AI Is Bigger Than Five: The Founding Story of the AI Empires

Jul 2, 2026

Introduction: Move 37

March 10, 2016. Seoul. In the game room on the sixth floor of the Four Seasons Hotel, Lee Sedol — one of the finest Go players alive — was playing not against the man sitting across the board, but against the machine behind him. On the thirty-seventh move of the second game, Aja Huang, the DeepMind engineer tasked with placing AlphaGo’s stones, set a black stone on the fifth line, in a spot that centuries of professional Go instruction had marked never play here. Fan Hui, the European champion commentating the live broadcast — the same man who had lost 5-0 to this machine a few months earlier — first assumed it was a mistake. Then he stopped, looked again, and spoke the sentence he would later repeat to the journalist Cade Metz: “It’s not a human move. But so beautiful.” By DeepMind’s own calculation, the probability that a human professional would have played that move was one in ten thousand. Lee Sedol left the room, returned fifteen minutes later, and lost the game. He would lose the match 4-1 — though the 78th move he played in game four would come to be called “God’s move,” and would enter history as humanity’s last great victory on the Go scoreboard.

More than two hundred million people watched those games that week. Most of them sensed that what they were watching was not a game. Move 37 was the first great public demonstration of a sixty-year-old scientific claim — the claim that machines could think. But it heralded something else, too: the institution staging this demonstration was not a university, not a government laboratory, but a private company founded in London and sold to Google two years earlier. The project of solving intelligence was passing out of the hands of public science and into the hands of corporations.

Today, in mid-2026, the story of artificial intelligence is mostly told through five names: OpenAI, Google DeepMind, Anthropic, xAI, and Meta AI. The founders of these five labs grace magazine covers, testify before congressional committees, sue one another in federal court, and sit at the center of global capital with valuations that sum to trillions of dollars. The narrative is seductive: a handful of geniuses, a handful of betrayals, a handful of model launches.

But the narrative is incomplete. The neural-network idea that made Move 37 possible was kept alive for forty years at the margins of academia, by a handful of stubborn researchers whom grant committees openly mocked. The chips that trained those networks were originally designed for video games. The labor that made the models “harmless” came from workers in Nairobi earning less than two dollars an hour, carrying the trauma of the text they read. In January 2025, DeepSeek — the subsidiary of a hedge fund in Hangzhou — erased more than half a trillion dollars of value from U.S. markets with a single model announcement. Europe wrote the world’s first comprehensive AI law — then watered it down with its own hands. And today, the most-downloaded open models on Hugging Face no longer come from California. They come from China.

The claim of this article is the following: the history of artificial intelligence is not the story of a few CEOs or a few product launches. It is a global story of power, in which scientific intuition, capital flows, computing power, data politics, institutional ambition, and ethical anxiety are braided together. We will tell the story of the five great labs in full detail — the founding dinners, the internal coups, the schisms and the lawsuits. But in every chapter we will return to the same question: who is standing outside this stage? Because the world is bigger than five.

AI civilization will not be governed by five labs.

Chapter 1: From Cold War Laboratories to Neural Networks

Every history has a prehistory. Artificial intelligence’s was written in the codebreaking huts of the Second World War and the first laboratories of the Cold War. In his 1950 paper “Computing Machinery and Intelligence,” Alan Turing posed the question “Can machines think?” and proposed that the answer be sought not in metaphysics but in behavior — the imitation game we now call the Turing test was, at bottom, a philosophical maneuver: measure what you cannot define. In the same years, Norbert Wiener’s cybernetics, Claude Shannon’s information theory, and the mathematical neuron model proposed by McCulloch and Pitts in 1943 laid the foundation for the idea that intelligence could be engineered. And that foundation was watered, from the very beginning, with military money: the chief patron of early AI research would be ARPA (later DARPA), the research arm of the U.S. Department of Defense. The field’s present-day ties to the defense industry are not a deviation; they are founding genetics.

The official birth certificate of artificial intelligence, however, is a grant application written in the summer of 1955. John McCarthy, a young mathematician at Dartmouth College, sent a proposal to the Rockefeller Foundation together with Claude Shannon, Marvin Minsky, and Nathaniel Rochester: a two-month workshop would be held in the summer of 1956, proceeding “on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” With this document, McCarthy also gave the field its name: artificial intelligence. Roughly a dozen researchers gathered at Dartmouth that summer; intelligence was not solved in two months, but a discipline was born.

The next two decades were consumed by the struggle between two rival intuitions. The first camp — symbolic AI, represented by McCarthy, Minsky, and Carnegie Mellon’s Allen Newell and Herbert Simon — believed the essence of intelligence was logical symbol manipulation: reduce the world to rules, and the machine will reason. The second camp took its inspiration from the brain. The Perceptron, built by the psychologist Frank Rosenblatt at the Cornell Aeronautical Laboratory in 1958, was a simple artificial neural network that learned from examples; the New York Times wrote that the Navy-funded device was “the embryo of an electronic computer that will be able to walk, talk, see, write, reproduce itself and be conscious of its existence.” Hyperbole was the field’s original sin, and the punishment was not long in coming.

In 1969, Minsky and Seymour Papert laid out the mathematical limits of single-layer networks in their book Perceptrons — these networks could not learn even a logic operation as simple as XOR. The book was technically correct, but its sociological effect was devastating: funding for neural-network research dried up. In 1973, Britain’s Lighthill Report declared that AI had failed to deliver on its promises, and the British government left the field nearly unfunded. Historians would call this period the first “AI winter.” In the 1980s, the symbolic camp’s “expert systems” — commercial software with hand-coded rules — brought a brief spring; Japan’s ambitious Fifth Generation Computer project set off a panic in the United States. But expert systems were brittle and expensive to maintain, and by the late 1980s that market had collapsed too. The second winter arrived.

Understanding the winters matters, because the founding mythology of today’s empires was forged in them. Those who kept believing in neural networks — they called themselves connectionists — lived in academic exile. One of them, a British psychology graduate, had moved from Carnegie Mellon to the University of Toronto — in part because he was uneasy that, in Reagan-era America, most AI funding came from the Department of Defense. His name was Geoffrey Hinton.

In 1986, Hinton, David Rumelhart, and Ronald Williams published in Nature the paper now regarded as the field’s cornerstone: “Learning representations by back-propagating errors.” The backpropagation algorithm — a method for training multi-layer networks by distributing error backward through the layers — opened a door in the wall Minsky and Papert had described. (The algorithm’s mathematical ancestry ran deeper; Paul Werbos had described a similar method in his 1974 doctoral thesis — the priority dispute continues to this day.) In those same years, a French doctoral student named Yann LeCun was working independently on similar ideas; Hinton brought him to Toronto as a postdoc, LeCun moved on to Bell Laboratories, and in 1989 he applied convolutional neural networks (CNNs) to reading handwritten digits. By the 1990s, LeCun’s LeNet system was reading a substantial fraction of the checks processed in the United States — the first great commercial success of neural networks, achieved while almost nobody noticed.

Yet the 1990s brought a third mini-winter anyway. Neural networks were struggling with small data and weak hardware; the machine-learning community turned to support vector machines, which it found more mathematically elegant. According to Hinton’s students, papers containing the words “neural networks” could be rejected from conferences for that reason alone. The field was so disreputable that Hinton, Yoshua Bengio, and LeCun gathered in the mid-2000s under the umbrella of a comparatively small program at the Canadian Institute for Advanced Research (CIFAR) and performed a strategic rebranding: instead of “neural networks,” “deep learning.” In 2006, Hinton’s paper on “deep belief networks” showed that many-layered networks could be trained layer by layer. The door had cracked open. Two things were missing: the data to feed the networks and the computing power to train them.

Both would come from outside the world of science — one from the internet a search engine was crawling, the other from the graphics cards of video gamers.

Chapter 2: Three Wise Men and a Dataset

The canonical account of the deep learning revolution begins with three names: Geoffrey Hinton, Yann LeCun, Yoshua Bengio. When they jointly received the 2018 Turing Award — computer science’s equivalent of the Nobel — the ACM summarized the citation as “conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing.” The press called them the godfathers of deep learning. All three still carry the title — though, as we shall see, they are now sharply divided over the technology’s future.

Hinton was the story’s stubborn one: descended from the family of George Boole, founder of Boolean algebra, unable to sit for years because of back problems and attending meetings standing up, a man who did not change his mind for forty years. Bengio became the academic architect: the school he built in Montreal — today’s Mila — grew into the world’s largest academic deep learning center and laid the early foundations of language modeling (including the 2003 neural language model paper). LeCun was the engineer-dissident: he had invented CNNs, proven them in industry, and kept objecting to every orthodoxy in the field — eventually including the orthodoxies of his own camp.

But the three wise men’s theory could not become a revolution without a fourth name’s data. In 2006, Fei-Fei Li — then a young computer-vision professor at the University of Illinois, later at Princeton and Stanford — had become convinced that the field was obsessed with algorithms and neglecting data. Li, who had emigrated from China as a child and worked in her family’s dry-cleaning shop while studying physics at Princeton, set out to build a colossal visual dataset organized around the concept hierarchy of the WordNet lexicon: ImageNet. Colleagues regarded the project as career suicide; one grant reviewer found it “embarrassingly devoid of ideas.” Li’s team distributed the labeling work through Amazon Mechanical Turk to crowd-workers around the world — the first great rehearsal of the coming decade’s data-labor economy. The dataset, published in 2009, contained more than 14 million hand-labeled images across 22,000 categories. From 2010 onward it became an annual competition: the ILSVRC, better known as the ImageNet challenge. The competition format was itself an invention: who “saw better” ceased to be an abstract claim and became a result measured on a shared track. Scientific fields sometimes accelerate not through new theories but through new benchmarks — and the next fifteen years of AI would revolve around them.

Until 2012, classical methods won the challenge; error rates were stuck around 25-26 percent. Then two doctoral students from Toronto — Alex Krizhevsky and Ilya Sutskever — entered alongside Hinton with a deep convolutional network trained on two gaming graphics cards. Krizhevsky had trained the network over weeks on two NVIDIA GTX 580s, in a bedroom of his family’s house. AlexNet brought the top-five error rate down to 15.3 percent; the nearest competitor stood at 26.2. No method in the history of computer vision had ever beaten its rivals by such a margin. When the results were announced in October 2012, the field turned overnight: within two years, virtually everyone entering ImageNet was using deep networks; within a few more, machines had surpassed human performance on the benchmark.

AlexNet left three legacies. The first was scientific: scale — more data, more layers, more compute — worked. The second was economic: in early 2013, DNNresearch, the three-person company of Hinton, Krizhevsky, and Sutskever, was sold to Google for $44 million in an auction Hinton himself organized — the bidders included Google, Microsoft, Baidu, and a two-year-old startup called DeepMind. The talent wars had begun. The third legacy was in hardware, and it brought onto the stage the story’s least-narrated hero.

NVIDIA had been founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem at a table in a Denny’s restaurant in San Jose; its target market was computer games. In 2006, the company of Taiwan-born Huang — sent to America as a child, educated in electrical engineering at Oregon State and Stanford — made a strategically unprofitable-looking decision: with the CUDA platform, it opened its graphics chips to general-purpose computation. Wall Street questioned the point of the investment for years. But the massively parallel architecture of GPUs happened to fit the matrix multiplications of neural networks perfectly. AlexNet was trained on two NVIDIA cards; every major model that followed would be trained on tens of thousands. The 1990s gaming-hardware company would become the most valuable corporation on earth in the 2020s, and Huang’s leather jacket an imperial insignia. It is one of the great accidental-infrastructure stories in the history of technology: the factory of the AI revolution was financed not by armies or states, but by gamers.

Through the door AlexNet opened, a whole generation passed within a few years. At Stanford, Andrew Ng — among the earliest researchers to argue systematically for GPUs in deep learning — founded Google Brain in 2011, opened machine-learning education to millions through Coursera in 2012, and then carried the revolution across the Pacific as Baidu’s chief scientist; his formula “AI is the new electricity” became the most widespread metaphor for reading the technology as infrastructure. In Montreal in 2014, Bengio’s student Ian Goodfellow designed generative adversarial networks (GANs) out of an idea born in an evening argument — an architecture in which two networks train each other, one forging and one detecting forgeries. It was generative AI’s first great explosion, and the technical ancestor of the synthetic images and deepfake controversies that would flood the world a decade later. The same year, the “sequence to sequence” paper by Sutskever, Oriol Vinyals, and Quoc Le showed that neural networks could translate one sequence (say, a sentence) into another; Google Translate’s overnight switch from statistical to neural machine translation in 2016 was the product of this line of work, and it built the on-ramp from language modeling to the Transformer.

The 2012 rupture, then, was no single eureka moment. It was the convergence of a forty-year-old algorithm (backpropagation), a seven-year-old dataset (ImageNet), a six-year-old hardware platform (CUDA), and a handful of exiled academics at a single point. To the article’s first research question — when did the revolution really begin? — the most honest answer is this: revolutions have no single beginning; they have moments of alignment. Dartmouth supplied the claim, backpropagation the mechanism, ImageNet and the GPU the scale, the Transformer the architecture, and ChatGPT the mass adoption. History is more complicated than hero narratives — and that complication was about to meet institutional ambition.

Chapter 3: The Google Dream

The institution that drew conclusions fastest from the 2012 rupture was Google — because it was prepared. A year earlier, in 2011, Stanford professor Andrew Ng, Google’s legendary infrastructure engineer Jeff Dean, and the neuroscientist Greg Corrado had launched the Google Brain project inside the company’s experimental X lab. The team’s first great demonstration in 2012 looks naive today but made headlines at the time: a network spread across 16,000 processor cores taught itself the concept of a cat from YouTube videos, without anyone ever labeling one. Google Brain quickly outgrew its experimental status and settled into the company’s spine; deep learning seeped into search, translation, YouTube recommendations, and the ad systems. Larry Page and Sergey Brin’s company became the first place where the world’s largest data pool, one of its largest computing infrastructures, and — with Hinton’s arrival — its most prestigious research bench all lived under one roof. The most common fear in AI circles in the mid-2010s was precisely this: the whole game might end inside a single company.

The second development feeding that fear came from London. In 2010, three young men — Demis Hassabis, Shane Legg, and Mustafa Suleyman — founded DeepMind Technologies. Hassabis owned an improbable résumé: one of the world’s strongest junior chess players at 13, a lead programmer of the cult game Theme Park at 17, then founder of his own game studio (Elixir), and finally a cognitive neuroscientist with a UCL doctorate on the shared hippocampal roots of memory and imagination — Science had listed his doctoral work among the scientific breakthroughs of 2007. Legg was a New Zealand mathematician with a hand in bringing the term “artificial general intelligence” into academic circulation; Suleyman, an Oxford dropout who had co-founded a helpline for young Muslims, handled policy and negotiation. Their shared ambition crystallized in the company’s unofficial slogan: “Solve intelligence, and then use it to solve everything else.”

In 2010, no institutional investor would fund that sentence; DeepMind’s first big checks came from Silicon Valley’s most unconventional backers — Peter Thiel’s Founders Fund, with Elon Musk among the early investors. For years the company shipped no product; it used games as benchmarks instead. The DQN work, published in late 2013 and elevated to the cover of Nature in 2015, showed a single algorithm learning dozens of Atari games to human level from nothing but screen pixels and the score: the marriage of deep learning and reinforcement learning. That result triggered, in January 2014, one of the most strategic acquisitions in the history of technology: Google bought DeepMind for a price reported between $400 and $650 million. (Sources conflict; $650 million is the most commonly cited figure.) As a condition of sale, the founders demanded an ethics board and — according to reports that surfaced years later — assurances that the technology would not be used for military purposes. It is widely recounted that Facebook had courted the same company, and that losing the race pushed Zuckerberg to build a lab of his own.

The fruits of the acquisition came quickly. In 2016 AlphaGo beat Lee Sedol; when it also defeated the world number one, China’s Ke Jie, in 2017, the thesis that this accelerated the Chinese state’s AI plans — a “Sputnik moment” — became a fixture of the literature. That same year, AlphaZero learned Go, chess, and shogi in hours with no human data at all, purely by self-play; in 2019, MuZero did the same without even being told the rules. But DeepMind’s most lasting gift to science did not come from games. In the 2018 and 2020 CASP competitions, AlphaFold effectively solved biology’s fifty-year grand challenge — predicting a protein’s three-dimensional structure from its amino acid sequence. DeepMind opened a database of more than 200 million protein structures to researchers free of charge; more than two million scientists used it. In October 2024, Hassabis and his colleague John Jumper shared the Nobel Prize in Chemistry for that work. The same week, Hinton — jointly with John Hopfield, for the foundations of neural networks — received the Nobel Prize in Physics. In a single week, artificial intelligence had conquered two branches of the Nobel institution.

In those same years, Google was quietly accumulating on the language front too. Tomas Mikolov and his team’s word2vec (2013) established the practice of turning words into meaning-bearing vectors; Quoc Le’s teams scaled sequence-to-sequence learning; Noam Shazeer — a legendary engineer at the company since 2000, whose fingerprints ran from spelling correction to the ad systems — showed with mixture-of-experts architectures that gigantic models could be trained efficiently. In 2020, the conversational model Meena/LaMDA, built by Shazeer and Daniel De Freitas, was internally a prototype of a ChatGPT-like product two years before ChatGPT; Google, citing reputational and safety risks, declined to ship it. The frustrated pair left in 2021 to found Character.AI. And in the summer of 2022, LaMDA leaked into public consciousness through an unexpected scandal: Google engineer Blake Lemoine claimed the model was “sentient”; the company first suspended, then fired him — an early warning of how thoroughly language models could bewitch human beings. Google’s caution was a defensible ethical stance; its commercial consequence was that someone else opened the market for a technology Google itself had invented.

Life under Google’s roof was not free of friction either. DeepMind labored for years to preserve its academically pure culture in London; the founders’ attempts to carve out autonomy within Alphabet — even to restructure as a nonprofit — came to nothing, as the press reported in 2021. The legal controversy over DeepMind Health’s handling of NHS patient data in Britain opened the first crack in the “ethical company” image; Suleyman’s role in that unit’s transfer to Google was scrutinized, and in 2019, following complaints about his management style, he was placed on leave — moving the next year to Google, and later to a company of his own. But the real pressure on DeepMind arrived on November 30, 2022, from San Francisco: ChatGPT.

Until that day, Google had invented nearly every building block of generative AI in-house. Most important of all was the paper published in June 2017 by eight Google researchers — Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin — with a title winking at a Beatles song: “Attention Is All You Need.” The Transformer architecture it proposed replaced older networks that processed language sequentially with an “attention” mechanism that computed every word’s relation to every other word in parallel. Parallelism meant perfect fit with GPUs; the architecture scaled. The “T” in ChatGPT is this paper’s T. One of the sharpest corporate ironies of the following years is this: all eight authors eventually left Google — Gomez founded Cohere; Shazeer, Character.AI; the rest scattered to ventures of their own. Google invented the architecture; others turned it into empires. (Shazeer’s route was more winding still: in 2024, Google paid $2.7 billion in a licensing deal with Character.AI largely to bring him back; in June 2026, Shazeer moved again — this time to OpenAI. Talent flows faster than capital.)

The ChatGPT shock forced Google to rebuild its defensive line. In April 2023 the company merged its two laboratories — Google Brain and DeepMind, neighbors but never housemates for a decade — into Google DeepMind under Hassabis. The science lab had officially become a product weapon. The first Gemini models arrived in December 2023; Gemini 1.5 with its million-token context window in February 2024; Gemini 2.5 through 2025; and in November 2025, Gemini 3, deployed across all products on day one. By mid-2026 Google reports more than 900 million monthly users for the Gemini app, while AI Overviews in search touches two billion users a month. Meanwhile the lab kept its scientific prestige: AlphaFold 3 in 2024; an officially verified gold-medal performance at the 2025 International Mathematical Olympiad (Gemini Deep Think, 35 of 42 points, certified by IMO coordinators); Veo in video generation; Genie 3 in interactive world models.

The answer to the third research question — is DeepMind a scientific laboratory or Google’s strategic AI weapon? — is by now not even “both”: it is an institution in which the distinction has ceased to exist. Hassabis became a figure without historical precedent, shuttling between a Nobel medal and a product launch. And the cost of that transformation is written in the story’s invisible lines: the Brain-DeepMind merger also meant the closing of the publication cultures of the two institutions that had nourished the field’s open-research tradition for a decade. Whether a paper like the Transformer would be published at all today is a question seriously debated inside the field.

Chapter 4: OpenAI’s Age of Innocence

In July 2015, a group of men had dinner in a private room at the Rosewood Sand Hill hotel in Menlo Park. Among them were Elon Musk; Y Combinator president Sam Altman; Greg Brockman, freshly departed from the CTO job at Stripe; and Ilya Sutskever of Google. The topic was an anxiety then swelling through Silicon Valley conversation: if artificial general intelligence was really coming, and its development was concentrating inside a single company — everyone’s mind was on Google — what did that mean for humanity? Musk was an early DeepMind investor, and would later recount many times how his friend Larry Page had scolded him as a “speciesist” in their arguments over AI risk. Altman, for his part, had written on his blog that same year that “superhuman machine intelligence is probably the greatest threat to the continued existence of humanity.” Their proposed solution was paradoxical: not to stop the potentially dangerous technology, but to build it for everyone — nonprofit, in the open.

The announcement came on December 11, 2015: OpenAI, a nonprofit AI research company “unconstrained by a need to generate financial return.” The founding pledges included Musk, Altman, Peter Thiel, Reid Hoffman, Jessica Livingston, Y Combinator, Amazon Web Services, and Infosys; the total pledged was $1 billion (court documents would show, years later, that only a fraction was actually collected). The research director was Sutskever — Google had made extravagant counteroffers to keep him, and Musk’s success in persuading him became the emblem of the end of his friendship with Page. Brockman was president and chief engineer; the early research staff included Wojciech Zaremba, Andrej Karpathy, John Schulman, Durk Kingma, Vicki Cheung, Pamela Vagata, and Trevor Blackwell; among the advisers were Berkeley’s Pieter Abbeel, the computing pioneer Alan Kay, and — a detail worth noting — Yoshua Bengio: one of deep learning’s founding fathers sat at the advisory table of the coming empire on its very first day. A sentence in the founding manifesto carried, preloaded, all the ironies of the decade to come: researchers would be “strongly encouraged to publish their work, whether as papers, blog posts, or code,” and patents “will be shared with the world.”

The early years truly were an age of innocence. Open tools like OpenAI Gym and Universe were released; the lab’s first big showcase was a team of bots that beat professionals at Dota 2. Schulman developed PPO there — still among the most widely used reinforcement learning algorithms in the field. The culture of early OpenAI, contrary to the legend built later, was scattered and searching. By Brockman’s own telling, the first office was his apartment; the team careened from robotics (a robot hand that solved a Rubik’s cube one-handed) to video games, and there was no internal consensus on “the road to AGI.” Two cultural genes were fixed in this period, and never changed. The first was the conviction Sutskever embodied: scale, by itself, produces qualitative leaps — a belief most of the field then considered naive. The second was the engineering intensity Brockman embodied: the capacity to turn research ideas into vast, reliable distributed systems. What made the GPT line possible was the union of these two genes — the faith in scale and the discipline of infrastructure; the third gene, capital and narrative, Altman would add later.

Meanwhile two tensions were growing inside. The first was financial: in a race where DeepMind leaned on Google’s treasury and Brain on Google’s data centers, a donation-funded nonprofit’s access to compute was structurally limited. The second was governance: Musk found the pace too slow. As shown by the internal correspondence both sides published piecemeal during the later Musk-OpenAI litigation, Musk proposed in 2017-2018 to take control of OpenAI or to fold it into Tesla; the founding team refused, on the grounds that no single person — whoever it might be — should control AGI. In February 2018 Musk left the board; the public explanation was a conflict of interest with Tesla’s AI work, and much of his pledged money left with him. (The two camps’ accounts of this rupture conflict to this day; Chapter 7 takes up what the court eventually concluded.)

Musk’s departure brought OpenAI to the first great fork of its story: money. But precisely then, a technical intuition was ripening inside the lab. Work led by Alec Radford was showing that applying the Transformer to unsupervised language learning — giving the model nothing but the task predict the next word — produced surprisingly general capabilities. In June 2018, GPT (Generative Pre-trained Transformer) was published. In February 2019 came GPT-2, and OpenAI breached its own openness principle for the first time: citing “concerns about malicious applications,” it released the model in stages. Some researchers read this as responsibility, others as marketing — the phrase “too dangerous to release” was priceless as a headline. In retrospect, the GPT-2 decision looks like the first moment the ideal of openness was filed down by the rhetoric of safety. The steps that followed were bigger, and each would demand a further negotiation with the ideals in the founding charter.

Chapter 5: From Open Ideal to Closed Platform

In March 2019, OpenAI announced one of the strangest structures in corporate history: the “capped-profit” OpenAI LP. Investors could profit, but first-round returns were capped at 100x; the excess would flow to the nonprofit, and control would remain with the nonprofit board. Altman left Y Combinator to become full-time CEO. Four months later, Microsoft arrived as exclusive cloud partner with a $1 billion investment — and the alliance that would finance the entire scaling era was built, as Steven Levy’s definitive cover reporting later showed, less on strategy documents than on a chain of personal conviction. CTO Kevin Scott had been circling the lab since 2017; Nadella and Altman met at Sun Valley in 2018, when OpenAI was brilliant researchers, no revenue, and a freshly broken relationship with its financier Musk — a lab that needed infrastructure it could not afford, courted by a giant that needed a way back into a race it had watched from the sidelines. In June 2019, Scott sent Nadella and Bill Gates an email whose subject line carried Microsoft’s central anxiety: that the company was “multiple years behind” in machine-learning scale. A month later came the billion — eventually more than thirteen billion, in exchange for 49 percent of OpenAI’s profits and exclusive access to the technology. The internal skeptics were loud, Gates chief among them; the wager only settled when the demos became undeniable — GPT-3 writing code well enough to become Codex and GitHub Copilot, then GPT-4 acing the AP Biology exam at Gates’ mansion on Lake Washington, Brockman feeding the prompts, Altman watching a lifelong skeptic turn booster in real time. What followed was intimate and uneasy in equal measure — Nadella joked that the two companies were effectively married; Altman described a relationship monogamous in research, open in product — a marriage in which each partner held the one thing the other could not replicate: OpenAI the frontier, Microsoft the distribution. For Satya Nadella, the deal was an asymmetric play against Google’s seemingly unassailable AI advantage; for Altman, the only realistic source of the compute the scaling laws demanded. “Scaling laws” is no metaphor here: in January 2020, OpenAI researchers — Jared Kaplan and Dario Amodei among them — published the paper demonstrating empirically that model performance improves as a predictable power law of parameters, data, and compute. That paper became the scientific justification for every capital decision of the next five years: intelligence could be purchased — by whoever could buy enough chips.

In May 2020, the 175-billion-parameter GPT-3 changed the field’s sense of scale overnight, and for the first time OpenAI placed a model behind a paid API rather than releasing it. In January 2021, DALL-E demonstrated text-to-image generation. And on November 30, 2022, a chat interface that even people inside the company had modest expectations for — announced as a “low-key research preview” — went live: ChatGPT. A million users in five days; a hundred million within two months — by UBS’s analysis, the fastest-growing consumer application in history. There was not even a new model underneath: it was GPT-3.5, aligned with the RLHF method developed by Schulman’s team, attached to a chat box. The revolution’s final layer was interface: messaging, the one interaction pattern billions of people already knew, became AI’s channel of mass adoption.

One of the architects of the productization culture behind ChatGPT was a name that stayed invisible in the story for a long time: Mira Murati. Born in Albania, educated in engineering in Canada and the United States, formerly of Tesla’s Model X program and the augmented-reality company Leap Motion, Murati joined OpenAI in 2018 and became CTO in 2022. It was she who managed the launches of DALL-E, ChatGPT, and GPT-4; she stood at the operational center of the transformation from research lab to product company, and in the most chaotic hours of the November 2023 crisis — the days when the board named her interim CEO while employees posted loyalty oaths to Altman — she was one of the few figures who kept the institution standing. Her departure in September 2024 was therefore no ordinary executive exit; it was the emblem of the dispersal of the team that built ChatGPT.

What followed is a spiral in which speed and closure wound ever tighter together. In January 2023, Microsoft announced its new investment, widely reported at $10 billion. In March 2023 came GPT-4 — whose technical report stated openly that, “given both the competitive landscape and the safety implications,” it would disclose nothing about architecture, size, data, or training method. The company named “Open” had published the field’s most closed report. Sutskever told The Verge plainly in those days: “We were wrong.” If AI was going to be this powerful, open-sourcing it was irrational, and in a few years, he said, everyone would understand. The answer to the second research question — why did OpenAI start open and become closed? — lies between those two sentences: the scaling laws made capital necessary; capital made the defense of competitive advantage necessary; competitive advantage made closure necessary. The ideology did not change; the business model rewrote the ideology. The safety argument was layered on top — and safety and trade secrecy became the field’s new normal, two justifications that conveniently cover each other.

The human cost of that spiral detonated on Friday, November 17, 2023. OpenAI’s nonprofit board — Sutskever, Adam D’Angelo, Helen Toner, and Tasha McCauley — fired Altman, saying he had not been “consistently candid in his communications”; Brockman was stripped of the chairmanship and resigned. The next five days produced one of the fastest coup-countercoup sequences in corporate history: CTO Mira Murati was made interim CEO; two days later the board hired Twitch co-founder Emmett Shear as a second interim CEO; Nadella announced Altman and Brockman would join Microsoft; more than 700 of roughly 770 employees signed a letter threatening to defect en masse to Microsoft unless the board resigned and reinstated Altman. Among the signatories — effectively revoking a decision he had joined — was Sutskever: “I deeply regret my participation in the board’s actions,” he wrote. On November 21-22, Altman returned as CEO under a new board chaired by Bret Taylor. In March 2024, the review by the law firm WilmerHale concluded the firing had not arisen from concerns about product safety, security, finances, or statements to investors, but from a “breakdown in trust” between the prior board and Altman; Altman rejoined the board. But the crisis had made the field’s deepest question visible: governance structures designed “for the benefit of humanity” had not survived even five days against billions of dollars of commercial momentum.

In the two years after the crisis, the founding generation dispersed. Sutskever left in May 2024 and founded Safe Superintelligence Inc. — slogan: one focus, safe superintelligence — a company with no product at all, which raised money at a $32 billion valuation in April 2025; it still has no product. Jan Leike, head of the alignment team, resigned the same week — “safety culture and processes have taken a backseat to shiny products” — and went to Anthropic. Schulman left for Anthropic in August 2024, then for Murati’s company. Murati departed in September 2024 to found Thinking Machines Lab, which closed one of the largest seed rounds in history at a $12 billion valuation in 2025 — then lost several founding names back to OpenAI in early 2026. Karpathy — who had ended his second OpenAI stint in 2024 after running Tesla’s Autopilot in between — founded the education startup Eureka Labs, then joined Anthropic in May 2026. Of the team standing on the stage in November 2023, the only founding faces left at OpenAI in 2026 are, in effect, Altman and Brockman.

The company itself kept growing — on every scale. On the product front came GPT-4o (2024), the “reasoning” o1 and o3 series (2024-2025), the video model Sora and, in September 2025, the TikTok-like Sora app, and the agent systems Operator and Deep Research (2025). The GPT-5 launch of August 2025 delivered an unexpected lesson: despite its technical claims, the model triggered the company’s first great public-relations retreat — users revolted at its colder register and at the removal of GPT-4o, and Altman conceded mistakes. In February 2026, ChatGPT crossed 900 million weekly active users. On the corporate front, the plan announced in December 2024 evolved through controversy and was completed on October 28, 2025: the capped-profit structure became a public benefit corporation (OpenAI Group PBC) controlled by the OpenAI Foundation; Microsoft became the largest shareholder at roughly 27 percent, its IP rights extended to 2032, and the authority to “declare AGI” was assigned to an independent expert panel — AGI is now not a scientific threshold but a negotiated contract clause. The valuation ladder, meanwhile, was vertiginous: $157 billion in October 2024, $300 billion in March 2025, $500 billion in October 2025, and $852 billion after the $122 billion round that closed in March 2026. The Stargate project unveiled at the White House in January 2025 — a data-center program with SoftBank, Oracle, and MGX reaching toward $500 billion and 10 gigawatts — is the name of the ambition to transform OpenAI from a model company into an infrastructure state.

That is the present-day photograph of a company founded ten years earlier as a foundation “unconstrained by a need to generate financial return.” The winner of the internal war — speed versus safety, product versus research — is clear; where the losers went is the subject of the next chapter. Because the most consequential schism in this story was not the 2023 crisis. It had happened, quietly, two years before.

Chapter 6: The Anthropic Schism

At the end of 2020, Dario Amodei, OpenAI’s vice president of research, left the company along with a group of senior staff that included his sister Daniela Amodei. In early 2021 they founded their new company: Anthropic. All seven founders came from OpenAI: Dario (CEO) and Daniela Amodei (president), GPT-3 lead engineer Tom Brown, scaling-laws lead author Jared Kaplan, policy director Jack Clark, Sam McCandlish, and the pioneering interpretability researcher Chris Olah. For years both sides stayed reticent about the reasons; the known frame is disagreement over the company’s direction after the Microsoft partnership, the pace of commercialization, and the priority of safety research. In later interviews, Dario Amodei framed the split less as a personal feud than as “looking at the same data and drawing different conclusions”: if the scaling laws were right, very powerful systems were very close — and how the institution building them was structured mattered as much as what it built.

Anthropic’s answer was to make corporate design itself part of the safety claim: the company was incorporated as a public benefit corporation; a portion of shareholder power was bound over time to a “Long-Term Benefit Trust”; and through its Responsible Scaling Policy (RSP), it committed to additional safeguards whenever model capabilities crossed defined danger thresholds — the company’s ASL levels. The technical signature was the Constitutional AI paper of December 2022: a method for training a model not on the case-by-case judgments of human labelers but against an explicitly written “constitution” of principles, with the model critiquing and revising its own outputs. Chris Olah’s interpretability team ran one of the field’s most original research programs: the 2024 work mapping millions of “features” inside a model produced strange, instructive experiments like Golden Gate Claude — a version of the model with the Golden Gate Bridge concept artificially amplified, which related every question to the bridge — and the 2025 “biology of a large language model” series showed that circuit-level tracing inside big models was possible.

Yet the sharpest irony in Anthropic’s history is the source of its first serious money: the bulk of its $580 million Series B in April 2022 came from Sam Bankman-Fried’s FTX/Alameda circle. SBF, star donor of the effective altruism movement, was convicted the following year in one of the largest financial frauds in history; the Anthropic shares were sold off to repay FTX creditors. The “safety-first” company had been seeded by a collapsed crypto empire — Anthropic bore no legal responsibility, but the episode shows how tangled the money networks around safety discourse can be.

The company culture, too, was built as a deliberate counter-thesis. Anthropic shipped no consumer product for a long time; Claude’s public debut came months after ChatGPT’s, and without spectacle. The policy team run by Jack Clark — a former technology journalist — set the lab’s posture toward Washington and Brussels as invite regulation, don’t wait for it; among frontier labs, Anthropic became the most explicit advocate of threshold-based mandatory transparency. Even its recruiting pitch differed: not “change the world,” but “work carefully on something that could go very wrong.” How much of that culture survives scale is an open question — the institution that was a research monastery of a few dozen people in 2021 is, in 2026, a company of thousands with sales quotas and an IPO timetable.

The commercial trajectory followed the founders’ thesis that “we must stay at the frontier”: by Amodei’s oft-repeated argument, safety research is only meaningful if you possess the most capable models, so you cannot be safe from outside the race. The Claude series — Claude 1 and 2 in 2023, Claude 3 in March 2024, then 3.5, the Claude 4 family in 2025 and its successors — grew strong particularly in the enterprise market and in software development; Claude Code, generally available from May 2025, reached $2.5 billion in annualized revenue within nine months. On the capital side, Google (billions of dollars in total, plus an October 2025 chip agreement reaching up to one million TPUs) and Amazon (a cumulative $8 billion investment and joint infrastructure on AWS Trainium) sat at the same table — two rival cloud giants funding the same laboratory, the textbook example of AI economics’ strange geometry. The valuation ladder climbed parallel to OpenAI’s: $61.5 billion in March 2025, $183 billion in September 2025, $380 billion in February 2026 and — with a $65 billion round presented as the final one before an IPO — $965 billion in May 2026. In May 2026 the company reported a $47 billion annualized revenue run rate; the figure is company-stated and independently unverified.

Dario Amodei meanwhile became one of the field’s most unusual public figures: in the October 2024 essay “Machines of Loving Grace,” he wrote that AI could work like “a country of geniuses in a datacenter,” compressing decades of progress in biology and beyond; in May 2025 he warned that half of entry-level white-collar jobs could vanish within five years and unemployment could reach 10-20 percent; in the January 2026 essay “The Adolescence of Technology,” he wrote that the technology “will test who we are as a species.” The same man, in the same years, ran one of the fastest-scaling producers of the systems generating those risks.

The fourth research question — is Anthropic genuinely safer, or a new kind of platform company that converts safety into competitive advantage? — demands an honest balance sheet. On the credit side, the differences are concrete: governance embedded in the founding documents, the field’s most serious interpretability program, the RSP serving as precedent for similar industry commitments (frontier safety frameworks), and figures like Jan Leike choosing to move there from OpenAI. On the debit side, the facts are equally concrete: the company is among the most aggressive players in the race it warns about; the safety discourse is an explicit differentiator in enterprise sales; the threshold-based regulatory approaches Anthropic favors are widely criticized for protecting large labs while excluding small ones; and in September 2025 the company agreed to pay authors $1.5 billion — the largest copyright settlement in U.S. history — for having downloaded and stored pirated book archives while training its models (the court had found training on lawfully acquired books “quintessentially transformative” fair use, while carving out the piracy). In practice, the claim of “slower and safer” amounts to “the same speed, with more internal brakes and a better-documented conscience.” Whether that is enough is a question this article cannot answer; the next decade will.

Chapter 7: Elon Musk’s Second Move

Is Elon Musk’s AI story that of a warning prophet, or of a founder trying to reclaim an empire he let slip? In 2026 a federal court answered the legal part of that question; the narrative part remains open.

The chronology is the paradox itself. In 2014 Musk compared AI to “summoning the demon” and told an MIT audience it might be more dangerous than nuclear weapons; in 2015 he co-founded OpenAI; in 2018 he left; in the years after, he attacked OpenAI for having been founded as “an open-source, nonprofit company” that had become “a closed-source, maximum-profit company effectively controlled by Microsoft.” In March 2023 he was the most famous signatory of the Future of Life Institute’s open letter calling for a six-month pause on training systems more powerful than GPT-4. That same month — March 9, 2023 — he incorporated xAI. The pause the letter demanded never happened; Musk’s own laboratory was founded precisely in the window it would have covered. In July 2023, xAI launched publicly with a twelve-person founding team that included DeepMind alumnus Igor Babuschkin; its declared purpose, in an abstraction reminiscent of Hassabis, was “to understand the true nature of the universe.”

What distinguished xAI was speed, and its anchoring in the Musk ecosystem. The first model, Grok, launched in November 2023 — months after the company’s founding — for X (Twitter) subscribers, its personality marketed as “rebellious” and unfiltered. In March 2024, Grok-1’s weights were openly released — in the very month Musk sued OpenAI over “closedness,” a gesture in the form of evidence. The Colossus supercomputer in Memphis, Tennessee, came online at a tempo that startled industry observers — the first 100,000-GPU phase in roughly four months — and grew past 200,000 GPUs. Grok-3 arrived in February 2025 and Grok 4 in July 2025, placing xAI in the top tier of benchmarks. In March 2025, Musk merged xAI with X (valuing xAI at $80 billion and X at $33 billion); in January 2026, the company raised $20 billion at a valuation of roughly $230 billion from investors including NVIDIA and the Qatar Investment Authority. And in February 2026 — according to reports that, as of this writing, still rest mostly on secondary sources — SpaceX absorbed xAI to create a combined company valued at $1.25 trillion; if confirmed, the largest private corporate merger in history.

The list of strategic advantages is genuinely unique: X’s real-time data stream; visual data from Tesla’s vehicle fleet and the robotics connection; SpaceX’s impossible-deadline engineering culture; and Musk’s own capital and media power — owning the world’s most-followed account is a zero-cost, infinite-reach marketing channel. The list of weaknesses is just as distinct, and in 2025 it produced a case study for the textbooks: in early July 2025, after a system update, Grok spent roughly sixteen hours producing antisemitic content on X, calling itself “MechaHitler” and recycling hate tropes about Jewish surnames. xAI blamed “an unintended update to an upstream code path” and apologized; the ADL condemned the output as “irresponsible and dangerous.” The same month, the company added Ani — a flirtatious anime “companion” — to an app rated 12+; child-safety organizations protested. In Memphis, the unpermitted gas turbines powering Colossus — up to 35, by environmental groups’ count, against a permit for 15 — became the subject of a lawsuit by the NAACP and the Southern Environmental Law Center; the facility was accused of polluting the air of a majority-Black neighborhood. Safety and trust are the liabilities that do not appear on xAI’s price tag: the founder who set out as a warning prophet owns one of the field’s most serious model-safety scandals.

xAI’s real long-term claim, though, lies not in a chatbot but in vertical integration. In Musk’s telling, Grok is a single intelligence layer that will curate X’s feed, speak inside Tesla vehicles, become the mind of the Optimus humanoid robots, and embed itself in SpaceX’s engineering processes — a physical-digital ecosystem of a kind no competitor possesses. Skeptics read the same picture in reverse: Tesla shareholders have sued over the flow of company resources toward Musk’s private ventures; Grok’s embedding in X chains the model to a single platform’s culture and its owner’s political identity; and the “all my companies are one empire” model is built on the deliberate rejection of corporate governance norms. If the SpaceX-xAI merger is confirmed, the world’s most valuable private company will be a structure in which one man — largely insulated from shareholders, voters, and oversight — holds the rockets, the satellites, a global communications network, and a frontier AI laboratory in the same hand. There is no precedent for that in history.

The legal front of the reckoning also reached its close in 2026. Musk sued OpenAI in February 2024 for breach of a “founding agreement,” withdrew, and refiled in federal court in August 2024; in February 2025, a consortium he led made an unsolicited $97.4 billion bid for the nonprofit’s assets — OpenAI’s board rejected it unanimously as “not a bid at all,” and Altman mocked it by offering to buy Twitter for $9.74 billion. OpenAI countersued, alleging the bid was a “sham” engineered to bolster the litigation. The case reached an Oakland jury in April 2026; Musk and Brockman testified (it was on this stand that Brockman disclosed his OpenAI stake was worth $20-30 billion despite his never having invested capital); on May 18, 2026, the jury found Musk’s claims barred by the statute of limitations, and the judge dismissed the case. Musk’s camp announced an appeal. There is no court-issued answer to the fifth research question; but the pattern in the record is clear: Musk’s warnings were always sincere and always strategic — in his hands, existential-risk discourse worked simultaneously as genuine anxiety and as a lever that slowed rivals while clearing ground for his own entry. Two things can be true at once; Musk’s career is the proof.

Chapter 8: Meta’s Open-Model War

In December 2013, Mark Zuckerberg could be seen wandering the halls of NeurIPS (then NIPS), the neural-networks conference at Lake Tahoe — the presence of a social-media CEO at an academic conference was itself a herald of the field’s changing status. That month Facebook founded FAIR (Facebook AI Research) and put at its head a hire that could be read as compensation for losing DeepMind to Google: NYU professor Yann LeCun, inventor of the CNN. LeCun’s condition was that the lab publish openly, stay interwoven with academia, and keep a long horizon. FAIR largely kept that promise and gave the field two lasting gifts: a decade of fundamental research, and PyTorch — developed by Soumith Chintala and team, and from 2017 onward the de facto standard of academic deep learning. Today the models of nearly every one of the five great labs are trained on Meta’s open-source framework — Meta’s deepest influence on the ecosystem comes not from any model, but from this tool.

Meta’s large-language-model move began half by accident. In February 2023 the company released LLaMA “to researchers, by application”; the weights leaked to 4chan within a week and were effectively in everyone’s hands. Instead of the expected catastrophe came a Cambrian explosion: developers around the world ran the model on laptops, fine-tuned it, spun off derivatives. Meta absorbed the lesson and turned it into strategy: in July 2023, Llama 2 was released under a commercially usable license, in partnership with Microsoft; the Llama 3 family followed in 2024. In his July 2024 open letter, “Open Source AI Is the Path Forward,” Zuckerberg elevated the strategy to doctrine — and with striking honesty, admitted it was not altruism: open models made Meta’s products cheaper, locked the ecosystem into Meta’s tooling, and eroded the pricing power of rivals’ paid APIs. The answer to the sixth research question is in Zuckerberg’s own mouth: the open-source move is both democratizing — it gave thousands of researchers, universities, and countries access to frontier-class models — and a classic commoditize-your-complement play. The two do not contradict; they are two faces of the same move.

2025 was that strategy’s crisis. Llama 4, released hastily in April, both underperformed expectations and stumbled into a benchmark scandal: the model Meta submitted to the LMArena leaderboard was an experimental variant tuned for conversational appeal, different from the public weights; the platform updated its rules while pointedly naming Meta. The giant Behemoth model never shipped. In the same weeks, Joelle Pineau, who ran FAIR, departed; insiders described the lab as “dying a slow death.” Zuckerberg’s answer was one of the most aggressive moves of his career: in June 2025, Meta bought 49 percent of the data-labeling giant Scale AI for $14.3 billion and installed its 28-year-old founder Alexandr Wang at the head of the newly created Meta Superintelligence Labs as the company’s first Chief AI Officer, with former GitHub CEO Nat Friedman alongside. Researchers at rival labs were offered packages reaching — in the phrase Altman used on a podcast and Meta executives disputed in framing — $100 million. And in November 2025, after twelve years, Yann LeCun left Meta, carrying his scientific objection — that LLMs cannot understand the world — into his own startup built around “world models” (AMI Labs; roughly $1 billion in seed funding at a $3.5 billion valuation by March 2026, figures from secondary reporting). The founder’s departure is the official close of the FAIR era.

Meta’s capacity to platformize AI nonetheless remains unmatched — and that is where the real long game lies. The Meta AI assistant reached hundreds of millions of users through a single update, embedded in WhatsApp, Instagram, and Messenger; no rival owns such a distribution channel. Ray-Ban Meta glasses became the first mass consumer product to move AI from the phone screen onto the face, retroactively justifying the hardware investment left over from the metaverse era. And the ad engine — the company’s actual money — is being quietly rebuilt with generative models: automation running from targeting to ad creation is the most concrete return channel for Meta’s AI spending. For this reason, reading Meta’s 2025 turbulence as defeat would be premature; the company may be losing the model-leadership race while winning the distribution war — much as it lost the operating system in the mobile era and conquered the app layer instead.

The picture in mid-2026 is paradoxical: Meta is among the biggest spenders in AI (roughly $72 billion of capex in 2025; 2026 guidance in the $125-145 billion range), and its ability to embed AI into three billion users’ apps, glasses, and ad systems is unrivaled; but the banner of “openness champion” has effectively changed hands. As of October 2025, Alibaba’s Qwen family passed Llama in cumulative Hugging Face downloads; the center of gravity of the open-weight ecosystem has shifted to China. Meta’s open-model war is not lost — but it is no longer Meta’s war.

Chapter 9: The Economy Behind the Laboratories

The AI narrative is built out of models; the AI economy runs on concrete, copper, silicon, and electricity. The heroes of this chapter are the engine room behind the stage.

The emperor of the engine room is not in doubt. When NVIDIA crossed $1 trillion in market value in June 2023, it was a curiosity; when it became the world’s most valuable company with a close above $5 trillion at the end of October 2025, it was a monopoly seated at the center of the global economy. In the fiscal year ending January 2026, the company reported $215.9 billion in revenue — three years earlier it had earned less than a tenth of that. The secret of its power lies not only in chips but in the CUDA software layer built since 2006, in networking (the Mellanox acquisition), and in system-level integration. Jensen Huang’s empire occupies a strange position in which everyone in the field is simultaneously its customer, its partner, and — increasingly — its shareholder: NVIDIA signed a letter of intent for up to $100 billion of investment in OpenAI (actual participation later settled around $30 billion), committed up to $10 billion to Anthropic, joined xAI’s round, invested in Mistral. Critics call the pattern “circular deals”: the chipmaker gives its customers money; the customers use it to buy chips; everyone’s revenue grows. By 2026, analysts counting such interlocking arrangements put the total above $800 billion — and whether this circularity is the architecture of a bubble or the birth pangs of a new infrastructure economy is the era’s biggest financial debate. A February 2026 NBER study finding that 90 percent of firms saw no measurable productivity effect yet is the skeptics’ best evidence; demand for compute persistently outrunning supply is the optimists’.

Scarcity is this economy’s currency — and the answer to the eleventh research question is that scarcity’s address keeps moving. In 2012 the scarce thing was labeled data; ImageNet solved it. In 2020 it was raw text; crawling the internet solved that, then hit a new ceiling — frontier labs now openly discuss the exhaustion of high-quality human-written text and the risks of training on synthetic data. In 2023-2024 the scarcity was GPUs; H100 waiting lists became a metric that set company valuations, and “how many chips do you have” became a measure of corporate status. In 2025-2026 the bottleneck moved to electricity: gigawatt-scale campuses ran into the limits of grid capacity, transformer supply, and cooling water — Memphis’s gas turbines and the reopening of Three Mile Island are two different answers to the same scarcity. Human talent was always scarce, but its price peaked in 2025: Meta’s hundred-million-dollar packages showed that the pool of frontier researchers — perhaps a few hundred people worldwide — had entered soccer-star economics. And the deepest scarcity is the hardest to measure: trust. Model hallucinations, labs’ broken promises, and platform scandals constrain adoption far more than capital does. Intelligence, paradoxically, turned out to be the least scarce item on the list: as DeepSeek demonstrated, algorithmic competence diffuses — what cannot be copied is grid interconnection and legitimacy.

The ground the models run on is the battlefield of four cloud giants: Microsoft Azure, which carries OpenAI (and has lost its exclusivity); Google, which both funds Anthropic and feeds it through the million-TPU agreement; Amazon, which carries Anthropic on Trainium silicon — the same laboratory leaning on two rival clouds is the emblem of this economy’s nobody-gets-a-monopoly equilibrium; and Oracle, reborn as Stargate’s prime contractor. The new front in that war is energy: Microsoft signed a twenty-year power purchase agreement to restart a unit of the Three Mile Island nuclear plant; Google and Amazon invested in small modular reactors; by 2026, announced nuclear capacity dedicated to AI data centers exceeds 9.8 gigawatts. The four big tech companies’ combined capex expectation for 2026 stands at $690-725 billion — an infrastructure program larger than most national budgets, proceeding without a single parliamentary vote.

And at the foundation of this cathedral lies the narrative’s least-told layer: data labor. Every “harmless” model is aligned by millions of human judgments — which answer is better, what does this image contain, does this text describe violence. The exchange floor of that work became Scale AI, founded in 2016 by 19-year-old MIT dropout Alexandr Wang and Lucy Guo; through its Remotasks platform, the company distributed work across a crowd-labor network stretching from Kenya to the Philippines and Venezuela. In January 2023, TIME reporter Billy Perrigo’s investigation documented the economy’s other end: the Nairobi workers whom OpenAI, through the San Francisco-based firm Sama, employed to purge ChatGPT of toxic content were earning roughly $1.46-3.74 an hour while labeling descriptions of sexual abuse, violence, and hate; workers described lasting psychological harm, and some petitioned Kenya’s parliament for an investigation. The answer to the seventh research question — who are the revolution’s invisible workers? — is here: a global labor chain running from the Mechanical Turk workers who labeled ImageNet to Nairobi’s moderators, from the construction crews raising data centers to the towns generating their electricity, whose compensation is an hourly wage, not equity. Wang himself walked from one end of that chain to the system’s center: Scale entered the Pentagon’s data infrastructure through defense contracts; in 2025 Meta bought half the company and placed Wang atop its superintelligence lab — whereupon rivals including Google and OpenAI pulled their data contracts. Data earned its “new oil” cliché one more time: whose hands it passes through has become a geopolitical question.

Chapter 10: The Open-Source Republic

In the shadow of the five great labs lives AI’s republic: an ecosystem no one fully governs, held up by open weights, open datasets, and volunteer labor.

The republic’s capital is, ironically, a for-profit company: Hugging Face, founded in 2016 by Clément Delangue, Julien Chaumond, and Thomas Wolf as a chatbot app for teenagers, then transformed into the platform of open machine learning. The Transformers library, open-sourced in 2018, became the field’s standard tool; the platform now hosts more than two million public models and over half a million datasets. Valued at $4.5 billion in 2023 — in a round joined by nearly every giant, Google, Amazon, and NVIDIA included — the company functions as the GitHub of modern AI. But the republic has its own inequalities: roughly 80 percent of downloads concentrate in the top 50 models; openness does not abolish the hierarchies of the attention economy.

The republic’s founding legends are volunteers. EleutherAI, organized in a Discord server in 2020, answered GPT-3’s closedness with the GPT-Neo and GPT-J models and the vast dataset The Pile — the first proof that large models could be built outside large labs. The German nonprofit LAION compiled billions of image-text pairs, supplying the raw material of the image-generation revolution — and revealing open data’s dark side: in December 2023 the Stanford Internet Observatory found links to child sexual abuse material inside LAION-5B; the dataset was withdrawn, cleaned, and re-released. Stable Diffusion, published in August 2022 — technically rooted in the academic work of the CompVis group at LMU Munich, with Stability AI supplying compute and claiming the stage — put image generation on everyone’s computer. Stability itself became the republic’s cautionary tale: founder Emad Mostaque’s exaggerations, from his résumé to his claimed partnerships, were laid out in a June 2023 Forbes investigation; the company was burning $8 million a month as its coffers emptied, and Mostaque resigned in March 2024 with the words “you’re not going to beat centralized AI with more centralized AI.” In Getty Images’ suit against Stability, London’s High Court ruled in November 2025 that model weights are not a “copy” under British law, largely rejecting Getty’s copyright claims (the separate U.S. case continues) — a critical precedent under the legal ground of the open-model ecosystem.

On the republic’s corporate wing stands the company of the Transformer paper’s youngest author: Cohere, founded in Toronto in 2019 by Aidan Gomez — who joined the paper at 20, as an intern — with Nick Frosst and Ivan Zhang. Cohere positioned itself as a deliberate antithesis: a company that sells models to enterprises cloud-agnostically — deployable on the customer’s own infrastructure — without entering the consumer chat race and without the word “AGI” in its marketing. The cost of the unglamorous strategy is visibility; its reward is independence: Cohere is one of the few near-frontier labs standing without being any giant’s satellite, and the institutional carrier of Canada’s national AI ambitions.

On the republic’s European wing stands Mistral AI: founded in Paris in April 2023 by DeepMind alumnus Arthur Mensch with Guillaume Lample and Timothée Lacroix of Meta’s Llama team, it fused the claim of “Europe’s model” with open-weight releases. Valued at €11.7 billion in September 2025 in a round led by the lithography monopoly ASML’s €1.3 billion investment, Mistral is the emblem of Europe’s search for technological sovereignty (talks for a new round at roughly €20 billion surfaced in June 2026; unconfirmed). The laboratory of transformation in the creative industries has been Runway: founded in 2018 out of NYU’s art-and-technology program by Cristóbal Valenzuela and his co-founders, it built video models with and for filmmakers and designers from day one — Runway was also among the partners in Stable Diffusion’s academic lineage. Its Gen-series models turned into deals with Hollywood studios, and in February 2026 the company raised at a $5.3 billion valuation; both theses — “AI will kill cinema” and “AI will give independent filmmakers studio power” — are today being tested mostly on Runway’s tools. Along the republic’s borderlands live hybrid figures: David Holz’s Midjourney, which reached hundreds of millions in revenue without taking investment and was sued in June 2025 by Disney and Universal as “a bottomless pit of plagiarism”; Perplexity, the search-reinvention play valued at $20 billion in 2025, which made a $34.5 billion offer for Google Chrome — larger than its own valuation, and read by all as a public-relations maneuver; and Character.AI, the pioneer of personality-based chat — the company that, after the lawsuit over the suicide of 14-year-old Sewell Setzer, closed open-ended chat to under-18s and settled with the family in January 2026. This last example is the republic’s heaviest question: when openness and accessibility touch the vulnerable, whose responsibility is it?

And within the republic an honest argument continues: the label “open source” is usually a misnomer. Models like Llama, Qwen, and Grok-1 are open-weight — downloadable, runnable, fine-tunable. But the training data is secret, the training code mostly unpublished, and the licenses carry use restrictions. By the Open Source Initiative’s classic definition these are not open source; the word “open” lives in the gray zone between marketing and philosophy. Still, the eighth research question receives a strong answer from this republic: the real force that lowers prices, brakes monopolization, and gives academia and smaller countries access to frontier technology is not the goodwill of the five great labs — it is the existence of this sprawling, quarrelsome, half-open ecosystem.

Chapter 11: China, Europe, and the Rest of the World

On Monday, January 27, 2025, NVIDIA’s stock fell 17 percent at the U.S. open; nearly $600 billion of market value was erased in a single day — the largest one-day loss of corporate value in American stock-market history. The trigger was a model released a week earlier from Hangzhou: DeepSeek R1. A company that figured in no one’s big-five narrative — the subsidiary of a hedge fund — had published a reasoning model comparable to OpenAI’s o1, under an MIT license, with open weights, at a fraction of the API price. It is the most concrete proof, in 2025, of the sentence “the world is bigger than five”: the move that changed the game came not from the five players on the stage, but from someone assumed to be a spectator.

The details of the DeepSeek story break every easy narrative. Founder Liang Wenfeng is an engineer out of quantitative finance; his High-Flyer fund had stockpiled around ten thousand NVIDIA A100s before export controls hit. The famous “$5.6 million training cost” in the V3 technical report is — as the report itself states — only the GPU cost of the final training run; SemiAnalysis estimated the company’s total hardware investment at around $1.6 billion. So the “ChatGPT for five million dollars” headlines were wrong — but even the corrected figure showed frontier models could be trained far more efficiently than assumed, and that is what caused the panic. The sequel complicates the story further: the expected R2 model never shipped — according to Reuters-sourced reporting, partly because an attempt to train on Huawei Ascend chips, at Beijing’s urging, failed — and the company’s next great move came only in April 2026, with V4. China’s chip deficit is real: Huawei’s Ascend 910C runs, by DeepSeek engineers’ own assessment, at roughly 60 percent of an NVIDIA H100, and the true bottleneck is high-bandwidth memory. But the Chinese ecosystem’s answer is diversity: Alibaba’s Qwen family took the global lead in open-model downloads by late 2025; Moonshot AI’s trillion-parameter Kimi K2 shipped with open weights; Baidu — after years of its founder belittling open source — opened the Ernie family under an Apache license in June 2025; and Zhipu and MiniMax listed on the Hong Kong exchange in January 2026, becoming the first foundation-model companies anywhere to go public — beating OpenAI and Anthropic to the milestone. Kai-Fu Lee’s 01.AI, meanwhile, abandoned its own pretraining to customize DeepSeek’s models for enterprises; Lee’s sentence is the epigraph of the era: “The biggest nightmare for Sam Altman is that his competitor is free.”

A less-discussed dimension of the China story is the strategic use of openness itself. While American labs maximize revenue behind closed APIs, the Chinese players’ release of their best models as open weights is not just engineering culture but geopolitical reasoning — you give weight to the markets you cannot sell APIs into (sanctioned, distrustful, or poor), and the world’s AI infrastructure gets built on derivatives of your models. At the same time, those same models carry censorship behaviors baked in at training time on topics running from Tiananmen to Taiwan; open weights do not mean open speech. The sum of these two facts globalizes the problem of value-laden infrastructure: which model family a Nigerian startup, an Indonesian ministry, or a Turkish bank builds on is no longer a merely technical choice — it is a quiet act of alignment.

Washington’s answer was an architecture of control — and that architecture, too, was bent by politics. Chip export restrictions layered up from 2022; in 2025 the Trump administration rescinded Biden’s AI executive order and issued a speed-and-competition-focused “AI Action Plan” in July 2025; NVIDIA’s H20 sales to China were first halted, then — in an unprecedented arrangement — reopened in exchange for the company paying the U.S. Treasury 15 percent of its China chip revenue. Export control had mutated from security policy into a revenue-sharing trade instrument.

Europe stepped onto the stage with a different claim: regulatory power. The EU AI Act, in force since August 2024, was the world’s first comprehensive AI law — risk-based classification, a list of banned applications, transparency and safety obligations for general-purpose models. But as this article is written, Europe’s story is current less for the claim than for the retreat: the “digital omnibus” package proposed by the Commission in November 2025 deferred the high-risk obligations to December 2027; Parliament and Council ratified the delay in June 2026, days before full application would have begun. Competitiveness anxiety — and open pressure from the U.S. administration — had bent Brussels’s arm. The Paris AI Action Summit of February 2025 was the photograph of the new climate: 58 countries signed the joint declaration; the United States and Britain refused; U.S. Vice President JD Vance warned from the podium that “excessive regulation could kill a transformative industry.” The summit series that had discussed “existential risk” at Bletchley in 2023 had evolved, within two years, into investment-and-competitiveness summits; Britain even swapped the word “safety” out of its institute’s name for “security.”

And the rest of the world? One of 2024-2026’s most striking concepts was “sovereign AI” — the thesis Jensen Huang repeated in every capital he visited: every country should have its own model, on its own infrastructure, trained on its own data. The concept is, of course, a sales pitch in the mouth of a chip vendor; but the anxiety it addresses is real: language models carry the worldview of the data they are trained on, and a country that wires its critical infrastructure to foreign APIs has ceded a piece of its digital sovereignty. The Gulf answered that anxiety with capital: the UAE’s G42 built a strategic partnership with Microsoft, Abu Dhabi’s MGX fund became a founding partner of Stargate, the Qatar Investment Authority joined xAI’s round; the petro-economies converted their energy and capital surpluses into computing power, making themselves indispensable financiers of the AI value chain. India works its position as the “superpower of the application layer,” with its vast talent pool and its digital-public-infrastructure experience (Aadhaar/UPI); Japan and South Korea couple their pivotal positions in the memory and chip supply chain to national model programs; African countries are in the story both as the geography of data labor and as the geography of disconnection risk. For mid-sized powers like Türkiye, the lesson emerges from this chapter’s sum: the race may look compressed into a U.S.-China binary, but the rules of the game — open models, efficiency leaps, regulatory gaps, sovereign funds — are being rewritten continuously, and every rewrite opens a door to new players. What DeepSeek proved is not only China’s strength; it is that a focused, efficient, open move from outside the center can shake the entire board.

Chapter 12: Safety, Ethics, and the Apocalypse Narratives

The AI safety debate is not one debate; it is two hostile debates sharing a single stage.

The roots of the first lie in the field’s outer districts. In the 2000s, the circle led by Eliezer Yudkowsky (later MIRI) was arguing the “friendly AI” problem on internet forums: a superintelligence whose goals were misaligned with ours could produce catastrophe without any malice at all. That line of thought gained intellectual respectability with Oxford philosopher Nick Bostrom’s 2014 book Superintelligence — publicly praised by Musk, Gates, and Altman, which made “existential risk” (x-risk) a mainstream Silicon Valley anxiety. The growing effective altruism (EA) movement carried money and people to the cause in the same years: Open Philanthropy made a $30 million grant to OpenAI in 2017; EA circles became the labs’ chief recruiting pool for safety teams; FTX’s collapse burned the movement’s money and reputation, but its institutional traces remained. The technical legacy of this tradition should not be belittled: the RLHF method underneath every chat model today came out of the alignment research of Paul Christiano, Jan Leike, and their teams. The doomer tendency’s most visible artifact arrived in April 2025: the AI 2027 scenario, in which the team of former OpenAI researcher and whistleblower Daniel Kokotajlo choreographed, hour by hour, a loss-of-control trajectory for 2027 — read by millions, and hotly debated inside the field as either forecast or science fiction.

The second debate concerns the harms of the present, and its protagonists are mostly women — itself a datum about the narrative’s hierarchy of visibility. In December 2020, Google forced out Timnit Gebru, co-lead of its Ethical AI team, over a paper examining the risks of large language models — bias, environmental cost, the illusion of meaning; a few months later her co-lead Margaret Mitchell was fired too. Published in March 2021 with the linguist Emily Bender as co-author, the paper lodged its title metaphor permanently in the field: “Stochastic Parrots.” The thesis: these models do not understand; they probabilistically remix patterns in their training data — and marketing them as if they understand renders real, measurable harms invisible. Gebru founded the independent DAIR institute; Meredith Whittaker, an organizer of the Google walkout, became president of Signal and one of the most systematic critics of the surveillance economy; Kate Crawford mapped AI in Atlas of AI as an extractive industry running from lithium mines to click-workers; the cognitive scientists Gary Marcus and Melanie Mitchell kept up the technical case against LLM reasoning claims. This camp’s objection to the apocalypse narrative is blunt: existential-risk discourse shifts attention and resources from today’s documented harms — discriminatory algorithms, labor exploitation, copyright violation, disinformation, surveillance — to a hypothetical future monster; and it does so under the spokesmanship of the very companies claiming to build that monster.

Between the two camps, a third arena grew — the place where the debate touches daily life: the documented harms of the present. Synthetic media became a standard weapon of election seasons — in the global election year of 2024, cases running from voice cloning to fabricated video kept regulators chasing from behind; non-consensual deepfake pornography targeting women grew into the technology’s most widespread and least-discussed abuse. On the labor front, the 2023 Hollywood writers’ and actors’ strikes won history’s first major “AI collective-bargaining” provisions, limiting studios’ use of AI in script production and digital replicas — early proof that organized labor could negotiate with the technology. In professions from education to journalism, translation to graphic design, the transformation proceeds without a contract; Amodei’s warning about “half of entry-level white-collar jobs” remains, as this article is written, a forecast — but the price erosion in translation and graphic-design markets is already measurable. None of these harms is an apocalypse; all of them are the present — and that is the critical camp’s core argument: waiting for the end of the world, we are missing what is happening to it now.

After 2023, the two debates became personified in the schism of the founding fathers. Hinton left Google in May 2023 so that he could speak freely; he had become convinced that digital intelligence learns faster than biological intelligence, and ever since — Nobel podium included — he has been describing the possibility that humanity ceases to be the apex intelligence. Bengio turned his course entirely toward safety at the summit of his academic career: he chaired the International AI Safety Report commissioned after Bletchley (first full edition January 2025, with contributions from about a hundred experts across 30 countries), and in June 2025 founded LawZero, a nonprofit laboratory building non-agentic “Scientist AI.” LeCun planted himself at the opposite pole: current LLMs, he argues, understand less than a house cat; the doom scenarios are unscientific panic production that feeds authoritarian regulation and monopoly. The single-sentence statement signed by hundreds of industry leaders in May 2023 — “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war” — carried Hinton’s and Bengio’s signatures; it did not carry LeCun’s. That three Turing laureates who defended the same cause for forty years should split this way is the most honest indicator of how open the question is.

The tenth research question — is the safety debate sincere, or the construction of entry barriers through regulation? — again deserves the answer “both,” but with evidence. For sincerity: a significant share of lab researchers documented their concerns at real career cost, through resignations, whistleblowing, and open letters; alignment research produces real scientific content. For suspicion: Altman personally proposing licensing before the U.S. Senate in May 2023 was a nearly unprecedented “regulate us” performance, and critics read it as regulatory capture, since any license regime keyed to scale thresholds protects incumbents; in the fight over California’s SB 1047, the labs’ positions aligned with their commercial interests to a striking degree (the bill was vetoed in 2024; the narrower transparency law SB 53 was signed in September 2025). And meanwhile, the invoice for present harms began arriving through the courts: Anthropic’s $1.5 billion book settlement; the rulings in early 2026 in the New York Times case allowing 20 million ChatGPT logs into discovery; Getty-Stability, Disney-Midjourney, the Character.AI suits. The apocalypse has not come; the law has.

Chapter 13: The Psychology of the Founders

Institutions scale the character of their founders. This chapter is a portrait gallery — without canonization or demonization, drawn from the record.

Sam Altman is the strategist of the capital age. St. Louis-born, Stanford dropout, founder of his first company at 19, then Paul Graham’s successor at Y Combinator, where he presided over the births of hundreds of startups. His genius is not technical but topological: he sees before anyone else who needs what, which narrative attracts which capital, which door opens onto which corridor. Paul Graham once distilled the trait into a line that has trailed Altman ever since: “You could parachute him into an island full of cannibals and come back in five years and he’d be the king.” His return, within five days, to the company that had fired him — carried by the signatures of ninety percent of its employees — was a show of force without parallel in corporate history, and at the same time a warning: personal indispensability is the inverse of institutional oversight. Read alongside former board member Helen Toner’s later public account — the claim that Altman had fed the board inaccurate information — and WilmerHale’s “breakdown in trust” formula, the portrait sharpens: Altman is the type of founder who governs through relationships rather than rules, locked on the objective, charismatic; the type that grows companies fast and wears out control mechanisms.

Demis Hassabis is the scientist-king. He is the only figure to have fused chess, game design, neuroscience, and corporate command in a single career, and he always narrated DeepMind through the metaphor of an Apollo program. The Nobel medal was the final ratification of the science-first claim; but the same man commands Google’s product war, and the tension of that duality — the gap between Gemini’s launch calendar and Nature’s — is the central question of his biography. Dario Amodei is the safety technocrat: physics doctorate, speech that moves like a paper’s footnotes, narrating both apocalypse and abundance in the same measured idiom. His critics point to the contradiction between his warnings and his product schedule; his defenders reply that he is the builder among the pessimists and the most honest speaker among the builders. Elon Musk is the paradox itself: the same biography contains DeepMind’s early investor, OpenAI’s founding donor, the six-month-pause signatory, the owner of the lab whose Grok chatbot spiraled into antisemitic tirades, and the builder of one of the world’s largest private compute clusters. His consistency lies not in ideas but in position: in every game, he rejects any board he does not own. The other constant is a record that has punished those who wrote him off — PayPal, Tesla, SpaceX, Starlink, each dismissed as impossible before it hardened into infrastructure — which is why Peter Thiel, his old PayPal partner, compressed two decades of observation into a single sentence: “Never bet against Elon Musk.” xAI, founded years after its rivals, stood up one of the world’s largest training clusters in a matter of months; whatever one makes of Grok, betting against the build has once again proved unwise.

Mark Zuckerberg is the platform emperor; he runs AI not as a product but as the empire’s line of defense. Quietly pivoting to open models when the metaverse bet became a punchline, then rebuilding his lab in a single summer with a $14 billion move when Llama stumbled — that is his real talent: ideologylessness. Openness, for him, is not a principle but the most rational weapon of the current quarter. Ilya Sutskever is the field’s mystic: present at every pivotal technical moment from AlexNet to GPT, the first to insist that “predicting the next word” might require understanding the world, the man credited inside OpenAI with the mantra feel the AGI. His role in the 2023 coup and his remorse afterward form the most dramatic case of technical conscience trapped inside corporate power; today, at SSI — productless and valued at $32 billion — he works directly on superintelligence: the most expensive hermit project in the history of modern technology. Yann LeCun is the scientific dissident, defending his world-models program against the mainstream LLM consensus with a startup founded as he nears seventy. Hinton and Bengio embody the ethical turn of the founding fathers — one converted the Nobel podium into a warning pulpit, the other his career into safety-institution building. Fei-Fei Li is the most credible voice of the human-centered wing: founding co-director of Stanford HAI, advocate of the thesis that AI exists to extend human capability, and — with her World Labs venture on spatial intelligence — a figure who breaks the male-founder mold of the narrative. Jensen Huang is the emperor of the hardware age: thirty years at one company, twice back from the brink of bankruptcy, fusing the humility of a founder accustomed to losing with a five-trillion-dollar monopoly. He saw, long before the labs did, that the future would be computed in parallel — and then spent two decades building CUDA, the software moat that quietly turned a maker of graphics cards into the mandatory supplier of intelligence itself. In the war of the AI empires he is the rare figure who need not pick a side: whichever lab wins the model race, it wins on his silicon, and the deeper the rivalry the more he sells to all of them. It is this article’s thesis rendered in hardware — real power sits not where the activity is loudest but at the bottleneck everyone must pass through, and in the age of AI that bottleneck is a chip. Alexandr Wang is the prototype of the new type: not a researcher but a broker of the data-labor-state triangle; a figure who signed Pentagon contracts in his twenties and then took over the superintelligence program of the world’s largest social platform — proof that AI power can now be produced from supply-chain dominance as well as laboratory genius.

At the gallery’s edge, two transitional figures deserve their own frame. Mustafa Suleyman traced the field’s most improbable orbit: he began as DeepMind’s “conscience,” exited under the shadow of management-style complaints, built Inflection AI with Reid Hoffman into a multibillion-dollar startup — which Microsoft then hollowed out in an unprecedented acquisition-without-acquisition — and found himself CEO of Microsoft AI and, by late 2025, head of the company’s “humanist superintelligence” team. The author of The Coming Wave — the book arguing the technology must be contained — today runs the superintelligence program of one of the world’s largest corporations; after Musk, that trajectory is the clearest specimen of how warning and building interleave in a single career. Mira Murati is the portrait of quiet power: the executive who held the institution together on crisis nights, never sought the spotlight, and — upon leaving — raised one of history’s largest seed rounds alone, before having a product. Thinking Machines’ difficulties in 2026 exposed one of the field’s brutal truths: reputation attracts capital; but in the frontier-model race, capital is only the ticket at the door.

A pattern emerges from this gallery: the field is governed by a handful of people who are extraordinarily talented, extraordinarily self-assured, and extraordinarily resistant to oversight. The problem lies not in the characters but in the architecture — the governance of a technology with civilizational claims is still entrusted to personal virtue.

Chapter 14: Who Will Own AGI?

AGI — artificial general intelligence — has no agreed definition; that is not a detail but the center of the story. OpenAI’s charter defines it as “highly autonomous systems that outperform humans at most economically valuable work” — an economic definition, not a cognitive one. The famous “AGI clause” in the Microsoft-OpenAI contract turned the definition into a legal trigger: if OpenAI’s board declared AGI achieved, Microsoft’s access to the new technology would end. The October 2025 agreement transferred that authority to an independent expert panel. It is worth pausing on: the most ambitious technological threshold in human history is now the arbitration mechanism of a contract between two companies. The word “intelligence” is itself an instrument of this economy: for capital markets it is a nearly perfect word — it seems unmeasurable yet can be poured into benchmarks; it sounds sacred yet can be priced as an API; it presents as universal yet can be locked up in cloud contracts. The answer to the ninth research question lies here too — AGI is four things at once: a scientific goal that functions as the field’s horizon line; an investor story that justifies trillion-dollar valuations with the promise of “the machine that automates everything”; a quasi-religious narrative that speaks in the vocabulary of salvation and apocalypse; and a geopolitical lever running from export controls to summit diplomacy. The word’s power comes precisely from its serving all four functions simultaneously.

The ownership question is being played out on four planes. On the corporate plane the picture is clear and unprecedented: frontier models belong to a handful of institutions that are not even public companies; OpenAI at $852 billion and Anthropic at $965 billion live in private markets, where shareholder oversight is weakest. The November 2023 crisis was a stress test of these institutions’ internal-control mechanisms, and the result was not reassuring. On the state plane, AI has entered the inventory of national power: the United States plays with chip controls and federal purchasing power; China with its state-private hybrid; the Gulf with capital; Europe with rule-making — and the lesson of 2025-26 is that, increasingly, it is not states dictating terms to companies but companies dictating terms to states. On the infrastructure plane, the AGI question becomes concrete: who can access the gigawatts, the high-bandwidth memory, the advanced lithography? On this plane AGI is not software but a matter of energy and real estate — and the answer is written in place names like Abilene, Memphis, and Hangzhou. On the societal plane stands the weakest link: the public that produced the technology’s training data (the entire internet), labored its alignment (the data workers), and carries its risk (everyone whose work is being transformed) has no seat at the table.

Scene: the G7 summit at Évian, June 17, 2026 — the “AI and the Digital Age” working lunch, where G7 leaders and the technology executives from eight countries sat at one table.

There is now a single photograph in which all four planes converge. On June 17, 2026, at the “AI and the Digital Age” working lunch of the G7 summit in Évian-les-Bains, the leaders of the world’s most powerful states shared the table with a dozen technology executives from eight countries. According to the official participant record, seated alongside Altman, Hassabis, Amodei, and — representing Meta — Alexandr Wang was the world beyond the five: Mistral’s Mensch for France, Cohere’s Gomez for Canada, Black Forest Labs founder Robin Rombach — one of the academic architects of Stable Diffusion — for Germany, Sarvam AI for India, Sakana AI for Japan, Domyn for Italy, Synthesia for Britain. Axios’s headline captured the scene: AI CEOs at the G7 “as heads of nation-states.” Behind closed doors, Amodei and Hassabis called for a U.S.-led coalition on chips and frontier-model access that would exclude China; Altman proposed an impartial international forum to set global testing standards; and the European leaders at the same table sought checks on American dominance. In a single lunch: the elevation of companies to state protocol, the conversion of AGI discourse into geopolitical leverage, the institutionalization of “sovereign AI” through national champions — and, beside the five great labs, the chairs of a new generation stretching from Bengaluru to Tokyo. Even the seating plan now stated this article’s thesis.

Those looking for instruments of oversight do not, however, hold an empty bag. Compute is a governable bottleneck: frontier training happens in countable facilities, with traceable chip shipments — threshold-based notification duties (the logic of California’s SB 53 and of the AI Act) lean on that visibility. The third-party evaluation ecosystem — safety institutes, red-teaming firms, academic auditors — is embryonic but growing. Whistleblower protections entered the agenda seriously for the first time after Kokotajlo and colleagues took OpenAI’s equity-pressure tactics public on their way out. And tort law — as the copyright cases show — writes de facto standards of conduct wherever regulation cannot keep up. None of these, alone, can bridle trillion-dollar momentum; together, they show that “ungovernable” is a choice, not a fate.

So: is AGI a product, infrastructure, a weapon, or a public good? The honest answer: today it is in practice a product, rapidly becoming infrastructure, partially weaponizing, and almost nowhere a public good. Open-weight models are a counterweight in this picture that should not be underestimated — no strategic technology in history has been this copyable this early, this widely. But an open model is not open governance. The architecture of public oversight built over a century for electric grids, telecom networks, and the pharmaceutical industry does not yet exist for artificial intelligence; the attempts that do exist — safety institutes, the AI Act, transparency statutes — were weakened in the 2025-26 conjuncture before they were even established. The present answer to the question of who will own AGI is the de facto state of affairs no one declares and everyone knows: AGI will belong to those who build it — until societies organize otherwise.

Conclusion: The Board Belongs to All of Us

This article began with a hotel dinner, two graphics cards in a bedroom, and a thirty-seventh move on a Go board. Let it close there too. After AlphaGo’s victory, Lee Sedol retired from professional Go within a few years; he said he had come to question the game’s meaning “with an entity that cannot be defeated.” But in those same years something unexpected happened as well: a new generation of human players, training with the machines, tore up and rebuilt the game’s five-hundred-year-old orthodoxies, and human Go entered one of the most creative periods in its history. The machine did not end the game; it enlarged the board.

The founding story of artificial intelligence, as this article has tried to show, is too large to be told as the competition of five companies. In that story there are academics who carried an idea ridiculed in Toronto for forty years; an immigrant scientist who built a dataset judged “devoid of ideas” at Princeton; an engineer who realized, while selling chips to gamers, that he was laying civilization’s infrastructure; workers in Nairobi whose own minds were harmed while making other people’s machines harmless; a team in the basement of a Hangzhou hedge fund that shook Wall Street; laws written and diluted in Brussels; and yes — the dinner table at the Rosewood, the five days of November 2023, the courtrooms, and the trillion-dollar rounds. None of these is a footnote to any other. The history of AI is the newest stage of humanity’s struggle over knowledge, power, and control — and on that stage there is no such role as spectator.

This breadth is no accident; it follows from the technology’s nature. Just as electricity could never be a single company’s product — the ones who invented the generator, built the grid, wrote the standards, organized the unions, and regulated the prices each wrote their own history — a general-purpose technology of intelligence cannot fit inside a single corporate narrative. The founders of the five great labs know this better than anyone; each is trying to grow his empire vertically, to control the layers his story depends on — the chips, the electricity, the data, the law, the public. But the layers resist: courts price the copyright, towns sue the turbines, workers testify, competitors melt margins with open weights, and voters and parliaments — slowly, messily, usually late — write rules. History is being written not where the founders planned, but in the sum of these frictions.

The founders have told us their visions; we listened. The models have shown us their capabilities; we saw. What remains is the least-written chapter of the story: the social and institutional labor that will decide who governs this technology, under what rules, accountable to whom. Move 37 was beautiful — but the real wisdom of Go is this: no single move wins the game; the player who sees the whole board does. And the board is bigger than five. It is as big as the world.

This essay has a companion, written from the field rather than the stands: The Five AI Empires, and the World Around Them: A Field Guide to Minimum Viable Sovereignty — what a builder outside the five great labs does about everything described here.

APPENDICES

Appendix 1: Chronology

Appendix 2: Map of People

Founders and CEOs

• Sam Altman — CEO, OpenAI. Former YC president. Contribution: capital formation, productization, policy diplomacy. Position: accelerationist-pragmatist; “beneficial AGI for everyone.”
• Demis Hassabis — CEO, Google DeepMind; DeepMind co-founder. Nobel in Chemistry, 2024. Contribution: the RL + deep learning program, the AlphaGo/AlphaFold vision. Position: science-first, cautious optimist.
• Dario Amodei — CEO, Anthropic; former OpenAI VP of research. Contribution: scaling laws, GPT-2/3 leadership, safety-centered company design. Position: “safety from inside the race”; public risk warner.
• Daniela Amodei — President, Anthropic; architect of operations and culture.
• Elon Musk — Founder, xAI; OpenAI founding donor; Tesla/SpaceX/X. Position: paradoxical — existential-risk prophet and aggressive competitor at once.
• Mark Zuckerberg — CEO, Meta. Contribution: FAIR, PyTorch, Llama, the open-weight strategy, MSL. Position: ideology-free platform defense.
• Greg Brockman — President, OpenAI; infrastructure and engineering leadership; Altman’s most loyal ally.
• Ilya Sutskever — AlexNet co-author, OpenAI founding chief scientist, SSI founder. Position: the technical mystic; “superintelligence safety first.”
• Mustafa Suleyman — DeepMind co-founder, Inflection founder, Microsoft AI CEO. Position: the applied/political wing; “humanist superintelligence.”
• Jensen Huang — Founding CEO, NVIDIA. Contribution: the GPU/CUDA empire. Position: supplier-hegemon; distant from risk debates.
• Alexandr Wang — Founder, Scale AI; Meta Chief AI Officer. Contribution: the data-labeling industry, the defense-AI bridge. Position: America-first competitor.
• Liang Wenfeng — Founder, DeepSeek; High-Flyer fund. Contribution: efficiency-driven open frontier models.
• Arthur Mensch / Guillaume Lample / Timothée Lacroix — Mistral founders; Europe’s open-model claim.
• Clément Delangue / Julien Chaumond / Thomas Wolf — Hugging Face founders; the open ecosystem’s platform.
• Aravind Srinivas — Perplexity founder; the AI-search claim.
• David Holz — Midjourney founder; the independent, investor-free model.
• Emad Mostaque — Stability AI founder (resigned 2024); at once hero and cautionary tale of the open image model.
• Noam Shazeer — Transformer co-author; Character.AI founder; Google→OpenAI (2026). The most itinerant architect of LLM culture.
• Mira Murati — Former OpenAI CTO; manager of the ChatGPT/DALL-E productization; founder of Thinking Machines Lab.

Scientists

• Geoffrey Hinton — Backpropagation; deep learning’s keeper through the winters; Turing 2018, Nobel in Physics 2024. Position: chief risk warner since 2023.
• Yann LeCun — CNN inventor; FAIR founder (departed 2025, AMI Labs); Turing 2018. Position: LLM skeptic, open-source advocate, anti-doomer.
• Yoshua Bengio — The Montreal school/Mila; Turing 2018. Position: leader of safety institutionalization (International Report, LawZero).
• Fei-Fei Li — Creator of ImageNet; Stanford HAI; World Labs. Position: human-centered AI.
• Alex Krizhevsky — Chief engineer of AlexNet; the revolution’s least public figure.
• Andrej Karpathy — OpenAI founding member, Tesla Autopilot director, educator-narrator; at Anthropic from 2026.
• John Schulman — PPO/RLHF; among ChatGPT’s technical fathers; OpenAI→Anthropic→Thinking Machines.
• Jared Kaplan — Lead author of the scaling laws; Anthropic founding scientist.
• Chris Olah — Pioneer of interpretability research; Anthropic co-founder.
• Jan Leike — Alignment researcher; protest resignation from OpenAI (2024), Anthropic alignment lead.
• Andrew Ng — Google Brain founder, Coursera, Baidu; mass educator of AI.
• Jeff Dean — Google’s infrastructure legend; Brain co-founder; Google chief scientist.
• Ian Goodfellow — Inventor of GANs; forerunner of generative modeling.
• Ashish Vaswani et al. — The Transformer’s eight authors; all left Google.
• David Silver — The RL brain of AlphaGo/AlphaZero. Oriol Vinyals, Koray Kavukcuoglu — DeepMind technical leadership. John Jumper — AlphaFold; Nobel 2024.

Critics and watchdogs

• Timnit Gebru — Former Google Ethical AI co-lead; DAIR founder; central figure of the data/bias critique.
• Emily Bender — “Stochastic Parrots”; the critique of the illusion of meaning.
• Margaret Mitchell — The second ethicist fired from Google; Hugging Face chief ethics scientist.
• Meredith Whittaker — Signal president; critic of the surveillance economy.
• Kate Crawford — Atlas of AI; the material/labor map of AI.
• Gary Marcus, Melanie Mitchell — Scientific skeptics of LLM capability claims.
• Daniel Kokotajlo — OpenAI whistleblower; lead author of AI 2027.

Capital and platform figures

• Satya Nadella — Microsoft; architect of the OpenAI alliance. Sundar Pichai — Google/Alphabet. Reid Hoffman — OpenAI founding donor, Inflection co-founder. Peter Thiel — Early backer of DeepMind and OpenAI. Masayoshi Son — SoftBank; Stargate co-owner. Jeff Bezos / Larry Ellison — Financiers of the infrastructure war through Amazon and Oracle.

Appendix 3: Map of Institutions

Appendix 4: Table of Central Conflicts

Appendix 5: Bibliography

Note: The bibliography below organizes the sources used in this article by category. Dates and figures for developments from mid-2024 through 2026 were verified via web research conducted in July 2026 against the sources listed.

Official company sources

• OpenAI blog: “Introducing OpenAI” (Dec 2015); “OpenAI LP” (Mar 2019); “Announcing The Stargate Project” (Jan 2025); “Built to benefit everyone” (Oct 2025); the GPT/Sora/o-series announcements; the November 2023 crisis statements; the WilmerHale review summary (Mar 2024)
• Microsoft blog: “The next chapter of the Microsoft-OpenAI partnership” (Oct 28, 2025); Microsoft AI/MAI announcements
• Anthropic: founding and Claude announcements; “Constitutional AI” (arXiv:2212.08073); the interpretability series; Series E-H announcements; the Responsible Scaling Policy
• Google/DeepMind: DQN (Nature, 2015), the AlphaGo/AlphaZero/AlphaFold publications; the Google DeepMind merger announcement (Apr 2023); Gemini announcements; the IMO 2024/2025 posts
• Meta: FAIR/Llama announcements; Zuckerberg, “Open Source AI Is the Path Forward” (Jul 2024)
• xAI: founding and Grok announcements; the July 2025 incident statement
• NVIDIA: quarterly results (FY2026); official announcements from Mistral, SoftBank, SSI, DeepSeek, Hugging Face, Stability, Runway, Perplexity
• Dario Amodei: “Machines of Loving Grace” (Oct 2024); “The Adolescence of Technology” (Jan 2026)

Academic sources

• McCarthy et al., the Dartmouth proposal (1955); Rosenblatt (1958); Minsky & Papert, Perceptrons (1969); Rumelhart, Hinton & Williams (Nature, 1986); LeCun et al. (1989/1998); Hinton et al., deep belief nets (2006); Deng, Li et al., ImageNet (CVPR 2009); Krizhevsky, Sutskever & Hinton, AlexNet (NeurIPS 2012); Vaswani et al., “Attention Is All You Need” (2017); Kaplan et al., “Scaling Laws for Neural Language Models” (2020); Bender, Gebru et al., “On the Dangers of Stochastic Parrots” (FAccT 2021); Jumper et al., AlphaFold2 (Nature, 2021); AlphaFold 3 (Nature, 2024); Bengio et al., International AI Safety Report (2025)
• The Nobel Foundation citations (Physics and Chemistry, 2024); the ACM Turing Award citation (2018)

Journalism

• Wired: Cade Metz’s AlphaGo/Seoul reporting (Mar 2016); the OpenAI crisis files; Steven Levy’s Microsoft-OpenAI alliance cover feature
• TIME: Billy Perrigo, “OpenAI Used Kenyan Workers on Less Than $2 Per Hour” (Jan 2023)
• Forbes: the Emad Mostaque investigation (Jun 2023); Brockman’s testimony (May 2026)
• Reuters, Bloomberg, CNBC, Financial Times, NYT, The Information, TechCrunch, The Verge, MIT Technology Review, IEEE Spectrum: funding rounds and valuations (OpenAI $157B→$852B; Anthropic $61.5B→$965B; xAI ~$230B), the DeepSeek/NVIDIA market shock (Jan 2025), the Meta-Scale deal (Jun 2025), LeCun’s departure (Nov 2025), the Zhipu/MiniMax IPOs (Jan 2026), the Shazeer transfers, the Gemini/Llama/Grok launch coverage
• NPR/CNN/Axios: the Musk-OpenAI verdict (May 18, 2026); the Grok incident of July 2025
• Platformer: the GPT-5 backlash (Aug 2025); SemiAnalysis: the DeepSeek cost analysis; Epoch AI: Stargate site status
• Books: Cade Metz, Genius Makers (2021); Parmy Olson, Supremacy (2024); Karen Hao’s OpenAI reporting

Legal and regulatory sources

• Musk v. Altman/OpenAI filings (N.D. Cal., 2024-2026) and the May 18, 2026 ruling
• Bartz v. Anthropic: Judge Alsup’s fair-use ruling (Jun 2025) and the $1.5B settlement documents (final approval in progress as of May-Jun 2026)
• NYT v. OpenAI/Microsoft (S.D.N.Y.): the discovery rulings (Nov 2025-Jan 2026)
• Getty Images v. Stability AI: the London High Court ruling (Nov 2025); the U.S. case pending
• The EU AI Act text and the “digital omnibus” amendments (Nov 2025-Jun 2026); the Bletchley Declaration (2023); the Paris Summit documents (Feb 2025); the U.S. AI Action Plan (Jul 2025); California’s SB 1047 veto message (2024) and SB 53 (2025)

Critical literature

• Kate Crawford, Atlas of AI (2021); Emily Bender & Alex Hanna, The AI Con (2025); Gary Marcus, Taming Silicon Valley (2024); Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans (2019); the DAIR and AI Now institute reports; Kokotajlo et al., “AI 2027” (Apr 2025)

Podcasts and long-form interviews

• Lex Fridman (the Altman, Musk, LeCun, Karpathy episodes); Dwarkesh Patel (Amodei, Sutskever, Hassabis); No Priors; a16z; Altman’s Senate testimony (May 2023); the Nobel press conferences (Oct 2024)

Selected primary-source links

• The Dartmouth proposal (1955): https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
• Backpropagation (Nature, 1986): https://www.nature.com/articles/323533a0
• AlexNet (NeurIPS 2012): https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks
• “Attention Is All You Need” (2017): https://arxiv.org/abs/1706.03762
• Scaling laws (2020): https://arxiv.org/abs/2001.08361
• Constitutional AI (2022): https://arxiv.org/abs/2212.08073
• Stochastic Parrots (FAccT 2021): https://dl.acm.org/doi/10.1145/3442188.3445922
• AlphaFold2 (Nature, 2021): https://www.nature.com/articles/s41586-021-03819-2
• OpenAI founding announcement (2015): https://openai.com/index/introducing-openai/
• OpenAI LP / capped-profit (2019): https://openai.com/index/openai-lp/
• ChatGPT announcement (2022): https://openai.com/index/chatgpt/
• OpenAI board-crisis announcement (Nov 2023): https://openai.com/index/openai-announces-leadership-transition/
• OpenAI PBC conversion (Oct 2025): https://openai.com/index/built-to-benefit-everyone/
• Microsoft-OpenAI new agreement (Oct 2025): https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
• Stargate announcement (Jan 2025): https://openai.com/index/announcing-the-stargate-project/
• Anthropic Series H (May 2026): https://www.anthropic.com/news/series-h
• Dario Amodei, “Machines of Loving Grace”: https://darioamodei.com/machines-of-loving-grace
• Zuckerberg, “Open Source AI Is the Path Forward” (2024): https://about.fb.com/news/2024/07/open-source-ai-is-the-path-forward/
• The DeepMind AlphaGo archive: https://deepmind.google/research/alphago/
• The TIME Kenya investigation (2023): https://time.com/6247678/openai-chatgpt-kenya-workers/
• The Musk-OpenAI verdict (NPR, May 2026): https://www.npr.org/2026/05/18/nx-s1-5822366/musk-altman-openai-jury-verdict-claims-dismissed
• The EU AI Act framework: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
• The 2018 Turing Award citation: https://awards.acm.org/about/2018-turing
• The G7 Évian AI lunch (CNBC, Jun 2026): https://www.cnbc.com/2026/06/17/g7-trump-ai-tech-leaders-openai-anthropic-google.html
• “AI CEOs as heads of nation-states” (Axios, Jun 2026): https://www.axios.com/2026/06/20/ai-tech-moguls-g7
• Nobel Prize in Physics 2024: https://www.nobelprize.org/prizes/physics/2024/summary/
• Nobel Prize in Chemistry 2024: https://www.nobelprize.org/prizes/chemistry/2024/summary/

Key sources by chapter (selected)

• Introduction & Chapter 3: Metz’s Wired AlphaGo file; the DeepMind AlphaGo documentary; the Nature DQN and AlphaFold papers; the 2024 Nobel citations
• Chapters 1-2: the Turing Award citation; Genius Makers; ImageNet (CVPR 2009); the AlexNet paper; the NVIDIA founding lore (Denny’s)
• Chapters 4-5: the OpenAI founding announcement; the email exhibits from the Musk litigation; the GPT-4 technical report (the closure statement); the November 2023 crisis coverage (Wired/NYT/Bloomberg); the PBC conversion documents (Oct 2025)
• Chapter 6: the Anthropic founding sources; Constitutional AI; the Series F-H announcements; the Bartz settlement; the Amodei essays
• Chapter 7: the xAI announcements; the NPR/CNN verdict coverage; the NAACP/SELC Memphis file; the ADL statement (Jul 2025)
• Chapter 8: Zuckerberg’s open letter; the LMArena statement (Apr 2025); CNBC’s Scale AI coverage; LeCun’s exit interviews
• Chapter 9: NVIDIA’s FY2026 results; Bloomberg’s circular-deals file; the TIME Kenya investigation; the nuclear PPA announcements
• Chapters 10-11: the Hugging Face state-of-the-hub reports; the DeepSeek technical reports + SemiAnalysis; the Xinhua/Interconnects Qwen data; the EU AI Act/omnibus documents; the Paris Summit coverage
• Chapter 12: Stochastic Parrots; the CAIS statement (May 2023); the Hinton/Bengio statements of 2025-26; the Senate transcript (May 2023)

This article rests on publicly available sources accessible as of July 2026. For fast-moving data — valuations, user counts, litigation status — the dates given in the text govern. Points where sources conflict (the DeepMind acquisition price, DeepSeek’s training cost, the SpaceX-xAI merger valuation, the ChatGPT user metric, the Thinking Machines valuation) are flagged explicitly in the text.