Four Senior Gemini Researchers Left Google in Six Days

Google just lost four senior AI leaders in six days — the transformer co-author who co-led Gemini, the Nobel laureate who built AlphaFold, Google's AI coding lead, and a pretraining specialist who helped train the models. The exodus landed the same week Gemini 3.5 Pro missed its June launch window and Alphabet shares shed roughly $225 billion in a single session. Demis Hassabis says Google still has the industry's "biggest and broadest research bench." Wall Street is pricing the direction that bench is moving.

The six-day bleed:

• June 18: Noam Shazeer — co-author of the 2017 transformer paper and Gemini co-lead — joins OpenAI, which filed confidentially for an IPO on June 8.
• June 19–20: John Jumper — 2024 chemistry Nobel laureate and AlphaFold lead — announces he is leaving for Anthropic.
• June 24: Bloomberg reports Jonas Adler (Google's AI coding lead) and Alexander Pritzel (pretraining specialist, AlphaFold contributor) are also heading to Anthropic.

Why Adler and Pritzel sting: Shazeer invented the architecture. Jumper proved AI can crack science. Adler and Pritzel run the two pipelines that actually ship capability — the coding stack that competes with Claude Code and the pretraining phase where frontier models are born. Anthropic is not poaching celebrities. It is taking the operating system.

The compute trigger: Bloomberg reported that shortly before Shazeer announced his OpenAI move, computing power dedicated to one of his projects was reassigned to a London-based DeepMind team consolidating pretraining work. The reallocation was meant to boost collaboration. From a researcher's chair, it looked like GPUs pulled mid-experiment. At frontier labs, compute access is velocity. Lose the cluster, lose the job.

The IPO lottery: Business Insider framed the simpler math: Google RSUs track a $4 trillion public company with predictable upside. Anthropic and OpenAI offer pre-IPO equity grants that could multiply if listings land in late 2026 or 2027. Shazeer already cashed hundreds of millions when Google paid roughly $2.7 billion to reacquire him through Character.AI in 2024. Twenty months later, he left again — this time for OpenAI's S-1 pipeline. SignalFire data cited by Bloomberg show DeepMind engineers are nearly 11 times more likely to leave for Anthropic than the reverse.

Gemini 3.5 Pro slips: Business Insider reported on June 24 that Google's next frontier model will not reach general availability in June as CEO Sundar Pichai promised at I/O on May 19 — "give us until next month." Google is now targeting July while gathering feedback from Antigravity testers and fixing token-burn issues flagged in Gemini 3.5 Flash. A Google spokesperson declined to comment. The delay is narratively manageable. Losing your coding lead and pretraining specialist the same week is not.

Markets noticed first: Alphabet shares fell as much as 7% on June 22 before closing down 5% — the stock's largest one-day percentage drop in more than a year and roughly $225 billion in erased market value, per MarketWatch and Indian Express reporting. Hassabis addressed the anxiety two days later at Cannes Lions: "There's a lot of talent movement between all the leading labs and we win our fair share of the top talent. We have by far the biggest and broadest research bench of any of the labs out there." He called it "the most ferociously competitive" hiring market tech has ever seen. A Google spokesperson pointed reporters to those remarks when asked about Adler and Pritzel.

What Google still has: Hassabis remains. Google DeepMind still employs thousands of researchers. Gemini 2.5 Pro with Deep Think posted strong benchmark scores on June 22. TPUs and custom silicon keep improving. Google's counter is scale — and scale is real. But scale without the people who defined your last two breakthroughs is a different kind of moat.

Convina's view: Individual departures are normal in a talent war. Four in six days — spanning architecture, science, coding, and pretraining — is a pattern, not noise. Anthropic is assembling the team that builds models while Washington holds the off switch on its best ones; OpenAI is assembling the team that sells them to the government customer-by-customer. Google is left insisting the bench is deepest while the people who staffed the starting lineup collect pre-IPO equity elsewhere. Hassabis may be right that Google still wins its share. The share that matters right now is walking out the door with a July deadline on the product they were building.