The Talent War

“Missionaries will beat mercenaries.” — Sam Altman, in an internal memo to OpenAI staff, late June 2025

In the early summer of 2025, the most valuable thing in technology was not a chip or a data center or a training run. It was a phone number. Specifically, it was the personal phone numbers of perhaps a few hundred people, scattered across a dozen labs, who knew how to build a frontier model and had done it before. Mark Zuckerberg had been collecting them.

He did the collecting himself, which was the part that startled people. The chief executive of a company worth more than a trillion dollars, with tens of thousands of employees and a recruiting apparatus that could fill an arena, was personally working a spreadsheet of names. He texted researchers directly. He invited them to dinner at his house in Palo Alto and to his estate at Lake Tahoe. He asked them who else was good, then added those people to the list and texted them too. Inside Meta the document acquired a nickname that leaked to the press: “The List.” It was a roster of the people Zuckerberg believed could pull his company back to the front of the AI race, and he was prepared to pay each of them more money than most of them had ever imagined a salaried employee could earn.

The reason for the urgency was a public humiliation. In April 2025, Meta had released Llama 4, the latest in the open-weights model family that had once made the company a hero to the open-source world, and the launch had gone badly, capped by a leaderboard embarrassment over a specially tuned variant that the previous chapter recounts in full. For a company that had bet its AI strategy on being the credible open alternative to OpenAI and Google, it was a bad few weeks. Zuckerberg concluded that incremental fixes would not close the gap. He decided to buy his way to the frontier.

The first and largest purchase was a person, wrapped inside a company. On June 12, 2025, Meta announced it would pay roughly $14.3 billion for a stake of about 49 percent in Scale AI, the data-labeling firm that supplied the human-annotated training data many of the labs depended on. The deal valued Scale at around $29 billion. But the structure was unusual: Meta took a large minority stake rather than acquiring the company outright, a shape that conveniently sidestepped the kind of antitrust review a full acquisition would have triggered. What Meta actually wanted was not the labeling business. It was Scale’s twenty-eight-year-old founder, Alexandr Wang.

Wang, the Scale AI founder introduced earlier in this book, had no record of building frontier models himself. What he had was a reputation as an operator, a network that touched every major lab, and a willingness to bet his career on Zuckerberg’s. He joined Meta as its first Chief AI Officer and the head of a new organization, Meta Superintelligence Labs, which absorbed the company’s older research groups, including FAIR, the Facebook AI Research lab that Yann LeCun had founded a decade earlier. The man who labeled other people’s training data now sat atop the research empire of one of the most powerful technology companies on earth. The signal to the market was unmistakable. Meta was no longer trying to win the talent war on the merits of its culture or its mission. It was going to win it with money.

The packages were the part that broke people’s sense of proportion. Reporting from The New York Times, Wired, and others described offers in tiers that climbed past anything a research scientist had ever commanded. There were packages worth more than $100 million in total compensation, structured over several years and weighted heavily toward Meta stock. There were a handful that reportedly reached around $300 million across four years for the most coveted individuals. And then there was the figure that became shorthand for the whole strange season: a package, reported by The Wall Street Journal, worth roughly $1.5 billion over six years, dangled in front of Andrew Tulloch, a respected engineer who had helped build PyTorch and had recently co-founded a new startup. Meta disputed the precise number, saying the figure was off. Tulloch turned it down and stayed where he was. That a person could be offered more than a billion dollars and say no told you something about how much the offers were worth and how little money had to do with the decision.

Sam Altman, whose company was the primary target of the raids, did the math out loud. Meta, he said on a podcast in June 2025, had been offering signing bonuses of $100 million to lure OpenAI researchers away. He framed it as a compliment that doubled as an insult: Meta had to pay that much because it could not attract people any other way. “Missionaries will beat mercenaries,” he told his own staff in an internal memo that quickly leaked, drawing the line he wanted drawn between OpenAI, which he cast as a place people came to because they believed in the work, and Meta, which he cast as a place people came to for the check. It was a good line, and it was not entirely fair, given that OpenAI paid its own researchers extraordinarily well and would, within months, run a secondary share sale that turned its employees’ paper into hundreds of millions in real cash. But it captured the anxiety. The poaching did not feel like ordinary competition. It felt personal.

Mark Chen, OpenAI’s chief research officer, made the feeling explicit. In a message to the research staff after Meta succeeded in hiring several of them away in late June 2025, Chen wrote that he felt “a visceral feeling right now, as if someone has broken into our home and stolen something.” He promised that leadership would fight to keep people. The metaphor was telling. A home is not a market. You do not expect your neighbors to make competing bids on your children. Chen was reaching for the language of violation because the ordinary language of compensation and retention no longer described what was happening. The labs had spent years cultivating the idea that they were families bound by mission, and Zuckerberg had just demonstrated that a family could be disassembled one nine-figure offer at a time.

Meta did pull people across. The most visible group came from OpenAI’s Zurich office: Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai, three researchers known for foundational work on vision transformers who had moved from Google DeepMind to OpenAI not long before and now moved again. Trapit Bansal, who had worked on the reinforcement-learning techniques behind OpenAI’s reasoning models, went to Meta as well. By midsummer the count of senior researchers Meta had recruited from rival labs was in the dozens, drawn from OpenAI, Google DeepMind, Anthropic, and the smaller shops. Zuckerberg also bought the people who could find more people. He effectively acquired the venture firm NFDG, run by Nat Friedman, the former chief executive of GitHub, and Daniel Gross, who had until recently been running Ilya Sutskever’s secretive startup. Both joined Meta Superintelligence Labs. Gross’s departure left Sutskever’s company without the chief executive who had helped found it, and it showed how far the gravitational pull of Zuckerberg’s checkbook could reach, into even the most committed missions. The man who had walked away from OpenAI to chase a purer version of the work had now lost his co-founder to the bluntest possible expression of the impure version, the one whose whole pitch was the size of the offer.

For the researchers being courted, the experience was disorienting in a specific way. These were people who had, for the most part, gotten into the field a decade or more earlier, when neural networks were an academic backwater and a faculty job at a good university was the prize worth wanting. They had taken the unfashionable path because they found the problem beautiful. Now the chief executive of Meta was texting them at night, and the numbers on the table exceeded what a founder might hope to clear from a successful exit. Some described feeling less like recruits than like assets being moved between balance sheets. The decision in front of them was rarely about whether they could use the money. Most of them were already wealthy by ordinary standards. It was about what the money said, and whether saying yes to it would change what they were. Altman’s missionaries-and-mercenaries framing landed because it named the exact discomfort: every researcher who took a giant package had to decide whether they had just confirmed they were the second thing.

That was the irony at the center of the season, because some of the largest valuations in AI in 2025 attached to companies that had no product at all and no plans to ship one soon. They were priced almost entirely on the names of the people who founded them.

The clearest example was Sutskever’s. After leaving OpenAI in the wake of the November 2023 board crisis, Sutskever had founded Safe Superintelligence in June 2024 with a stated philosophy of having “one goal and one product” and a refusal to release anything incremental along the way. The company would build a safe superintelligence and ship nothing until it had. By April 2025, Safe Superintelligence had raised about $2 billion at a valuation of roughly $32 billion. It had no revenue, no public model, no demo, nothing a customer could touch. Its valuation had multiplied roughly sixfold in seven months. Investors including Greenoaks, Andreessen Horowitz, Lightspeed, and DST, with Alphabet and Nvidia among the backers, were paying $32 billion for the proposition that Ilya Sutskever, who had been the technical conscience of the deep-learning era, knew something the rest of them did not and would eventually reveal it.

Mira Murati’s company followed the same logic. Murati had been OpenAI’s chief technology officer and had briefly served as its interim chief executive during the board crisis. In February 2025 she founded Thinking Machines Lab and pulled in a roster of senior people from OpenAI and elsewhere: John Schulman, one of the architects of the reinforcement-learning methods behind ChatGPT; Barrett Zoph; Lilian Weng; Andrew Tulloch, the same engineer who would later turn down Meta’s billion-dollar offer. In July 2025 the company closed a seed round of about $2 billion at a valuation of roughly $12 billion, led by Andreessen Horowitz with Nvidia, Accel, AMD, Jane Street, and others participating. It was, at the time, one of the largest seed rounds in the history of venture capital, and it was raised before the company had said publicly what it intended to build. By November, reports had Thinking Machines in talks at a valuation near $50 billion. The product, when a piece of it finally appeared, was a research tool and an API, not the consumer juggernaut the valuation seemed to imply.

The pattern had a name now in the venture world, half admiring and half nervous: the no-product round. It rested on a simple wager, which was that the binding constraint in artificial intelligence had shifted. For most of the field’s modern history the constraints had been ideas, then data, then compute, in roughly that order. By 2025 the labs had access to staggering quantities of compute, much of it described in the previous chapter, and the data and the core ideas were broadly understood across the industry. What was genuinely scarce, the investors had concluded, was the small number of people who could combine all of it into a model that actually worked at the frontier. There were not thousands of such people. There were, by most insiders’ reckoning, a few hundred. And if you believed that one of them, a Sutskever or a Murati, could assemble and lead a team that would build something worth a hundred billion dollars, then $32 billion or $12 billion for the privilege of backing that bet was not obviously crazy. It was a venture bet on a person, priced the way a sports franchise prices a generational athlete, less for what they have produced this season than for the championships they might deliver over a contract.

Whether the bet would pay was, as of the spring of 2026, unknown. Neither Safe Superintelligence nor Thinking Machines had shipped anything resembling the world-changing product their valuations implied. The companies were young and the timelines they had set themselves were long by design. But the valuations had already done their work in the labor market. They reset every researcher’s sense of their own worth. If a person who had not yet built anything could be valued at tens of billions on the strength of a résumé, then a person who had built something at OpenAI or Google could reasonably ask what they were worth, and the answer kept climbing.

The most poignant casualty of the season was the man whose lab Wang had just inherited. Yann LeCun had won the Turing Award in 2018 alongside Geoffrey Hinton and Yoshua Bengio for the work that made deep learning possible. He had founded FAIR in 2013, when Zuckerberg first called him, and had spent twelve years building it into one of the premier research organizations in the field. He was a Meta institution, and he was also, by temperament, a contrarian. While the rest of the industry poured its capital into ever-larger language models, LeCun had spent years arguing publicly that large language models were a dead end on the road to real intelligence, that they could not plan, could not reason about the physical world, and would never get there no matter how much text they consumed. They were, in the house-cat line he had been repeating for years, less intelligent than the family pet. He had been pushing instead for what he called world models, systems that learn how the world behaves by watching it rather than by reading about it, an approach his team pursued under the name V-JEPA.

The trouble was that the company he worked for had just spent $14.3 billion to bet on exactly the opposite. FAIR had been subordinated, after years of budget cuts, to an organization run by a twenty-eight-year-old whose company sold data labeling and who believed, as Zuckerberg did, that the path forward ran straight through bigger models. In mid-November 2025, LeCun confirmed he was leaving Meta. He was not poached, and he did not chase a package. He left to start his own company, AMI Labs, in Paris, to build the world models he had been arguing for, on the conviction that the entire industry had picked the wrong road. In March 2026, AMI Labs was reported to have raised about $1.03 billion at a pre-money valuation of $3.5 billion, one of the largest pre-launch rounds in European history, with Jeff Bezos’s investment vehicle, Eric Schmidt, and Tim Berners-Lee among the backers. The godfather of the field had, at sixty-five, joined the no-product club himself, raising a billion dollars on his name and a thesis that the people who had just been paid hundreds of millions to scale language models were building the wrong thing.

It was a strange tableau, taken together. Zuckerberg paying $14.3 billion to install a data-labeling founder over the lab that one of the field’s founders had built, while that founder walked out the door to bet against the whole enterprise. Sutskever and Murati commanding tens of billions for companies that had shipped nothing. Researchers turning down a billion and a half dollars because they believed more in a startup than in Meta’s stock. Mark Chen comparing a rival’s hiring to a home invasion. The numbers had detached from anything an outsider could anchor to, and the people inside the field talked about the offers the way earlier generations had talked about lottery tickets, with a mix of disbelief and the uneasy sense that the rules had changed and no one had decided to change them.

There was a coherent logic underneath the spectacle, even if it looked like mania from the outside. If a frontier model, once built, could generate revenue at the rate the labs were beginning to report, and if the difference between a model at the frontier and a model a year behind it was worth tens of billions of dollars in enterprise contracts and consumer subscriptions, then a researcher who could move a lab from second place to first was, in pure financial terms, worth an enormous amount. The math of the talent war was the math of leverage. A team of a few hundred people, paid in aggregate a few billion dollars a year, was building products that the same companies were spending hundreds of billions to run. Against that backdrop, a $100 million package for someone who could meaningfully improve the model was not extravagant. It was rounding error on the compute bill.

What the season could not settle was whether the wager on individuals was sound. The history of the field cut both ways. Some of its largest leaps had indeed come from small groups of identifiable people: the eight authors of the transformer paper, the handful of researchers who built the first version of ChatGPT, the team that cracked reasoning models. But many of those leaps had also depended on being in the right institution at the right moment, with the right compute and the right colleagues, in a way that did not obviously transfer when you wrote one of those people a giant check and moved them somewhere else. Meta had bought a great deal of talent by the spring of 2026. It had not yet bought its way back to the frontier. The model that would vindicate Zuckerberg’s spending had not shipped, and the leaderboards were still topped by the companies he had been raiding.

All of that money and all of those people were, in the end, an input. They were assembled in the service of one output, which was a model good enough to win the month. And in 2025 and 2026, the months came fast, each one bringing a new release that claimed the crown from the last, in a cadence the field had never seen and could barely keep up with.