On Thursday, June 18, Noam Shazeer posted on X that he was leaving Google. For most companies, one researcher's exit is a footnote. Shazeer is not most researchers. He is one of the eight authors of "Attention Is All You Need," the 2017 paper that introduced the Transformer, the architecture underneath nearly every large language model shipping today. Until that post, he was also a co-lead on Gemini, Google's flagship model line.
Then it happened again. Within days, John Jumper, the DeepMind scientist who shared the 2024 Nobel Prize in Chemistry for AlphaFold, announced he was leaving too. His destination was Anthropic.
Two of Google DeepMind's most decorated researchers walked out in roughly 48 hours, one to OpenAI and one to Anthropic, the two rivals Google has spent three years trying to catch.
Investors read the message instantly. Alphabet shares fell more than 5 percent on Monday, June 22, with analysts tying the drop to fresh doubt about whether Google can hold onto the people who build its best models. By midweek the story had grown. Bloomberg reported that two more Gemini researchers, Jonas Adler and Alexander Pritzel, were also heading to Anthropic. That makes four senior departures to direct competitors inside a single week.
This is not a gossip item about who works where. It is a question about the scarcest input in AI, which is not compute or data but the small number of people who know how to turn both into a frontier model. Right now, that talent is moving in one direction, and it is away from Google.
The Pioneer and the Nobel Laureate
Shazeer's history with Google is its own saga. He helped build LaMDA, Google's earliest LLM-based chatbot, in 2021, then left in frustration that the company was too slow to ship it. He and fellow Googler Daniel de Freitas founded Character.ai, a chatbot startup that went viral. In 2024, Google lured him back through a deal that licensed Character's technology for a reported $2.7 billion and installed him atop Gemini. He has now left a second time.
Jumper's exit lands differently. He is not a serial founder chasing the next thing. He shared the Nobel with DeepMind CEO Demis Hassabis for AlphaFold, the system that predicted the three-dimensional shape of proteins from their sequences and solved a problem biologists had chased for 50 years. After the prize, he stayed at DeepMind to push protein modeling further. Leaving for Anthropic is a statement that the scientific opportunity there looked better than the one at the lab where he won the Nobel.
Anthropic CEO Dario Amodei recently told Bloomberg's Emily Chang that the company intends to do more in biology. Hiring the most recognized protein-AI researcher alive is how that intention becomes a roadmap. Anthropic had already pulled OpenAI founding member Andrej Karpathy onto its pretraining team a month earlier, a sign of how aggressively it is recruiting marquee names ahead of a public listing.
Neither man has said publicly why he left. The likeliest reasons have little to do with base salary.
How One Week Unspooled
Why the Startups Keep Winning the Bidding War
The pull is partly financial, but not in the way it first looks. Both Shazeer and Jumper were already wealthy, and Shazeer reportedly made hundreds of millions from the Character.ai deal. The real prize is equity in a company about to go public. Anthropic and OpenAI are both expected to file for IPOs within months, and a researcher who joins before the listing can capture gains that are difficult to match at a company Google's size.
Anthropic has the currency to spend. In late May it raised $65 billion at a valuation near 965 billion dollars, passing OpenAI to become the world's most valuable private AI company. That valuation is also a recruiting tool: it turns pre-IPO equity into a number that even a Nobel laureate has to think about.
The fit is strategic too. Anthropic is expanding past general-purpose chatbots into coding, healthcare, and science, which is exactly the terrain where the departing researchers built their names. Adler worked on Google's AI coding efforts. Pritzel worked on pretraining, the early stage where a model learns from raw data. Both contributed to AlphaFold alongside Jumper, whose move dovetails with Hassabis's own spinout, Isomorphic Labs, the AlphaFold-derived drug company, now competing with Anthropic for the same scientific talent.
There is friction inside Google as well. Shortly before Shazeer announced his move, Google reportedly reassigned computing capacity from one of his projects to another DeepMind team in London, a reallocation meant to consolidate pretraining work. As advanced chips stay scarce, who controls them has become a live source of tension.
What the Exodus Says About Google DeepMind
The departures would sting less if Google were clearly ahead. It is not. Its top models, Gemini 3.5 Flash and Gemini 3.1 Pro, frequently sit outside the top five on public benchmark leaderboards, trailing systems from Anthropic, OpenAI, and Chinese labs such as Z.ai and MiniMax. Its release cadence has slowed: roughly four months separated Gemini 3.1 Pro in February from the wide release of Gemini 3.5 Pro. Over the same stretch, Anthropic shipped two Claude Opus updates and introduced an entirely new model family, Mythos, built for long-horizon autonomous work.
Current and former employees describe a familiar culprit: size. They portray a culture that is bureaucratic and highly risk-averse, the same critique Shazeer is thought to have made in an anonymous internal memo years ago, before his first departure. The memo leaked. ChatGPT launched soon after. The critique aged well.
The Other Side
Google's defenders make a real case. Hassabis has said that movement between leading labs is normal in the current market, and he argued that Google still has the largest and broadest research organization in the field. The company said it remains confident in its ability to attract and keep AI talent. Not every recent exit is a defection, either: David Silver, one of DeepMind's earliest employees and a leading reinforcement-learning researcher, left to start his own company, Ineffable Intelligence, rather than to join a rival.
There is a structural argument too. Alphabet has billions of users and real profits to defend, which means it cannot make the same reckless bets that venture-funded labs can. As Fortune's Jeremy Kahn framed it, Google may be playing not to lose rather than playing to win, a strategy its enormous distribution advantage can sustain even if it never reclaims the top leaderboard spot. The weakness in that logic is cultural: playing not to lose is rarely what keeps Nobel laureates in the building.
The Bottom Line
Strip away the names and the stock move, and the situation is simple. The rarest resource in AI is the handful of people who can build a frontier model, and in one week four of them chose Google's rivals. A Transformer co-author went to OpenAI. A Nobel laureate and two of his AlphaFold collaborators went to Anthropic. Money explains part of it. The chance to ship faster, with fewer layers of approval and a clearer path to generational wealth, explains more.
Hassabis, a former chess prodigy who hates losing, will not take this quietly. The open question is whether he can move a company large enough to slow down its own stars before more of them update their profiles. For now, the people who helped invent the Transformer and AlphaFold have already placed their bets on where the frontier is headed. At the moment, it is not pointed at Mountain View.
Sources
- Defections from Google DeepMind prompt questions about Alphabet's efforts to stay at the forefront of AI — Fortune, Jun 23, 2026
- Google set to lose two AI researchers to Anthropic and OpenAI amid growing talent exodus — Crypto Briefing, Jun 24, 2026
- Google Poised to Lose Two More High-Profile AI Staffers to Anthropic — Bloomberg, Jun 24, 2026
- Alphabet falls after DeepMind's John Jumper exits to join Anthropic — Seeking Alpha, Jun 2026
- Google Loses Two Top AI Researchers To OpenAI & Anthropic — Search Engine Journal, Jun 2026
- Noam Shazeer announcement — X, Jun 18, 2026
- John Jumper announcement — X, Jun 2026