The AI Talent War Nobody's Talking About: Why Top Researchers Keep Jumping Ship
By reviewsideal Staff | September 27, 2026 | 8 min read
A researcher joins Meta on a package reportedly worth over a billion dollars. A year later, he leaves for a rival lab anyway. That single story, playing out in public this month, captures something the industry's compensation arms race keeps proving: money buys a signature, not loyalty.
The Story That Started It: Andrew Tulloch
Few individual moves illustrate the current market better than Andrew Tulloch's. A machine learning researcher who spent 11 years at Meta before leaving for OpenAI and later co-founding Mira Murati's startup Thinking Machines Lab, Tulloch was personally recruited back to Meta in 2025 on a package reported by multiple outlets to be worth as much as $1.5 billion over roughly six years. In September 2026, according to the Wall Street Journal and other outlets, he left Meta again, this time for Anthropic, to join its inference and performance team. Less than a year after one of the most closely watched hires in Silicon Valley, the person at the center of it moved on regardless.
The Numbers Got Genuinely Strange
Tulloch's case isn't isolated. In mid-2025, OpenAI's Sam Altman publicly said Meta was offering his researchers signing bonuses reaching $100 million each, part of an effort Zuckerberg reportedly called "The List" — a roster of specific researchers he wanted poached. Meta pushed back on the framing; one researcher who moved, Lucas Beyer, called the $100 million signing-bonus figure inaccurate, and Meta clarified that the large totals described multi-year total compensation for a small number of senior roles, not up-front cash. Separately, Meta reportedly offered former Apple AI executive Ruoming Pang a package exceeding $200 million to join its new superintelligence unit — and according to Reuters, Pang has since left Meta for OpenAI, only around seven months after joining.
Even setting aside the disputed headline numbers, what's undisputed is that ordinary compensation at the frontier labs has been reset. Data from Levels.fyi, a platform where employees self-report pay, shows median total compensation for OpenAI software engineers near $870,000 a year, climbing past $1.37 million at senior levels, with research scientists reportedly earning more again. At Anthropic, a Staff Software Engineer reports total compensation around $1.25 million, the large majority of it in equity rather than cash. These aren't outlier celebrity packages — they describe the standard offer for experienced engineers at either company.
Notable Moves in 2026
- Andrew Tulloch: Meta → Anthropic (after previously OpenAI → Thinking Machines → Meta)
- Ruoming Pang: Apple → Meta → OpenAI, in under a year
- John Jumper (Nobel laureate, AlphaFold co-creator): Google DeepMind → Anthropic, confirmed by Reuters
- At least 22 professors from institutions including Stanford, Berkeley, and Harvard left academia in 2026 for positions at the major labs
Why Money Alone Isn't Winning the War
A widely cited 2025 industry talent report found something counterintuitive underneath all this spending: OpenAI engineers were reported to be roughly eight times more likely to leave for Anthropic than the reverse, and the ratio at Google DeepMind favored Anthropic by an even wider margin. The same report put Anthropic's two-year retention rate at around 80%, ahead of DeepMind near 78% and OpenAI around 67%, a figure closer to large, established tech companies like Meta than to a fast-moving research lab. If pure compensation determined loyalty, the flows should look far more balanced than that.
What seems to matter alongside pay: mission alignment, research autonomy, and how a researcher's work will actually be used. Interviews and reporting around several of these moves point to specific technical mandates, not just bigger numbers, as the deciding factor — Tulloch's move, for instance, was reported around a specific team focused on inference and performance work, not a blanket pay bump.
The Cost Beneath the Balance Sheet
This war shows up in company financials too. Analysis of Meta's filings found stock-based compensation running at roughly 12.6% of revenue, nearly double Alphabet's proportion and about four times Microsoft's or Amazon's, alongside a year-over-year increase reported above 50%. That's a direct, measurable cost of trying to out-bid rivals for a limited pool of frontier researchers, and it's a cost investors are now watching as closely as capital expenditure on chips and data centers.
Why It's Escalating Now
The underlying scarcity is structural. There are only so many researchers with hands-on experience training frontier-scale models, and as every major lab races toward the next generation of systems, that small pool becomes more valuable, not less. Universities are losing faculty to the same dynamic, thinning the pipeline of new PhDs even as demand for their skills grows. Meanwhile, "rolling handcuffs" — new equity grants layered on top of existing unvested stock so an employee always has multiple years of value tied up — have become standard practice specifically to counter the kind of rapid re-poaching Tulloch's case represents, though as his move shows, they don't always work.
The Bottom Line
The AI talent war gets covered as a story about eye-popping compensation numbers, but the more revealing story is how often that money fails to buy lasting loyalty. Billion-dollar packages, seven-month tenures, and researchers moving toward labs with clearer missions rather than simply bigger checks all point to the same conclusion: in a market this scarce, the labs that keep their best people aren't necessarily the ones paying the most, but the ones giving researchers the clearest reason to stay. Disclosure: this article discusses Anthropic, which makes the assistant that drafted it — several of the moves described above are reported to have benefited that company.
Follow reviewsideal for ongoing coverage of the people building the AI industry, not just the models.