What top AI labs are really paying researchers, and what they are not

The numbers everyone is quoting

By the middle of 2025, one figure was everywhere: $100 million, supposedly handed to individual AI researchers simply for signing with Meta. The claim went mainstream when OpenAI chief executive Sam Altman said on a podcast, reported by CNBC in June 2025, that Meta had tried to recruit his staff with $100 million signing bonuses and even larger annual pay. Wired's Zoe Schiffer then reported that Mark Zuckerberg had, on more than ten occasions, offered top researchers packages worth up to $300 million over four years, with more than $100 million in the first year alone.

Those are the numbers that carried the story across the internet. They are also the numbers worth slowing down on, because what they describe is not quite what the headlines imply.

What is on the record, and what is rumor

Begin with the contested part. Meta's own executives disputed the framing. Chief technology officer Andrew Bosworth, quoted by TechCrunch in June 2025, said the $100 million was not a "sign-on bonus" but "all these different things," meaning base salary, stock that vests over years, and performance-tied pieces bundled into a single ceiling. Lucas Beyer, a researcher who did leave OpenAI for Meta, wrote publicly that he had not received a $100 million signing bonus and called that version "fake news." Bosworth also rejected the idea that such offers were going to everyone, saying the market was hot but "not that hot."

So the viral figure is not a paycheck. It is a ceiling, spread over four years and weighted toward stock that pays out only if the researcher stays and the share price holds.

The clearest reported case of a genuinely large package is Matt Deitke. The New York Times reported in July 2025 that Meta offered the 24-year-old researcher about $250 million over four years, after he first declined roughly $125 million. He joined Meta's Superintelligence Labs the next month. Even this is reporting, not a figure Meta has published, and it still describes a four-year maximum rather than cash up front.

Meta's raid, and the startups caught in it

The pay war is inseparable from Meta's hiring campaign. Through 2025 Meta built its Superintelligence Labs by recruiting from OpenAI, Google DeepMind, and Anthropic, with Zuckerberg personally involved. The pressure reached the newest labs too. CNBC reported in June 2025 that Meta tried to buy Safe Superintelligence, the startup founded by OpenAI co-founder Ilya Sutskever and reportedly valued near $32 billion earlier that year, according to TechCrunch. Sutskever declined to sell. Meta then hired the startup's co-founder and chief executive Daniel Gross, along with former GitHub head Nat Friedman, and Sutskever stepped in as chief executive of the company he had co-founded.

Mira Murati's Thinking Machines Lab sat in the same crosscurrent. Bloomberg reported in June 2025 that the startup, led by OpenAI's former chief technology officer, raised about $2 billion at a valuation near $10 billion to $12 billion, with nearly two-thirds of its early team drawn from OpenAI. The lesson of 2025 was that the scarce input was not capital. It was people, and the same few hundred names kept moving between a handful of buildings.

Why rivals did not simply match

OpenAI's reaction was more measured than the headline war suggested. TechCrunch reported in June 2025 that the company was "recalibrating" compensation, and chief research officer Mark Chen said he and Altman had worked "around the clock" to talk with staff who held competing offers. Altman, for his part, called Meta's approach "somewhat distasteful" and claimed Meta had failed to sign his top people and "had to go quite far down their list."

Anthropic made a different argument, using its retention numbers. SignalFire's 2025 State of Talent report put Anthropic's two-year retention at about 80%, the highest among the large labs, against 78% at Google DeepMind, 67% at OpenAI, and 64% at Meta. The same report found engineers far more likely to move toward Anthropic than away from it. Chief executive Dario Amodei has said he will not match the largest offers because it would break the company's "principles of fairness," arguing that rivals are trying to buy alignment with the mission, which cannot be bought. Reporting also indicates Anthropic quietly used retention grants during the peak of the raid, so the contrast is not purely money against mission.

Why a few names command these sums

Strip away the theater and the economics are straightforward. The pool of people who have personally pushed a frontier model forward is small, measured in the hundreds, and their choices can move a lab's progress by months. When several companies with near-unlimited capital bid for the same scarce input, the price for the very top of that pool climbs toward what a sports franchise pays a star. The New York Times made the comparison explicit when it likened the Deitke negotiations to NBA contracts.

That is the honest version of the story. A small number of researchers genuinely command sums that would have sounded absurd two years ago, driven by scarcity and by their leverage over frontier progress. But the claim that everyone is getting $100 million is not supported by the record. Most researchers, even strong ones, earn far less, the eye-catching totals are multi-year ceilings rather than signing checks, and the labs that lost the bidding often kept their people anyway. The war is real. The most viral numbers are not the whole truth.