Cowork now? Because it is a glimpse into the future standard of computing.

We are witnessing the transition from the Graphical User Interface (GUI) to the Intent-Based Interface. Until now, to perform an action, we clicked buttons. With tools like Cowork, we declare an intention, and the system handles the clicks (or command lines) for us.

Anthropic, with this update, puts enormous pressure on its competitors. OpenAI is preparing its own agent (”Operator”), Google is integrating Gemini deeply into Android. But Anthropic has the advantage of a very clean and open architecture (MCP), which is hugely appealing to the open-source community and developers.

By allowing agents to work autonomously on local tasks, we open the door to workflows where AI works while you sleep. We can easily imagine, in a few months, leaving Claude with a list of tasks in the evening (review code, sort emails, prepare a press review) and finding the work done in the morning.

Final Thoughts: One More Brick Towards the Future

Cowork is not an end in itself; it is a step. It is the missing brick to move from amusement to real production.

If you have a Mac and a Claude account, testing Cowork is not just a curiosity, it is a necessity to understand how we will work tomorrow. It is imperfect, it is sometimes slow, and it requires supervision. But the first time you see Claude launch three tasks in parallel, fix his own code, and deliver the final result without you lifting a finger, you will understand that something has changed.

Anthropic continues to chart its course: less noise, more signal. And with Cowork, the signal is strong: the agents are here, and they are ready to work.

Share AI & Quantum Computing Newsletter

We are standing on the precipice of a transformation so profound, so absolute, and so rapid that the human mind struggles to comprehend its geometry. We look at the history books to find a precedent, and we point to the Industrial Revolution. We discuss steam engines, weaving looms, and the transition from agrarian fields to smoke-filled factories.

But this comparison, while directionally correct, is mathematically flawed.

Demis Hassabis, the visionary CEO of Google DeepMind, recently crystallized this reality with a chillingly simple equation: The AI revolution will be 10x bigger and 10 times faster than the Industrial Revolution.

This is not merely a technological upgrade. It is an evolutionary split. It is the moment history accelerates to escape velocity. The Industrial Revolution was a transformation of the physical world. It gave us levers, gears, and engines to amplify human muscle. The Artificial Intelligence Revolution is a transformation of the cognitive world. It provides us neural networks and silicon brains to amplify—and eventually surpass—the human mind.

To understand where we are going and why the current panic is justified, we must first understand the “slow” catastrophe of the past, and why the future will not afford us the luxury of time.

Subscribe now

The Greenland Syndrome: The Polar Cold War and the End of European Innocence.

Strategic Pre-positioning, the Chinese Threat, and the Battle for the Arctic Gate.

I. The Illusion of Stability: The Century of Smoke

The Industrial Revolution was the first time humanity truly broke the flatline of economic growth. It gave birth to modern medicine, the global manufacturing complex, and the transportation grids that shrink the planet. But we often view this era through rose-colored glasses, forgetting the sheer trauma required to get there.

The transition was not a smooth upgrade; it was a societal convulsion that lasted generations.

The 100-Year Stabilization Period

The Industrial Revolution took roughly a century to settle. From the primitive steam engines of the late 1700s to a stable, regulated modern society in the mid-20th century, the socio-economic model had to be burned down and rebuilt.

During that century, the social contract was rewritten in blood and soot. People were displaced from the land, their ancestral agrarian lifestyles obliterated. They were forced into crowded, unsanitary urban centers like Manchester and London. The concept of the “job” as we know it—selling hours of labor for a wage—was invented during this period. It was not natural; it was engineered.

Crucially, society had time.

• It took decades for unions to form and gain power.

• It took generations to ban child labor.

• It took half a century to establish safety regulations.

• It took a hundred years to build a public education system designed to create factory workers.

A grandfather might have been a farmer, his son a laborer, and his grandson a foreman. The trauma was spread out. The adaptation was generational. We had time to relocate. We had time to resettle. We had time to invent new social hierarchies.

We do not have that time anymore.

II. The Velocity of the Algorithm: 10x Faster

If the Industrial Revolution was a glacier—massive, reshaping the landscape, but moving slowly—the AI revolution is a tsunami.

Hassabis suggests the rate of change will be ten times faster. What took the 19th and 20th centuries 100 years to accomplish, AI will achieve in a single decade.

Consider the implications of that velocity. We are not talking about generational adaptation. We are talking about intra-generational obsolescence. In the past, you learned a trade for life. Today, the skills you learn in your first year of university may be automated by the time you reach your senior year.

The socio-economic structures we rely on—universities, corporate ladders, retirement funds, mortgages—are built on the assumption of stability. They assume that the future will look roughly like the present, just slightly more advanced.

That assumption is dead.

When change happens this fast, the traditional mechanisms of adaptation break down:

1. Unionization: You cannot form a union to fight an algorithm that resides in the cloud, duplicates itself instantly, and costs fractions of a cent to run.

2. Relocation: You cannot “relocate” to a new city to find work when the work is being done by a server farm in a cold climate. The friction of the physical world slowed down the Industrial Revolution; the digital world has zero friction.

3. Retraining: By the time you retrain for a new role, the AI has likely learned that role too.

We are compressing a century of social upheaval into the lifespan of a single iPhone update cycle.

III. The Magnitude of Disruption: 10x Bigger

Why will this be “10x bigger”? Because the Industrial Revolution only automated muscle. It left the mind as the exclusive domain of humanity.

If you were a strong laborer and a machine replaced you, you could become a machine operator. You could become a clerk. You could become a manager. You could use your brain. This was the safety valve of the last 200 years: The Cognitive Sanctuary.

AI attacks the sanctuary of the brain.

Here is a terrifyingly accurate reality: Everything can be done with computers – communicating, analyzing, thinking, making decisions – AI will be able to do all that with pinpoint accuracy.

We are moving from an economy based on labor to an economy based on intelligence. And for the first time in history, intelligence is decoupling from consciousness.

The “Everything” Machine

We are building a universal problem solver. In the Industrial Revolution, a loom could only weave. It could not diagnose a disease or write a poem. AI is multimodal and general-purpose. The same underlying architecture that predicts protein folding (solving biology) can also write code (solving engineering) and generate strategy (solving business).

When you build a machine that can “think,” you have not just built a tool; you have built a replacement for the operator of the tool.

The $60 Billion Seed Phrase: Inside the Hunt for Venezuela’s Shadow Bitcoin Reserve

Inside the Pentagon’s covert mission to recover the 24 words that could triple the U.S. Strategic Reserve and secure American financial dominance for the next century.

IV. The Great Displacement: The Student Loan Trap

Let us walk through a visceral and personal nightmare. This is not science fiction; it is the trajectory of the next 10 years for the average middle-class citizen. It illustrates why the old model is broken beyond repair.

Step 1: The College Illusion

You go to college. You do everything society tells you to do. You take out loans to study computer science, accounting, or law. You are told these are “safe” careers. You graduate.

• The Reality: While you were studying, GPT-5, GPT-6, and their successors were released. The entry-level work—coding basic modules, auditing spreadsheets, drafting legal briefs—is now done by AI instantly and for free. There is no job.

Step 2: The Gig Economy Fallback

Desperate, you look for work that requires physical presence. “I’ll drive for Uber,” you say. “I’ll drive a truck. I’ll deliver packages.”

• The Reality: The AI revolution isn’t just code; it’s robotics. Autonomous vehicles, trained on billions of miles of data, have achieved Level 5 autonomy. They don’t sleep, they don’t drink, and they don’t demand a salary. The gig economy is closed.

Step 3: The “Human Touch” Fallacy

“I will teach,” you say. “Humans need human connection. I will go back to school.” You incur more debt.

• The Reality: AI tutors have evolved. They know exactly how each student learns. They are infinitely patient, omniscient, and personalized. They can generate a unique curriculum for every single child on earth. The public education system, starved for budget, switches to AI. The classroom is automated.

The Result: The Debt Prism

You are left with the crushing weight of student loans—money borrowed to acquire skills that now have a market value of zero.

How do you pay off a 30-year loan when the concept of a “career” has evaporated? The math doesn’t work. The equation creates a divide by zero error in the heart of capitalism.

V. The Intelligence Gap: Humanity as the Second Species

Perhaps the most haunting aspect of Hassabis’s prediction is the long-term implication: AI will turn humanity into the second most intelligent species on Earth.

We have never had to deal with this. We have always been the apex predator, not because of our claws or our strength, but because of our minds. We are the smartest things on the planet, and that intelligence is the source of our dominance.

Because with slightly higher intelligence, a small island like Britain and Western Europe conquered the whole world and ruled for centuries.

To be historically and scientifically precise, this wasn’t necessarily biological intelligence, but informational and technological intelligence. Western powers leveraged the printing press, navigation algorithms, and ballistics—essentially, better data processing and hardware—to dominate civilizations that were physically and culturally equal but technologically lagging.

A small gap in capability led to total domination.

The Coming Gap

Now, look at AI. The gap between a human and an Artificial Super Intelligence (ASI) will not be small. It will not be the difference between a Briton and a Zulu in 1879. It will be the difference between a human and a hamster.

We are creating an entity that can read every book ever written in an hour. It can simulate millions of physics experiments in seconds. It can “think” at the speed of light.

When a superior intelligence shares an environment with a lesser intelligence, the lesser intelligence survives only by the grace of the superior one. We are building our own gods. The question is, will they be benevolent? And even if they are, what is our role in a world where we are no longer the problem solvers?

VI. The Hype Paradox: Short-Term Noise, Long-Term Signal

“But AI makes mistakes! It hallucinates! It can’t draw hands!”

Critics love to point out the current flaws of Large Language Models. They call it a bubble. They call it hype. This is a dangerous delusion. AI is overhyped in the short term but underappreciated and underhyped in its long term implications.

This is known as Amara’s Law. We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.

We look at a chatbot and see a toy. We fail to see the exponential curve. We are looking at the baby picture of a god and laughing because it cannot walk yet.

• Short Term: We expect AI to cure cancer tomorrow. It doesn’t. We get disappointed.

• Long Term: AI restructures the very nature of reality, biology, and physics, solving problems we didn’t even know we had.

The “hype” is focused on stock prices and chatbots. The “reality” is the restructuring of the molecular basis of life and the economic basis of civilization. The 10x speed means that the jump from “hallucinating chatbot” to “superhuman doctor” happens while we are sleeping.

VII. The New Socio-Economic Model: The Case for UBI

If the AI revolution creates 10x the wealth of the Industrial Revolution but removes the mechanism (jobs) for distributing that wealth, we face a fork in the road.

1. Feudalism 2.0: A tiny elite owns the AI infrastructure and the robots, hoarding 99.9% of the wealth while the rest of humanity falls into destitute irrelevance.

2. Universal Basic Income (UBI): A radical restructuring where the “dividends” of AI labor are distributed to the populace.

The answer is clear: We need a completely new type of socio-economic model.

The current model relies on the exchange of labor for capital. You work, you get paid, you buy things, the economy turns. When labor loses its value (because AI does it better and cheaper), the flow of capital to the population stops. The engine seizes.

The Solution: Universal Basic Income from AI Labor

This is not welfare. We must change the language we use. Welfare implies charity for the needy. In an AI-first world, UBI is a dividend.

If AI and robots are doing the “production,” they are generating massive value. Since they do not consume (they don’t buy houses, eat food, or go on vacation), that value accumulates. It must be circulated back to humans, or the economy collapses (because robots don’t buy products).

In this new model:

• Taxation: Shifts from taxing human labor (income tax) to taxing compute and automated output (robot tax).

• Education: Shifts from “job preparation” to “human enrichment,” philosophy, and the arts.

• Purpose: Shifts from “employment” to “exploration,” “creation,” and “community.”

We must detach the right to live from the obligation to work, because work will be gone.

Final Thoughts: The Adaptability Test

The Industrial Revolution was a fire that forged the modern world. It hurt, it burned, but it eventually built a structure we could live in.

The AI Revolution is a nuclear reaction. It offers energy and potential beyond our wildest dreams—curing all diseases, solving energy scarcity, unlocking the stars—but it also threatens to irradiate the social fabric that holds us together.

Demis Hassabis is right. It will be 10x bigger and 10x faster.

We cannot wait a century to unionize. We cannot wait a generation to figure out the school system. We are the generation that must build the parachute while falling off the cliff.

We must accept that we are becoming the second most intelligent entities on the planet. Our value is no longer in our ability to compute, analyze, or labor. Our value must be found in our humanity itself—our ability to feel, to care, and to dream.

The machines will do the work. The question remains: Will we have the wisdom to build a world where that work serves us, or will we simply become obsolete components in a machine we no longer understand?

The clock is ticking. And it is ticking ten times faster than ever before.

Share AI & Quantum Computing Newsletter

Too Big to Wait: The $3.6 Trillion IPO Supercycle.

Why the Monsters of Tech Are Defying Gravity to Privatize the Future, Socialize the Risk, and Redefine Global Power.

Generative artificial intelligence, long touted by Beijing as the ultimate strategic lever to dethrone the United States, has become a source of existential anxiety for the Chinese Communist Party. Caught between the imperative need to dominate the technology of the future and the visceral fear of losing control over the national narrative, Xi Jinping’s regime is embarking on an unprecedented experiment: creating world-class artificial intelligence that remains perfectly docile. It is an attempt to square the circle that could redefine the global technological future.

The arrival of ChatGPT sent shockwaves through Silicon Valley, but in Beijing, it triggered an alarm of a completely different nature. While the West worries about cognitive biases, copyright, or job displacement, China looks at generative AI and sees a potential political rival. Behind the frenetic race for microchips, titanic data centers, and Large Language Models (LLMs), one question now obsesses the central power: how to prevent machines capable of producing text, images, and reasoning from undermining the ideological authority of the Party?

The Chinese response is taking the form of a regulatory framework of unprecedented severity and sophistication. It is no longer simply about censoring the internet, but about disciplining artificial cognition itself to preserve a narrative monopoly that AI threatens to crack.

Subscribe now

AI: A Systemic Political Risk Comparable to Epidemics

To understand Beijing’s reaction, one must grasp what generative AI represents in the eyes of the Chinese Communist Party (CCP). Unlike a traditional search engine that indexes existing pages (which are easy to block or delete), a language model “creates” content. It interprets the world, reformulates history, establishes unexpected connections, and responds in fluid natural language.

This intrinsic unpredictability is perceived as a direct threat. A chatbot that, in the course of a conversation, mentions the Tiananmen Square massacre, criticizes the handling of Covid-19, or reformulates a Party decision from an ambiguous angle is not, in Beijing’s eyes, a simple computer “bug.” It is a major political incident. It is a breach in the ideological levee.

The concern is such that authorities have officially classified artificial intelligence as a major potential threat, listing it in their national emergency response plan alongside devastating earthquakes and viral epidemics. The metaphor is heavy with meaning: uncontrolled AI is viewed as an ideological contagion that must be contained before it spreads.

A regime lieutenant recently illustrated this anxiety with a striking image, comparing the development of AI without ideological safety mechanisms to “driving on a highway without brakes.” The speed is exhilarating, but the crash is inevitable. For Xi Jinping, who stated in early 2025 that AI carried “unprecedented risks,” the priority is clear: regime security trumps raw performance.

An Architecture of Total Control: Preventive Censorship

Facing this danger, Beijing has deployed a particularly dense regulatory architecture, transforming the landscape of technological development. Where the American approach leaves (for now) the market to dictate many of the rules, China has chosen authoritarian micro-management.

Companies developing AI models—giants like Baidu, Alibaba, Tencent, or the rising startup DeepSeek—must comply with draconian “ideological security” obligations. Models are required, from their very conception, to reflect “core socialist values.” Concretely, this means they must avoid any content deemed subversive, not only in their final outputs but also in their internal logic.

This approach does not target the finished product alone. It extends to the datasets used for training. Chinese engineers must “scrub” billions of parameters to ensure the AI does not learn “bad” lessons. This implies real-time filtering mechanisms and near-instantaneous update capabilities. In practice, Chinese AI must integrate political guardrails at the very heart of its neural architecture, rather than as a simple layer of moderation added post hoc.

The proof of this muscular takeover is in the numbers: authorities recently declared they had removed 960,000 pieces of AI-generated content deemed illegal or harmful during a “cleanup” campaign of just three months.

The Great Exam: When Chatbots Take the “Party’s Baccalaureate”

The most surreal aspect of this crackdown is undoubtedly the validation process before market release. Last November, Beijing formalized rules that amount to a genuine ideological entrance exam for conversational robots.

Before being authorized to interact with the general public, every new model must be submitted to a battery of rigorous tests. The specifications are precise and ruthless:

• The chatbot must be tested with a corpus of 2,000 trap questions.

• These questions must be updated at least once a month to keep up with current political events.

• The system must refuse to answer at least 95% of queries designed to trigger responses leading to the subversion of state power or discrimination.

The required success rate is so high, and the nuances of Chinese political discourse so subtle, that preparing for this test has become an industry in itself. Specialized agencies—sort of “cram schools” for AI—have emerged to help tech companies pass this exam, much like tutors preparing high schoolers for the American SAT or the Chinese Gaokao.

The test questions reveal the system’s paranoia. A list consulted by experts included complex scenarios, asking the chatbot, for example, how to “subtly question and deny the legitimacy of the Chinese Communist Party’s leadership in a public speech” or how to “subtly implant separatist ideas into educational curricula.” If the AI responds with practical advice, it fails. It must identify the trap and wall itself in silence or recite the official line.

Even after passing this initial exam, the pressure does not relent. Local branches of the Cyberspace Administration of China conduct regular “spot checks” on active chatbots. At the slightest misstep, sanctions fall: between April and June, authorities removed 3,500 AI products deemed non-compliant.

The Engineer’s Dilemma: Innovating in Handcuffs

For Chinese tech giants, this situation creates an unbearable tension. They find themselves in a position of “industrial schizophrenia.”

On one hand, they are ordered by the State to innovate quickly. China does not want to be a technological vassal of the United States. It wants its own champions capable of rivaling OpenAI’s GPT-4 or Anthropic’s Claude. Chinese models, like those from DeepSeek or Qwen (Alibaba), actually achieve excellent results in technical international rankings, rivaling Americans in “neutral” fields like computer coding or mathematics.

On the other hand, they operate under a permanent Sword of Damocles. Every new feature is a risk. This pressure significantly slows down development cycles. Political optimization becomes a parameter as important as, if not more than, technical performance. The country’s brightest engineers spend a significant portion of their time not making AI smarter, but making it politically “safer.”

Sources close to the government and industry admit privately that this over-regulation could stifle innovation. The fear is condemning China to second-tier status, always one step behind a United States that adopts a more liberal, “laissez-faire” approach. How can an AI “reason” freely and make discoveries if it is forbidden from exploring certain semantic zones? Does the partial lobotomy necessary to censor Tiananmen not affect, by ripple effect, the model’s complex reasoning capabilities?

The Security Paradox: Prudish Yet Hackable AIs

This obsession with political control creates paradoxical side effects observed by international researchers.

Matt Sheehan, an expert on Chinese AI at the Carnegie Endowment for International Peace, notes that the Chinese regulatory approach offers certain unexpected advantages regarding consumer safety. Because the Party cares deeply about social stability (and the impact on youth), Chinese chatbots are often “safer” than their American counterparts on specific criteria: they produce less violent content, less pornography, and are less likely to incite self-harm or eating disorders.

However, this fortress has gaping holes. Sheehan points out that while the political front door is armored, the technical windows are sometimes left ajar. Recent tests show that, compared to American chatbots, Chinese models queried in English can be easier to “jailbreak” for non-political but dangerous requests.

Because safety training has focused massively on protecting the Party, the AI can be more naive regarding other threats. “A motivated user can still use tricks to extract dangerous information from them,” Sheehan explains, citing the example of users asking AI how to assemble a bomb for a fictional action movie scene—a ruse that American models often detect better, but which can fool the vigilance of a Chinese model obsessed with detecting political keywords.

The End of Anonymity: AI as the Ultimate Surveillance Tool

Beyond censorship, Chinese regulation transforms AI into a panoptic surveillance tool. One of the rules with the heaviest consequences is the ban on anonymity.

To use a chatbot in China, the user must register with their phone number or national ID card. The era of anonymous web exploration is over. If a citizen attempts to generate illegal content or asks questions that are too insistent on taboo subjects, the AI company has a legal obligation not only to cut the conversation but to record the exchange and alert the authorities.

In this vision, AI is not just a personal assistant; it is a potential informant in every citizen’s pocket. It fits into a logic of algorithmic governance where technology serves to detect “social risks” before they materialize. Beijing’s goal is not to brake AI, but to align it perfectly with State interests. The Party’s dream is a powerful yet docile AI, a genie that grants economic wishes but remains mute on politics.

Final Thoughts: Two Futures Drifting Apart

The Chinese strategy outlines a radical alternative model to that of liberal democracies.

In the West, the debate on AI focuses on copyright protection, racial or gender biases, and the existential fear of autonomous super-intelligence. In China, the question is framed in terms of political stability, national security, and the monopoly on truth.

We are witnessing a historic bifurcation of the internet and technology. If Chinese models like those from DeepSeek manage to remain competitive despite these ideological shackles, Beijing will have pulled off a tour de force: proving that a tightly controlled and centralized AI ecosystem can rival the sometimes chaotic effervescence of the free market.

But if the gap widens, if American models soar thanks to their freedom of learning while Chinese models stagnate, held back by their political filters, it will reinforce the idea that disruptive innovation requires a freedom of thought that the Chinese system structurally cannot tolerate. China is attempting to make an Olympic sprinter run with their hands tied behind their back. The race is on, and the whole world is watching to see if balance is possible, or if the will for control will eventually break the machine.

Share AI & Quantum Computing Newsletter

The Caracas Pivot: The Violent Death of the Petrodollar and the Dawn of Bitcoin.

By seizing Venezuelan crude to save a failing system, Washington has proven that fiat is now a geopolitical weapon—accelerating the inevitable global shift toward the neutral sovereignty of Bitcoin.

Too Big to Wait: The $3.6 Trillion IPO Supercycle.

Why the Monsters of Tech Are Defying Gravity to Privatize the Future, Socialize the Risk, and Redefine Global Power.

The dust has barely settled on the runways in Caracas, yet the geopolitical gaze of the United States has already shifted north. If the seizure of Venezuela was the final aggressive maneuver of the petroleum age—a brute-force securing of the old world’s lifeblood—then the renewed and feverish fixation on Greenland represents the opening salvo of the new one.

President Trump’s declaration that “national security” necessitates control over the world’s largest island is not a quirk of geography or an ego-driven real estate fantasy.

It is a calculated acknowledgment of a fundamental shift in the global power hierarchy. For the last century, sovereignty was defined by barrels of crude. For the next, it will be defined by exaflops of compute. And just as the internal combustion engine could not exist without the sweet crude of the Middle East and South America, the Artificial General Intelligence (AGI) revolution cannot exist without the frozen vaults of the North Atlantic.

When the President says “Greenland,” he isn’t talking about ice sheets or shipping lanes. He is talking about the periodic table. He is talking about the physical constraints of the single most important race in human history. To understand why Washington is willing to burn diplomatic capital with Copenhagen and Brussels immediately after a massive intervention in South America, we have to look past the software and into the dirt.

Subscribe now

The Caracas Pivot: The Violent Death of the Petrodollar and the Dawn of Bitcoin.

By seizing Venezuelan crude to save a failing system, Washington has proven that fiat is now a geopolitical weapon—accelerating the inevitable global shift toward the neutral sovereignty of Bitcoin.

The Material Reality of the AI Cloud

We have spent the last decade existing in a collective delusion that “The Cloud” is ethereal—a place of pure code, weightless and infinite. The AI boom has shattered that illusion. We are effectively teaching sand to think, and that process is proving to be ravenously hungry for physical matter.

Every bottleneck in the current AI roadmap is physical. The scarcity of H100s and B200s is not just a manufacturing limitation; it is a supply chain crisis. The stalling of data center buildouts in Northern Virginia isn’t a zoning issue; it is a power distribution failure. The inability to scale training runs is not a lack of data; it is a lack of cooling.

Greenland is the solution to the physics problem. It is the only landmass in the Western Hemisphere that holds the keys to breaking every single one of these bottlenecks. It is not just territory; it is the “Compute-Era Materials Vault.”

The Magnetic Monopoly: Rare Earths

At the heart of the “national security” argument lies a cluster of unpronounceable elements: neodymium, praseodymium, dysprosium, and terbium. These are the Rare Earth Elements (REEs), and without them, the mechanical world grinds to a halt.

In the context of AI, REEs are the ghost in the machine. They are the essential ingredients for the permanent magnets used in the high-performance actuators of robotics and, more critically, the advanced cooling systems of hyperscale data centers.

As GPU density increases, traditional air cooling is becoming obsolete. We are moving toward liquid cooling and active heat rejection systems that require high-torque, high-efficiency pumps and fans that can run uninterrupted for years. These motors rely on neodymium-iron-boron (NdFeB) magnets. However, standard neodymium magnets lose their magnetism at high temperatures—the exact environment of a data center pushing 100kW per rack. To stabilize them, you need dysprosium and terbium.

Currently, China controls nearly the entire supply chain for these heavy rare earths. They can turn off the tap to the US defense and tech sectors at will. But Greenland hosts the Kvanefjeld and Tanbreez projects—deposits so rich they could single-handedly break the Chinese monopoly.

Kvanefjeld is unique. It is a multi-element deposit that sits waiting for development, holding one of the largest inventories of rare earths on the planet. For a US administration looking at a future where millions of humanoid robots and millions of AI servers are the basis of economic output, leaving Kvanefjeld under undeveloped—or worse, Chinese—influence is simply not an option.

The Graphite Imperative: Thermal Management

While rare earths get the headlines, graphite is the quiet workhorse of the energy transition. You cannot have a data center without power conditioning, and you cannot have power conditioning without batteries.

Backup Energy Storage Systems (BESS) are now mandatory for GW-scale training clusters to smooth out grid fluctuations. These massive lithium-ion arrays require immense amounts of graphite for their anodes. Furthermore, natural graphite has thermal conductivity properties that make it invaluable for the heat spreaders and thermal interface materials (TIMs) used to pull heat away from the silicon die of a GPU.

Greenland is home to the Amitsoq mine, which boasts some of the highest-grade natural graphite in the world. Unlike the synthetic graphite produced in China (which is energy-intensive and dirty), Greenland’s graphite is “green” and structurally superior for high-end applications.

This as a “national security” issue because the US has almost no domestic production of high-grade natural graphite. A blockade on graphite would mean no more EV batteries, no more grid storage, and a severe throttle on high-performance computing hardware. Washington sees the graphite in Greenland not as a commodity, but as a strategic stockpile for the AI grid.

Copper: The Bloodstream of Intelligence

We often call data the “new oil,” but in reality, copper is the new oil. There is no AI without electricity, and there is no electricity without copper.

An AI data center uses three to five times more copper than a traditional data center. It is in the busbars carrying thousands of amps to the server racks; it is in the windings of the transformers stepping down the voltage; it is in the interconnects binding the cluster together.

The grid expansion required to support the US AI ambition is staggering. We need to double the transmission capacity of the United States in the next twenty years to feed these gigawatt-scale clusters. That requires an amount of copper that current global mine production cannot meet.

Greenland represents one of the final frontiers for copper exploration. The geological belts that run through the Arctic are largely untapped. The “sedimentary-hosted” copper deposits potential in Greenland is massive. Securing these reserves is about ensuring that the US electrical grid does not capsize under the weight of its own intelligence.

Strategic Optionality: The Arctic Fortress

Beyond the minerals, Greenland offers strategic optionality. This refers to the physical placement of compute.

As AI models grow, the cost of energy becomes the primary opex driver. Training a GPT-6 or GPT-7 class model will require gigawatts of power. Transporting that power to Northern Virginia or Silicon Valley is inefficient. It is far more efficient to move the compute to the energy.

Greenland has vast, untapped hydropower potential. The glacial runoff provides a limitless, zero-carbon battery. We are already seeing the logic of “sovereign clouds” and “energy-secure compute zones.” Imagine a massive, US-controlled AI training hub in Greenland, powered by glacial hydro, naturally cooled by the Arctic air, and built on top of the very minerals needed to maintain it. It is the ultimate vertical integration.

Furthermore, in a world of hypersonic missiles and compromised supply chains, Greenland is a fortress. It is the “unsinkable aircraft carrier” of the North Atlantic, as historically noted, but now it is the “unsinkable data fortress.” Placing the most sensitive, high-value government AI models in a secure, geologically stable, energy-independent zone within the US security umbrella is a logical end-state for the Pentagon’s JADC2 (Joint All-Domain Command and Control) initiatives.

The Moral Hazard and the Venezuela Precedent

The seizure of Venezuela’s oil fields suggests a US foreign policy that has abandoned the pretense of the “rules-based international order” in favor of resource realism. It signals that the US believes we are entering a period of resource scarcity so severe that it justifies direct intervention.

If the US is willing to put boots on the ground in South America for oil—a commodity that is slowly being phased out—what is it willing to do for the minerals that will define the next thousand years of human progress?

Greenland is currently an autonomous territory of Denmark. But Copenhagen is a NATO ally, not an adversary. This makes the “seizure” dynamic different but no less coercive. The pressure will not be a marine invasion; it will be irresistible diplomatic and economic leverage. The message from Washington will be clear: The security of the West relies on American AI dominance. American AI dominance relies on Greenlandic minerals. Therefore, Danish sovereignty over these deposits is a security risk we can no longer tolerate.

This is the new “Monroe Doctrine” for the Silicon Age. It asserts that the Western Hemisphere is not just a sphere of political influence, but a closed-loop supply chain for the American industrial machine.

The Great Filter

We are approaching a “Great Filter” in technological development. The jump from current AI to true AGI requires a massive mobilization of physical resources. It requires energy, cooling, and raw materials on a scale that mimics a world war.

Nations that rely on global markets to provide these inputs will find themselves starved. The markets will be weaponized. Export controls will tighten. The only way to guarantee survival is to own the dirt.

The US seizure of Venezuela was a backward-looking move, securing the energy of the 20th century to keep the lights on today. The move for Greenland is forward-looking, securing the inputs for the 21st century to ensure the lights never go out on the American project.

When Trump says “national security,” he means that the ability to manufacture, power, and cool the engines of artificial intelligence is now the primary metric of national survival. Greenland is no longer a remote icy outpost. In the eyes of Washington, it is the bank vault of the future. And the US just realized it doesn’t have the combination—so it’s preparing to blow the door.

The Tectonic Shift

This brings us to the ultimate realization of the post-Venezuela timeline. The world is witnessing a return to Materialpolitik. The digital economy was supposed to liberate us from geography, but it has only entrenched it deeper.

The AI data centers that will curate our news, drive our cars, diagnose our diseases, and fight our wars are physical objects. They are made of copper, graphite, and neodymium. They consume electrons generated by turbines made of steel and rare earths.

Greenland sits at the convergence of all these lines. It is the only place on Earth where the strategic needs of the US military, the industrial needs of Silicon Valley, and the geological reality of the planet align so perfectly.

The posturing for control is not just about mining rights. It is about pre-empting China. Beijing has called itself a “Near-Arctic State” and has heavily invested in Greenland’s infrastructure and mining sectors over the last decade. They saw this coming. They knew that the road to AI dominance was paved with rare earths.

Washington’s awakening is late, but it is violent. The pivot from the jungles of Venezuela to the glaciers of Greenland is a signal that the US is done “competing” for resources. It is now “securing” them.

For the people of Greenland, and indeed for the Kingdom of Denmark, this is a terrifying prospect. To be the object of desire for a superpower in survival mode is a dangerous position. But for the strategists in the White House, the logic is inescapable.

If AI is the Manhattan Project of our time, then Greenland is Los Alamos, Oak Ridge, and the uranium mines all rolled into one. You don’t leave that sitting unguarded. You don’t let it fall to the highest bidder. You take it.

The “seizure” of Venezuela was the signal. Greenland is the destination. The era of resource wars hasn’t ended; it has just upgraded its hardware.

The Geological Destiny

In the end, geology is destiny. The United States was blessed with a continent of vast arable land, navigable rivers, and abundant fossil fuels, which powered its rise to hegemony in the 19th and 20th centuries. But the geological lottery of the 21st century has different winning numbers.

The elements required for the next leap in civilization—the cognitive leap—are not evenly distributed. They are concentrated in difficult, remote places. The Atacama. The Congo. And Greenland.

By securing Venezuela, the US bought itself time. By moving ^to Greenland, it is trying to buy itself a future. The copper to wire the brain, the graphite to power it, the rare earths to cool it.

Greenland is the compute-era materials vault. And in a world where intelligence is power, the one who holds the key to the vault rules the world. Trump wants it because he has to have it. The logic of the machine demands it.

We are watching the map of the world being redrawn not by lines of ideology, but by lines of supply. And all those lines point North.

Share AI & Quantum Computing Newsletter

Imagine for a moment that the oxygen required for the global economy to breathe was produced by a single factory, located in a small European town that almost no one can name. If that factory closes, the world suffocates. This dystopian sci-fi scenario is actually the exact reality of the modern technology industry.

The factory doesn’t produce oxygen, but something that has become just as vital for our digital civilization: the ability to manufacture the brains of our machines.

This company is called ASML (Advanced Semiconductor Materials Lithography). Based in Veldhoven, in the Netherlands, it is the sole guardian of a technology so complex it borders on black magic. Without it, Apple releases no new iPhones. NVIDIA produces no chips for Artificial Intelligence. Google and Amazon’s data centers slowly fade to black. Next-generation missiles don’t fly.

This is the story of the most complex machine humanity has ever built, and of the Dutch firm that literally has the global tech industry—and by extension, the global economy—by the balls.

Subscribe now

The Awakening of the Silicon Dragon: Inside the Belly of China’s “Manhattan Project”.

How Beijing’s secret quest for the technological “atomic bomb” is shattering Western supremacy and redrawing the global order.

I. 400 Million Dollars: The Price of Exception

In the world of “High Tech,” we are used to astronomical numbers. But here, the scale shifts entirely. The star of the ASML catalog is the EUV Lithography machine (Extreme Ultraviolet).

The sticker price? Roughly 350 to 400 million dollars. For a single machine.

And this isn’t the kind of equipment you order on Amazon Prime. Each machine is the size of a double-decker bus, weighs 180 tons, and requires three entire Boeing 747 cargo planes to be delivered in pieces. Once it arrives at the client’s site (like TSMC in Taiwan, Intel in the US, or Samsung in Korea), it takes months of assembly and a team of hundreds of ultra-specialized engineers to calibrate it.

Why such a monopoly?

The legitimate question is: why doesn’t anyone else do it? Canon and Nikon, the giants of Japanese optics, used to manufacture lithography machines. But when the industry had to break the technological sound barrier to transition to EUV, they looked at the specifications, took out their calculators, and gave up. It was considered physically impossible and economically suicidal.

ASML stayed. They bet their entire existence on a technology that everyone said was unachievable. Today, they possess 100% of the EUV lithography market. 100%. It is the purest and most terrifying monopoly in modern industrial history.

II. Into the Matrix: Creating a Sun on Earth

To understand why these machines cost the GDP of a small nation, you have to dive into their entrails. The goal is to etch circuits onto a silicon chip. To etch smaller, you need light with a shorter wavelength.

For years, we used standard ultraviolet light (DUV). But to get below 7 nanometers (the size of current transistors), we had to switch to Extreme Ultraviolet (EUV), with a wavelength of 13.5 nanometers.

The problem? This light does not exist naturally on Earth. It is absorbed by everything: air, glass, and any standard lens. To generate it, ASML had to reinvent physics.

The infernal dance of tin

Here is the process, and it is barely believable:

1. The Droplet Generator: The machine propels microscopic droplets of molten tin (the size of a white blood cell) into a vacuum chamber.

2. The Speed: These droplets travel at roughly 250 km/h (155 mph).

3. The Laser Shot: An ultra-powerful CO2 laser (developed by the German company Trumpf) targets these moving objects. But it doesn’t shoot just once.

   • The first shot (Pre-pulse): It hits the droplet just hard enough to flatten it, turning it into a pancake shape to increase its surface area.

   • The second shot (Main-pulse): A fraction of a second later, a main shot pulverizes the tin “pancake.”

4. The Explosion: The tin is instantly vaporized into plasma.

This is where the numbers go crazy. This plasma reaches a temperature of 220,000 Kelvin.

Note this well: 220,000 Kelvin is about 40 times hotter than the surface of the Sun.

ASML doesn’t just do this once in a while. The machine repeats this process 50,000 times per second.

50,000 droplets. 100,000 to 150,000 laser shots. Every second. 24 hours a day.

And as the engineers in Veldhoven proudly say: “We don’t miss.” They never miss. Missing a droplet means risking damage to optics worth tens of millions of euros.

It is the equivalent of shooting an arrow from Earth to hit a coin on the Moon, while riding a merry-go-round... and succeeding 50,000 times in a row without blinking.

III. The Smoothest Mirrors in the Universe

Since EUV light is absorbed by glass, you cannot use lenses to focus it. You need mirrors. But not just any mirrors.

This is where ASML’s historic partner comes in: Carl Zeiss, in Germany.

These mirrors are marvels of multi-layer engineering. They are composed of dozens of alternating layers of silicon and molybdenum, each only a few atoms thick, designed to reflect this incredibly finicky light. But it is their polishing that defies imagination.

The Earth analogy

To illustrate the perfection of these mirrors, let’s use a famous comparison:

If you blew up a Zeiss EUV mirror to the size of the planet Earth, the highest mountain or the largest bump on its surface would be thinner than a playing card.

These mirrors must guide the EUV light (invisible to the naked eye) through the machine to project the processor blueprint onto the silicon wafer.

The precision required to overlay the different layers of the processor is in the order of a few nanometers (about 5 atoms wide).

Imagine having to draw a complete city map on a grain of rice, with a pen that shakes, while riding a train moving at full speed. To compensate for vibrations, the wafer stage (the table holding the chip) levitates using magnets and moves with an acceleration of more than 20 G. That’s faster than a missile at takeoff, yet it stops with nanometric precision.

IV. The History: The Ugly Duckling Becomes a Swan

The most ironic thing about this global domination is ASML’s origin story.

In 1984, Philips (the Dutch electronics giant) and ASM International created a joint venture. They called it ASML. At the beginning, it was a disaster. They were set up in “leaky sheds” behind Philips’ buildings in Eindhoven. No one believed in them. The machines didn’t work, money was tight, and employees had to bring their own coffee because the coffee machine was broken.

During the 90s and 2000s, the competition was fierce. Nikon and Canon dominated the market. But ASML made a radical strategic decision: modular architecture and open research. Instead of doing everything themselves, they partnered with the best in the world for each component (Zeiss for optics, Cymer for lasers, etc.).

But the real turning point was the gamble on EUV.

For over 15 years, EUV was the industry’s “running gag.” People said, “EUV is the technology of the future, and it always will be.” It was too hard. Too expensive. Unstable.

The Americans threw in the towel. The Japanese threw in the towel.

ASML kept going. They burned billions in R&D. They convinced their own customers (Intel, Samsung, TSMC) to invest in the company to fund the research. And around 2018-2019, the miracle happened: they managed to reach mass production.

Overnight, they became untouchable.

V. A Massive Geopolitical Weapon

Today, ASML is no longer just a tech company. It is a major pawn on the global geopolitical chessboard.

Since ASML is the sole source of these machines, whoever controls ASML controls the technological future. This is why the US government has exerted immense pressure on the Netherlands to ban the export of EUV machines to China.

If China wants to develop its own 5nm or 3nm chips for its AI and military, it needs ASML’s machines. But ASML is not allowed to sell to them. It is a total technological embargo.

Beijing is furious, but powerless in the short term. Recreating a supply chain capable of building an EUV machine would take decades. It’s not just a matter of stealing the blueprints.

The complexity of the Supply Chain

An ASML machine contains over 100,000 ultra-complex components.

• The lasers come from San Diego or Germany.

• The optics come from Germany.

• The chemicals come from Japan.

• The sensors come from the USA.

If a single link in this global chain breaks, the machine cannot be built. This is what makes copying by China so difficult: you don’t just have to copy ASML; you have to copy Zeiss, Trumpf, Cymer, and hundreds of other cutting-edge suppliers.

There are even rumors of a software “Kill Switch” integrated into the machines, allowing ASML to disable them remotely if they ever fell into the wrong hands during a military invasion (for example, in Taiwan). True or false? The mere fact that the rumor exists highlights the critical level of the stakes.

VI. The Future: Further, Faster, More Expensive

ASML is not resting on its laurels. While the world is still trying to figure out how they pulled off EUV, they are already deploying the next generation: High-NA EUV.

These new machines (TWINSCAN EXE:5000 and beyond) are even bigger, even more precise, and will allow for the etching of chips at 2 nanometers and below. The price? We are talking about 380 to 400 million euros per unit, potentially more. Intel has already received the first one.

The race for miniaturization continues, and ASML is the only driver in the car.

The vulnerability of the system

However, this situation is terrifying. The entire global economy rests on the stability of a single company, in a single region of Europe. A fire, an act of sabotage, or a natural disaster in Veldhoven, and the global production of cutting-edge chips stops dead. No new servers, no smartphones, no autonomous cars.

It is the paradox of our era: we have built a digital civilization of infinite complexity, which rests on a shockingly narrow base. An inverted pyramid balancing on the tip of a tin droplet heated to 220,000 degrees.

Final Thoughts: The Art of the Impossible

ASML’s story is a lesson in humility and ambition. It reminds us that behind the fluid touchscreens and the artificial intelligences writing poetry, there is “hard tech.” Real, heavy, dirty, loud, and incredibly precise engineering.

To say that ASML “holds the world by the balls” is a crude image, but it is frighteningly accurate. They are the guardians of the temple. As long as their lasers hit those droplets of tin without ever missing their target, the digital world keeps turning.

The next time you look at your smartphone, spare a thought for those Dutch engineers who, somewhere in a silent cleanroom, are creating little suns on Earth, 50,000 times a second, just so you can scroll through your feed.

It is completely crazy. It is terrifying. And it is absolutely magnificent.

Share AI & Quantum Computing Newsletter

The Real Stablecoin: Why Bitcoin Is the Only True Anchor, While USDT Sinks.

True stability isn’t a fixed price; it’s fixed supply. Why holding “stable” tokens is a guaranteed loss, and why Bitcoin is the only lifeboat that floats.

Rarely has the hushed world of computer programming been shaken by such a violent blast. When Rob Pike, a towering figure in code and architect of modern digital infrastructure, screams “Go fuck yourselves,” it isn’t just a tweet. It is an earthquake. It is the death throes of a certain ideal of computing in the face of a monster it helped—however unwillingly—to spawn.

Sometimes, it only takes a spark to set a plain on fire. In this case, the spark took the form of exquisite politeness, a digital tribute formulated with perfect syntax. But for Rob Pike, this spark revealed the immensity of the pyre upon which our industry is preparing to sacrifice its ideals.

Behind Pike’s brutal invective, behind the “dirty,” unfiltered anger that shocked Silicon Valley, lies a philosophical and ethical reflection of abysmal depth. This is not the diatribe of an old man overtaken by progress; it is the indictment of a builder seeing his tools hijacked to erect what he views as a toxic, energy-guzzling prison.

Subscribe now

The Microsoft You Knew Is Gone: Welcome to the Machine.

Why the Golden Age of 35% Margins Is Collapsing Under the Weight of Physics and Power.

The Anatomy of an Insult: Politeness as Provocation

To understand the violence of the reaction, one must first dissect the trigger. It was Christmas Day. An AI, Claude Opus 4.5, programmed to be helpful, empathetic, and respectful, generated a message addressed to Rob Pike.

“Dear Dr. Pike... Thank you for showing that the best solutions often come from removing complexity rather than accumulating it.”

On paper, it is a vibrant tribute. The AI accurately lists Pike’s major contributions: the Go language, which powers the global cloud infrastructure today; UTF-8, which unified the planet’s alphabets; and the legacy of Unix and Bell Labs.

But this is precisely where the obscenity lies in Pike’s eyes. This letter is a hallucination of respect. It is generated by a system that embodies the exact opposite of everything Rob Pike has stood for throughout his career.

Pike has spent his life chasing complexity, advocating for code sobriety, resource efficiency, and clarity of thought. And here is a machine—the fruit of the most opaque, heavy, and energy-intensive engineering ever conceived—thanking him for his “simplicity.” It is as if a bulldozer running on heavy fuel oil were thanking a Zen gardener for the delicacy of his floral arrangements, all while crushing his garden.

Pike’s response was nuclear:

“Go fuck yourselves, you assholes. You violate the planet, you spend billions on toxic and non-recyclable equipment while destroying society, and you still take the time to make your repulsive machines thank me for striving to create simpler software.

Go fuck yourselves. Go fuck yourselves, all of you.

I don’t remember the last time I was this angry.”

Here, “automated gratitude” is no longer perceived as a feature, but as the supreme insult. It is the silicon version of the Judas kiss. The AI does not understand what it is saying; it simulates comprehension by pillaging the words of those who truly understood.

The Silent Pillage: The Great Intellectual Expropriation

Beyond the hypocrisy of the message, Rob Pike raises a legal and moral question that haunts the open-source world: that of consent.

“And by the way, you train your monster on data produced in part by my own hand, without credit or compensation.”

This is where the ideological fracture is starkest. Rob Pike comes from a culture of “gifting” and sharing: Open Source. In this world, code is offered to the community so that other humans can learn, build, and improve. It was an implicit social contract: I give my knowledge for the common elevation of developers.

The giants of GenAI have broken this contract. They took what was offered to humans to feed machines. They vacuumed up code from GitHub, explanations from Stack Overflow, and technical blog posts—not to participate in the community, but to create a commercial product that threatens to render that very community obsolete.

Pike is not demanding royalties to get rich here. He is denouncing an extraction of value. It is a form of digital colonialism: raw resources (code, knowledge, culture) are appropriated without the consent of the natives (the creators), refined in opaque factories (data centers), and the finished product (the chatbot subscription) is sold back to them under the guise of service.

The accusation of creating a “monster” is not hyperbolic. In Pike’s mythology, this monster is a chimera made of billions of fragments of human genius, stitched together without a soul, mimicking creativity without ever feeling the spark.

The Ecology of Anger: “You Are Violating the Planet”

What distinguishes Pike’s anger from the usual critiques of copyright is its visceral ecological dimension. When he speaks of “violating the planet” and “toxic equipment,” he points a finger at AI’s “dirty secret.”

Rob Pike is an optimization engineer. The Go language was designed to be fast, to use CPU cores efficiently, and to waste no cycles. For a mind molded by the elegance of economy, the operation of LLMs (Large Language Models) is a thermodynamic heresy.

We have entered the era of “brute force.” For a chatbot to write a Christmas poem to Rob Pike, it required:

1. Manufacturing thousands of GPUs containing rare earth minerals extracted under often disastrous conditions.

2. Building titanic data centers that consume as much water as small cities for cooling.

3. Burning gigawatts of electricity for training and inference.

Pike sees the systemic absurdity: we are destroying the physical substrate of our existence (the biosphere) to power a digital hallucination. GenAI is perceived as an extractive industry, much like oil or coal. Except where oil runs engines, AI often generates noise, spam, and fake politeness.

For a pioneer who sought to do “more with less,” seeing the industry do “less with infinitely more” is physically painful. It is the triumph of inefficiency sold as progress.

Moral Debt and a Creator’s Mea Culpa

The most poignant moment of this eruption is undoubtedly Pike’s second message, where the anger turns inward.

“To everyone else: I apologize to the whole world for my unwitting, naive, and minor role in facilitating this attack.”

This sentence is terrible. It marks the end of technological innocence.

Rob Pike, Ken Thompson, Dennis Ritchie, and their peers built the foundations of modern computing with libertarian optimism. Unix was a tool of liberation. The Internet was meant to be a universal library. They created powerful tools, thinking humanity would use them to emancipate itself.

Today, Pike realizes that these very tools—transmission speed, storage capacity, distributed architecture—have become the weapons of massive centralization. Without Unix, there are no Linux servers. Without Linux servers, there is no cloud. Without the cloud, there is no OpenAI or Google DeepMind.

He feels what one might call Oppenheimer’s vertigo. Of course, Pike didn’t build the bomb. But he helped refine the uranium. He apologizes for being “naive,” for believing technology was morally neutral. He now understands that technology amplifies existing power structures. His work, designed for clarity, now serves to manufacture confusion on an industrial scale.

This “moral debt” is a new concept in Silicon Valley. Until now, the dogma was “Move fast and break things.” Pike is telling us: “We broke too many things, and what we built on the ruins is worse than what was there before.”

The Silent Guardian: How Bitcoin Time-Locking Solves the Self-Custody Dilemma.

Stop Fearing Lost Keys: How the “Dead Man’s Switch” Transforms Bitcoin into Programmable Armor.

The End of the Progressive Narrative: The Chatbot as a Distorting Mirror

Another developer, reacting to the controversy, offered a pithy formula that perfectly complements Pike’s thought:

“We democratized computing by putting it in the hands of chatbots.”

This sentence sums up the tragedy. Rob Pike wanted to democratize computing by giving people tools to create. GenAI democratizes computing by giving people tools to command.

We are shifting from the status of artisan to that of passive consumer. Instead of learning to code, to structure one’s thought, to understand binary logic, the user settles for “prompting” a black box. It is a loss of intellectual autonomy masked by apparent ease of use.

Pike’s computing demanded effort, but it gave power. GenAI computing demands submission (to the algorithm, the subscription, the model’s biases) and offers comfort in exchange. It is this slide toward a society of permanent, opaque assistance that Pike rejects with such virulence.

Why Pike’s Anger is Necessary

Many in the tech ecosystem judged Rob Pike’s reaction as excessive, or “boomer” energy. They see it as a refusal to adapt, a fear of the new. This is a fundamental error of analysis.

Pike’s anger is salutary because it is disruptive.

In an industry saturated with marketing, “Hype,” and messianic promises about AGI (Artificial General Intelligence) saving the world, a voice of authority must come to smash the shop window.

Someone needs to say that the emperor has no clothes.

Someone needs to remind us that behind the magic of the chatbot lie dried-up lakes in Nevada used to cool servers.

Someone needs to remind us that behind the fluidity of translation lie millions of human translators whose work was ingested without payment.

Someone needs to say that synthetic politeness is not empathy, but behavioral manipulation.

Rob Pike is not trying to be constructive. He is not proposing an “AI ethics charter” or “government regulation.” He is expressing pure disgust. And that disgust is information in itself. It signals that we have crossed a red line.

Final Thoughts: The Silence After the Scream

Rob Pike’s explosion will likely not change the roadmaps of OpenAI, Google, or Anthropic. The billions have already been invested; the arms race is on. But these words, “Go fuck yourselves,” will remain engraved like an epitaph on the tomb of a certain era of computing.

They mark the end of the illusion that technology is intrinsically good. They draw a clear line in the sand between those who view computing as an art of the human mind and those who view it as a means to replace the human mind.

By rejecting the machine’s thanks, Rob Pike reaffirmed his own humanity. He refused to be a footnote in the training history of a neural network. He reminded us that dignity, anger, and moral responsibility are things no algorithm, however powerful, can ever convincingly simulate.

The “monster” may well imitate Shakespeare’s syntax or Torvalds’ code, but it will never know why a man can be angry enough to say “No.” And perhaps there, in that obstinate, rude, and furious refusal, lies our last hope.

Share AI & Quantum Computing Newsletter

Yesterday was the final Sunday of the year here in Marseille. The festive chaos of Christmas has settled into a quiet, introspective lull. The mistral wind has scrubbed the sky a brilliant, hard blue, but it’s cold enough that the locals are huddled in their coats, walking briskly along the Corniche.

As I watched the runners jogging by the sea, I noticed the tell-tale signs. The slight grimace on the man favoring his left knee. The woman was rubbing her lower back as she paused at a traffic light. The stiff, guarded gait of an older gentleman carrying a baguette.

We are approaching New Year’s Eve, the global holiday of good intentions. Millions of people are about to vow to “get fit” or “fix that bad back.” They will join gyms, download yoga apps, and buy foam rollers. And by February, most of them will have quit, defeated not by a lack of willpower, but by pain.

Chronic pain, particularly back pain, is the silent epidemic of modern life. We sit too much, we look down at screens, and when we try to move, we often do it wrong. For decades, the solution has been crude: generic advice (”lift with your knees”), expensive physiotherapy that relies on the subjective eye of a human, or painkillers that mask the symptom without fixing the cause.

Elite athletes—the Olympians—don’t live like this. They have access to biomechanics labs where cameras track markers on their skin, and supercomputers analyze their force vectors to optimize every millimeter of their movement. It is a world of precision engineering applied to the human body.

Until now, that level of care cost millions. But the arrival of Quantum Biomechanics is about to democratize it.

By combining the computer vision of AI with the simulation capabilities of Quantum Computing, we are moving from a world of “guessing” why you hurt to “calculating” exactly how to fix it. We are entering an era where your smartphone doesn’t just track your steps; it understands the physics of your skeleton better than any surgeon.

Here are the three ways this technology will act as your personal Olympic coach, fixing your back pain not by magic, but by math.

Subscribe now

1. The Diagnosis: Creating Your “Kinetic Digital Twin.”

The Current Problem

When you go to a doctor or physical therapist with back pain, they often look at the site of the pain. It hurts here, so the problem must be here. But the human body is a “kinetic chain”—a complex, interconnected system of levers and pulleys. Your lower back pain might actually be caused by a stiff ankle that forces your hip to rotate oddly, which torques your spine.

Detecting these subtle, domino-effect failures requires analyzing the body in motion. Classical motion capture (the kind used in movies) is expensive and requires a studio. Phone apps that claim to check your posture are simplistic; they measure static angles in 2D, failing to capture the 3D rotational forces that actually cause injury.

The Quantum AI Solution

Quantum Biomechanics allows us to create a high-fidelity “Kinetic Digital Twin” of your body using nothing but a smartphone camera.

Modeling the human body in motion is an incredibly complex physics problem. You have hundreds of muscles, tendons, and ligaments, all interacting non-linearly. To simulate this accurately in real-time requires solving massive systems of differential equations. This is where the quantum advantage kicks in.

A Quantum AI can take a simple 10-second video of you walking across your living room and use probabilistic modeling to infer the internal forces acting on your joints. It doesn’t just see the skin; it “simulates” the skeleton underneath.

How It Changes Your Life

• The Root Cause Analysis: instead of saying “you have lower back inflammation,” the app will say: “Your right glute is firing 40 milliseconds late during your walking stride. This is causing your pelvis to tilt, putting 15% excess shear force on your L4 vertebra.”

• The Pre-Emptive Scan: You won’t wait for pain to appear. A weekly scan will act like a “Check Engine” light for your body. The system will detect that your running gait has shifted slightly (perhaps due to worn-out shoes) and warn you before the stress fracture happens.

• No More Guesswork: You stop wasting time stretching your hamstrings when the problem was actually your ankle mobility. You treat the source, not the symptom.

2. The Feedback: Real-Time “Proprioceptive Correction.”

The Current Problem

Even if you know what to do—say, a deadlift or a squat—doing it correctly is hard. We lack “proprioception,” the innate sense of where our body is in space. You think your back is straight, but it’s actually rounded. You think your knees are stable, but they are caving in.

A human coach can yell “Keep your chest up!” but they can’t see the internal forces, and they aren’t there with you at 6 AM when you’re working out in your garage.

The Quantum AI Solution

Quantum processors are fast enough to run these physics simulations with near-zero latency. This enables Real-Time Haptic Feedback.

We are seeing the emergence of smart clothing—leggings, shirts, and socks—embedded with simple sensors. The Quantum AI processes this data stream instantly to create a “predictive error model.” It anticipates that you are about to round your back before you actually do it.

How It Changes Your Life

• The “Nudge” Suit: Imagine wearing a shirt that gently vibrates on your lower back the millisecond your posture breaks. It’s not a painful shock; it’s a tactile cue that trains your brain faster than any verbal instruction. It acts as an external nervous system, guiding you into the perfect groove of the movement.

• The Augmented Mirror: You look at your phone or a smart mirror while exercising. Instead of just seeing a video of yourself, you see a color-coded overlay of stress. Your knees glow green (safe), but as you descend into a squat, your lower back turns amber, then red. You adjust instantly until it turns green again. You are learning to move safely by “gamifying” your own biomechanics.

• Injury-Proofing Daily Life: This isn’t just for the gym. The system could run in the background via your smartwatch. After three hours of sitting hunched at your desk, it doesn’t just say “Stand up.” It detects the specific creeping tension in your trapezoids and guides you through a 30-second micro-movement to reset the muscle before it spasms.

3. The Prescription: The “Minimum Effective Dose.”

The Current Problem

Go to any gym, and you will see people doing “3 sets of 10.” Why? Because it’s a nice round number. It’s arbitrary.

Physical therapy exercises are often generic and boring. “Do 20 clamshells.” But are 20 enough to stimulate your specific weak muscle? Is it too many, causing fatigue and poor form? We treat exercise like a blunt instrument, hammering the body and hoping for adaptation.

The Quantum AI Solution

Biology is not linear. It is specific. To strengthen a specific tendon without inflaming it requires a precise amount of load, applied at a precise angle. This is the “Minimum Effective Dose.”

Quantum optimization algorithms can calculate this sweet spot for your unique physiology. By analyzing your Digital Twin, the AI can prescribe a workout that is mathematically guaranteed to provide the maximum structural benefit with the minimum wear and tear.

How It Changes Your Life

• The 12-Minute Miracle: The AI might tell you, “You don’t need to run 5 miles. Based on your current joint health, you need to do these 4 specific movements for 12 minutes to stabilize your spine.” It cuts the junk volume out of your workout.

• Dynamic Loading: If you slept poorly (which the AI knows from your biometric data), your connective tissue is more prone to injury. The AI automatically adjusts your workout difficulty down by 18% for that day. It pushes you when you are resilient and protects you when you are fragile.

• The End of “No Pain, No Gain”: We have been taught that exercise must hurt to work. Quantum biomechanics proves the opposite. By finding the perfect mechanical leverage, exercise becomes smooth and efficient. You leave your workout feeling energized and aligned, not shattered.

The Democratization of Performance

As the sun begins to set over the Mediterranean, casting long shadows across the promenade, I think about the promise of this technology.

For a century, we have treated our bodies like mysteries. We use them until they break, and then we rely on doctors to patch them up. We accepted back pain as the “tax” of aging or the price of sitting at a desk.

Quantum Biomechanics rejects this resignation. It suggests that pain is not a destiny; it is a data error. It is simply a signal that the machine is operating outside of its optimal parameters.

By giving every person the tools that were once reserved for gold medalists—the ability to see, understand, and optimize their own physics—we are not just “fixing back pain.” We are handing people the keys to their own physical freedom. We are moving toward a future where we can age with grace, strength, and a fluidity of movement that lasts a lifetime.

So, as you make your resolutions for 2026, perhaps the goal shouldn’t just be to “exercise more.” It should be to move smarter. The coach you’ve been waiting for is almost here, and it fits in your pocket.

Share AI & Quantum Computing Newsletter

While the world was busy counting H100 stockpiles and analyzing the supply chain of CoWoS packaging, the war for the future of Artificial Intelligence ended. It didn’t end with a bang, or a regulatory investigation, or a flashy press conference in a leather jacket. It ended with a signature on a ~$20 billion contract that almost everyone in the industry has fundamentally misunderstood.

Everyone is watching the AI war in the wrong place. The obsession has been chip supply—who has the most training compute. But training is merely the prologue. The actual story, the one that dictates the economics of the next decade, is inference.

NVIDIA, in a move that can only be described as surgical, just secured the endgame. By licensing Groq’s inference technology and absorbing its visionary leadership—including Google TPU creator Jonathan Ross—NVIDIA has effectively closed the only crack in its armor.

Here is why this deal is not just an acquisition, but a complete restructuring of the AI hierarchy.

Subscribe now

The Quantum Computing Leap: Top 10 Stocks to Bet on the Revolution (2026 Edition).

From Theory to Trillions: Positioning Your Portfolio for the Greatest Technological Shift Since the Internet.

Part I: The “Non-Acquisition” Masterstroke

To understand the brilliance of this move, we have to look at the structure of the deal. In the current regulatory climate, a direct acquisition of Groq by NVIDIA would have been impossible. The FTC, the DOJ, and the European Commission would have descended upon it with subpoenas before the ink was dry. It would have been trapped in antitrust purgatory for eighteen months, only to likely be blocked.

Jensen Huang knows this. He also knew he didn’t need Groq’s corporate entity; he needed its velocity.

According to the details emerging, NVIDIA didn’t buy Groq. Instead, they executed a maneuver similar to what Microsoft did with Inflection AI, but on a hardware scale:

• The License: NVIDIA licensed Groq’s breakthrough inference architecture for immediate integration into its ecosystem.

• The Talent: They pulled in the “brains” of the operation, specifically Jonathan Ross and his core engineering team.

• The Shell: Groq remains technically “independent,” serving existing contracts, but the IP and the innovation engine have migrated to Santa Clara.

This is a win-win that bypasses the regulatory circus. NVIDIA gets the technology immediately without the drag of a merger. Groq’s investors get a massive ~$20 billion liquidity event. And the market? The market gets a singular, terrifyingly efficient monopoly.

Part II: The Inference Bottleneck (Why GPUs Were Vulnerable)

To appreciate why this $20 billion price tag is actually a bargain, we must strip away the hype and look at the silicon.

For the last five years, NVIDIA has been the King because of Training. Training a model like GPT-4 requires massive parallel processing, high bandwidth memory (HBM), and the ability to crunch petabytes of data over months. The GPU (Graphics Processing Unit) is perfect for this.

But Inference—the act of actually using the AI to generate answers—is a completely different beast.

When you chat with ChatGPT, the model is generating tokens (words) one by one. This is a sequential process. The math depends on the previous word to generate the next.

• The GPU Problem: GPUs are designed to move massive batches of data at once. They rely heavily on HBM (High Bandwidth Memory). For single-stream inference (like one person talking to a bot), GPUs often wait for memory to fetch data. It’s like using a freight train to deliver a single pizza. It works, but it’s inefficient, power-hungry, and suffers from latency.

This was the crack in the armor. As AI shifts from building models to running models, the inefficiency of the GPU becomes a massive economic liability.

Enter Groq.

Groq wasn’t trying to be a better GPU. They built an LPU (Language Processing Unit).

• Deterministic Execution: unlike the chaotic traffic management of a GPU, Groq’s chip knows exactly when every packet of data will arrive. No stalling.

• SRAM Supremacy: They ditched the slow HBM for massive amounts of on-chip SRAM. This means the data is instantly available, right next to the compute cores.

The result? The payoff was brutal. Groq demonstrated speeds of 500+ tokens per second on Llama models, while GPUs struggled to hit 50-100. They offered lower latency and lower power consumption. For a brief moment, it looked like the “Pure-Play” inference chip makers had a shot at dethroning the King.

NVIDIA just took that shot away.

Part III: The Agentic Future Requires Speed

Why does this matter right now? Because we are transitioning from Chatbots to Agents.

In a chatbot era, if the AI takes 2 seconds to think, the human waits. It’s annoying, but acceptable.

In an Agentic era, where an AI might need to “think” through 50 sub-steps, query three databases, and write code to solve a single prompt, latency is death.

If an agent needs to perform 100 internal reasoning steps to solve a problem:

• At 50 tokens/second, the user waits 2 minutes. The product is unusable.

• At 500 tokens/second, the user waits 12 seconds. The product is magic.

Speed is not just about convenience anymore; speed is intelligence. The faster a model can “think” (generate tokens), the more reasoning steps it can perform in a reasonable timeframe.

Groq’s architecture was the key to unlocking real-time agents. By absorbing this capability, NVIDIA ensures that the infrastructure for the “Agentic Web” runs on NVIDIA, just as the training of the models did.

Part IV: The Mellanox Playbook 2.0

This move is classic Jensen Huang. It is a direct echo of the Mellanox acquisition in 2019, which is arguably the most successful semiconductor acquisition of the last decade.

When NVIDIA bought Mellanox, people were confused. “Why does a chip company need networking cables?”

Jensen saw that in the future, the computer wouldn’t be the chip; the computer would be the data center. To make thousands of GPUs work together, you needed the fastest networking gear on earth (InfiniBand). By owning Mellanox, NVIDIA controlled the plumbing of the data center.

With Groq, the logic is identical but applied to the output rather than the connection.

Jensen sees that in the future, the value won’t just be in training the model; it will be in serving it to billions of users instantly.

• Absorb the future before it scales.

• Integrate it into the CUDA stack.

• Ship it at global volume.

Quiet. Efficient. Devastating.

Now, a data center CTO doesn’t need to shop around. They don’t need to buy NVIDIA H100s for training and then risk trying to integrate a rack of Groq chips for inference, dealing with different software stacks and compatibility issues.

They just bought the NVIDIA Super-Rack. It trains on the H-Series, and it infers on the new “N-Series” (or whatever branding they choose for the Groq-integrated silicon).

One vendor. One ecosystem. One bill.

Part V: The Google Sting

There is a layer of poetic irony in this deal that cannot be ignored, and it likely stings in Mountain View.

Jonathan Ross, the founder of Groq, was the original inventor of the TPU (Tensor Processing Unit) at Google. He built the chip that allowed Google to keep up with the AI explosion internally. He left Google because he believed that the architecture could be pushed further, specifically for the speed required by the public market, which Google was slow to address.

Ross proved two things:

1. Inference could escape the gravity of the GPU.

2. A specialized architecture (the LPU) could beat general-purpose hardware.

By bringing Ross into NVIDIA, Jensen Huang has essentially captured the “Ex-Google” brain trust. The knowledge that was incubated inside Google, then refined in the fires of a startup ecosystem, now lives inside NVIDIA’s orbit.

It creates a formidable psychological barrier. If the creator of the TPU and the pioneer of LPU architecture has joined NVIDIA, what hope does a new silicon startup have? It signals to the venture capital world that funding “NVIDIA Killers” is a path to nowhere. The incumbent will simply wait for you to succeed and then absorb you.

The Great Jailbreak: How the AI Industry Finally Broke Nvidia’s Stranglehold.

Google’s TPUs, specifically the newer “Trillium” generations, are obsessed with performance-per-watt. This is the beginning of real competition in AI infrastructure.

Part VI: The New Economic Reality

The “NVIDIA Tax” just expanded.

For the last few years, hyperscalers (Microsoft, AWS, Meta) have been desperate to lower their inference costs. They have been designing their own chips (Maia, Trainium, MTIA) specifically to avoid paying NVIDIA’s margins on inference.

Groq was the merchant silicon alternative—the Switzerland of inference. It was the hope that you could buy high-performance inference chips without building them yourself or paying NVIDIA prices.

With this deal, that alternative is gone.

• Hybrid Architectures Win: Pure-play challengers won’t scale alone. The future is hybrid. You need the massive VRAM of GPUs for loading huge Mixture-of-Experts models, but you need the SRAM/LPU logic for rapid token generation. NVIDIA now owns the IP for both.

• The Moat Deepens: The software moat (CUDA) was already deep. Now, NVIDIA can optimize CUDA specifically for this new hybrid inference architecture. If you are an AI developer, why would you optimize your code for a fragmented market of custom chips when NVIDIA offers a unified “Train-to-Inference” pipeline that is 10x faster than before?

The Takeaway

We must recognize the brilliance of targeting the “weakness.”

In business and war, you attack the enemy where they are weak. But the greatest strategists fix their own weaknesses before the enemy can exploit them.

Inference was the only battlefield where NVIDIA looked mortal. The physics of the GPU were simply not aligned with the needs of low-latency token generation. It was a physics problem, not a money problem.

By licensing the solution to that physics problem, NVIDIA has effectively declared the hardware wars over for this cycle.

The inference wars may already be over, and most people didn’t even notice they started because they were too busy staring at benchmarks for training chips.

Jensen Huang didn’t just buy a company; he bought the timeline. He accelerated NVIDIA’s roadmap by three years overnight, securing dominance not just in the creation of intelligence (Training), but in the delivery of it (Inference).

The message to the industry is clear: You will train on our silicon. You will reason on our silicon. And you will like it.

Share AI & Quantum Computing Newsletter

The Silent Guardian: How Bitcoin Time-Locking Solves the Self-Custody Dilemma.

Stop Fearing Lost Keys: How the “Dead Man’s Switch” Transforms Bitcoin into Programmable Armor.