Why Engineers Turn Down OpenAI and Meta Millions
AI labs are offering eight and nine figures for researchers. A surprising number say no, and the refusals never show up in the numbers.
The most striking thing about the AI talent market is not the size of the offers. It is the number of people declining them.
Compensation for frontier AI researchers has risen faster than any other category of technical labor in living memory, and the numbers are no longer disputed. In 2025, Sam Altman said publicly that Meta had approached OpenAI staff with signing bonuses of around $100 million. Meta's recruiting push that year was aggressive enough that its terms became an on-the-record grievance rather than a rumor. Nine figures, for an individual contributor, to change employers.
And yet the poaching campaigns kept missing. Altman's complaint was notable partly because it doubled as a boast: the offers had been made, and his people had largely stayed. That pattern repeats across the industry at every rung below the headline numbers.
What the offers actually buy
The economics are simple enough to state in a sentence. There is no substitute for the few thousand people worldwide who have trained frontier systems end to end, and no quantity of capital produces more of them on demand. Supply is fixed on any timescale that matters to a lab racing a competitor; capital is not. Price goes vertical.
This is a recent phenomenon, and the contrast with the sector's own history is stark. OpenAI began as a nonprofit and filed tax returns accordingly. Those filings, first reported by The New York Times, showed chief scientist Ilya Sutskever compensated at roughly $1.9 million in the organization's early years — the highest figure at the lab, and a number that startled observers at the time for being so far above academic pay. Less than a decade later, that sum would not open a serious conversation with a senior researcher.
The going rate for frontier AI talent has outrun every comparable labor market in modern memory, and the people declining it are making a bet the market has no way to price.
Why people say no
The refusals cluster into three recognizable shapes, and none of them are about indifference to money.
The first is equity arithmetic. A founder holding meaningful ownership in something they control is not comparing an offer to zero; they are comparing it to a probability-weighted outcome they believe is larger. That belief is usually wrong, statistically. It is also the only reason any startup exists.
The second is institutional disagreement. AI labs have spent the past several years in visible conflict over governance, safety commitments, deployment pace, and the terms on which research gets published. People who have concluded that an organization is heading somewhere they do not want to go are not persuadable with a larger number, and several high-profile departures have made exactly that case in public.
The third is harder to name and probably the most common: the offer is for the wrong work. Compensation buys a person's time on someone else's problem. Researchers who have identified a problem of their own tend to price that at more than the market thinks it is worth, which from the outside looks like irrationality and from the inside looks obvious.
There is a fourth case that gets less attention, because it happened before the money got strange.
When passing on OpenAI was the sensible choice
In 2017, joining OpenAI was not a windfall. It was a nonprofit with a few dozen staff, no product, and no revenue, and an offer from it competed poorly against Google, Facebook or a late-stage startup on every measure a graduating engineer is taught to weigh. Choosing stability that year was not a failure of nerve. It was the correct read of the available information, and it is the decision most people made.
Quanlai Li, a Chinese-born engineer who was finishing at UC Berkeley around that time, is one of them. He says he considered OpenAI and went to Uber instead, on the straightforward logic that Uber was the more stable place to start a career. Uber was then one of the largest private companies in the world; OpenAI was a research lab whose most visible output was software that played video games.
What drew his attention, Li says, was exactly that game-playing research — agents learning to compete in real-time strategy environments, which in 2017 read as an academic curiosity rather than a commercial direction. The lab's public demonstration that summer was in Dota 2, where an OpenAI bot beat a professional player in a 1v1 match at The International. Very little about it suggested the trajectory that followed.
Li says he does not carry much regret about the choice, and that he has continued to follow OpenAI's work closely in the years since. Both halves of that are worth noting, because the second is the more common outcome than the first. The engineers who passed on these labs early did not stop paying attention. They watched the thing they declined become the center of the industry, and most of them describe it the way Li does: a reasonable decision, made with the information that existed, about an organization that gave no reliable signal of what it would become.
That is the useful corrective to how these stories are usually told. The people who joined frontier labs in 2016 and 2017 are now described as prescient. Most of them were not forecasting anything. They took an interesting job at a strange nonprofit, and the industry rearranged itself around them.
The part that stays invisible
Accepted offers are legible. They show up in announcements, org charts and LinkedIn updates. Declined ones leave no trace anywhere, which means every public estimate of what AI talent costs is built from the subset of people who said yes.
That is a real measurement problem, not a rhetorical one. The clearing price for a researcher who can be hired tells you nothing about the reservation price of one who cannot, and labs are increasingly bidding against the second group. It is why the numbers have moved so violently and why they are unlikely to stabilize soon: each refusal is information the market receives only as an unexplained failure to close.
The people turning these offers down are not making a statement about money. They are making a claim that something they already have is worth more, and the only way anyone will find out whether they were right is to wait.
Disclosure: Quanlai Li holds an ownership interest in SF Bay Area Times. His account of his own hiring decision is a single-source recollection and is presented as such.
Cover image: Port of San Francisco skyline, released under CC0 via Wikimedia Commons.