Google DeepMind Lost the Transformer’s Author in a Week

Noam Shazeer co-wrote the eight-page paper that every large language model on earth is descended from. Last week he left Google for OpenAI. A day later John Jumper, who won a Nobel for cracking protein folding at DeepMind, left for Anthropic. Google didn’t lose a couple of employees. It lost the people who made it the place everyone else was trying to catch.

The week’s tally is brutal in the specifics. Shazeer announced his move on June 18. Jumper announced his the next day, after roughly nine years at DeepMind, as Search Engine Journal walked through. Bloomberg reported that two more Gemini researchers, Jonas Adler and Alexander Pritzel, were also headed to Anthropic, which puts four senior departures to direct competitors inside a single week. Fortune tied the run of exits to real doubt about whether the lab can hold its position at the front of the field. The market reacted faster than the press did.

Alphabet shares fell about 5 to 6% on June 22, and the reporting tied the slide to exactly this mix: AI-spending fears stacked on top of a fresh question about whether Google can keep its own people, according to TechTimes. That drop wasn’t about one quarter’s numbers. It was the market repricing something it had quietly treated as permanent: the assumption that the best researchers alive would default to Google because Google was where the compute lived. The moment that stopped being automatic, a few hundred billion in market cap remembered it was only ever theoretical.

Look at where they landed. Shazeer to OpenAI, Jumper and at least two others to Anthropic. A Chemistry laureate choosing Anthropic over the lab where he did the prize-winning work is a statement about where the frontier feels alive right now, and it is not inside the company that invented the Transformer and then watched its co-author leave to scale the idea somewhere else. Google still has more compute, more data, and a deeper bench than anyone in the race. What it apparently can no longer promise is that the person who builds the next breakthrough wants to build it under Google’s roof.

The why is the uncomfortable part. People at this level don’t move for the salary, because the salary is already absurd everywhere. They move for autonomy, for the feeling that the most important work in the field is happening at their own desk and not three reorgs away. Google spent the past two years folding its research groups together, shipping Gemini on a corporate cadence, and reminding everyone that a search-ad business still pays for the science. For someone deciding where the next decade of their career goes, that reads as a place where the work serves the ad revenue instead of the other way around. OpenAI and Anthropic are smaller, less stable, and worth far less on paper, and that is exactly the pitch: come build the thing rather than maintain it.

Be concrete about what Google actually lost, because “four researchers” undersells it badly. Shazeer didn’t just co-write the Transformer paper and move on; he kept building, and the architecture sitting under every modern chatbot traces through work he led. Jumper’s AlphaFold rewired how an entire scientific field approaches protein structure and won a Nobel doing it. People like this are not interchangeable headcount; the next decade of recruits will read their résumés as a map of where the real work is happening. Google can backfill the seats, but the map is the part it just lost control of.

The gain compounds on the other side of the trade. Anthropic now has the Transformer’s lineage and a Nobel laureate in a single recruiting cycle, and that pulls the next tier of researchers along behind it. Talent clusters; it always has. The first defection is a headline, the fourth is a pattern, and the fifth through tenth are people quietly revising their own assumptions about where they ought to be sitting. The bigger risk to Google is the people who stayed: they watched four of the best walk out to the competition in a week and clocked that nobody could stop it.

The bull case for Google is real and worth saying out loud. Stars leave; deep organizations keep shipping. DeepMind’s pipeline runs hundreds of researchers deep, its infrastructure advantages still tower over the value of any single hire, and Gemini keeps coming out no matter whose name is on the next paper. Labs have survived the loss of a famous founder before. All of that holds. But losing a celebrity researcher is ordinary attrition, and losing four senior people to the two labs you are directly racing, in seven days, while your stock drops in real time, is the kind of thing a competitor’s recruiter quotes back to candidates for the next year.

There’s a deeper irony in the timing. Google did more than employ these people; it produced the ideas that made them valuable. The Transformer came out of Google in 2017. AlphaFold came out of DeepMind. Google has been the lab that turns blue-sky research into the tools everyone else commercializes, and its reward has been watching the commercializers hire its researchers away to do exactly that, faster. The next architecture, whatever finally succeeds the Transformer, now has better odds of being born at OpenAI or Anthropic than at the company that birthed the last one. That is what the stock drop was really pricing: a question about who owns the future of the field Google arguably started, not a single soft quarter.

Google spent a decade as the gravity well of AI research, the place ambition naturally fell toward. Last week it learned gravity can run backward. The papers will keep coming and Gemini will keep shipping, because that is what large, rich labs do. But the story Google told about itself, that the smartest people in the room would always be its people, just lost four of its best witnesses in a single week. This is the talent front of a race nobody is actually winning yet, and money buys a lot at this altitude. Loyalty, it turns out, isn’t reliably on the menu.

Sources: Fortune · Search Engine Journal · TechTimes