Why DeepMind researchers are leaving Google for AI rivals

In the space of a single week in June 2026, Google's artificial-intelligence division lost four of its most senior researchers to its two fiercest rivals. The names carry unusual weight: one co-invented the Transformer, the architecture beneath nearly every modern large language model; another holds a Nobel Prize. Individually, each exit is a story. Compressed into seven days, they read as a structural signal about where frontier talent now believes the value — and the equity — is accumulating.

The most symbolically loaded departure is Noam Shazeer, a co-author of Attention Is All You Need, who is heading to OpenAI; the move was announced on 18 June. Two days later, John Jumper — a DeepMind director and, since 2024, a Nobel laureate in chemistry for the AlphaFold protein-structure breakthrough — confirmed a move to Anthropic. Then, on 24 June, Bloomberg reported that two further researchers, Jonas Adler and Alexander Pritzel, were also bound for Anthropic. A fifth, Arthur Conmy, said on X that he would join Anthropic to work on AI safety.

The exits are not all immediate. Jumper may not begin at Anthropic until 2027, held back by the non-compete terms that apply to DeepMind leadership — a reminder that a headline departure and an actual desk change can be many months apart, and that Google retains some contractual friction even as the announcements land.

The clearest incentive is money that has not yet been minted. Anthropic recently raised $65 billion at a $965 billion valuation, making it the world's most valuable private AI company — ahead of OpenAI — and both labs are widely seen as approaching public listings. For a researcher weighing a move, pre-IPO equity at that scale is a lever Google's public stock cannot easily match. A Nobel laureate and a Transformer author are not leaving for a title; they are leaving for a cap table before the offering.

Compensation is not the whole story. Reporting points to internal DeepMind tensions — compute allocated to one of Shazeer's projects was reassigned to a London team before he left — alongside strategic fit. Anthropic is expanding aggressively into coding, health and science, precisely the areas where Jumper (protein science), Adler (AI coding) and Pritzel (pretraining) do their strongest work. The pull is as much about mission surface area as about the cheque.

The market registered the risk quickly. Alphabet shares fell roughly 5-6% on 22 June, with analysts tying the drop to a combination of heavy AI-spend concerns and questions about talent retention. Losing four decorated researchers in a week is not, by itself, a capability collapse — but it is the kind of signal that reprices a narrative about who is winning the frontier.

Google's leadership pushed back directly. Speaking at an event in Cannes, DeepMind chief executive Demis Hassabis argued that talent moves in every direction between the leading labs and that Google still wins its share, framing its research depth as a durable advantage rather than a leaking asset. The honest read sits between the two claims: bench depth is real, but so is the pattern of marginal, high-signal exits.