SAN FRANCISCO – John Jumper spent nine years at Google DeepMind building the AI system that earned him a Nobel Prize. On June 19, he walked out. The week that followed produced three more departures to the same rival, a delay to Google’s most anticipated AI model, and a single-session wipeout that cost Alphabet more than $270 billion in market value. It was the company’s worst trading day in over a year.
The exits accelerated over six days. Noam Shazeer, the vice president of engineering who co-led Gemini and co-authored the 2017 paper “Attention Is All You Need” that introduced the transformer architecture underlying most modern AI, announced he was leaving for OpenAI on June 18. Google had spent an estimated $2.7 billion in 2024 to license his startup Character AI and bring him back inside the company. That investment lasted less than two years. Jumper followed the next day, heading to Anthropic. Bloomberg reported on June 24 that Jonas Adler, who worked on Google’s AI coding tools, and Alexander Pritzel, who was involved in pretraining, were also moving to Anthropic. On June 25, Arthur Conmy, a senior research engineer who contributed to Gemini 2.5 and had worked on AI safety, posted on X that he too was joining Anthropic.
Four of those five departures now share the same destination. Their collective history points directly at Gemini’s architecture. Jumper led the AlphaFold team that won the 2024 Nobel Prize in Chemistry alongside DeepMind chief Demis Hassabis; Adler was among the AlphaFold collaborators before moving to AI coding; Pritzel contributed to model training. These are not peripheral figures leaving for lifestyle reasons. They are the researchers whose work defined what Google could point to as frontier science.
The financial architecture behind the exits is straightforward even if the personal motivations are not. Anthropic and OpenAI are both approaching initial public offerings. Pre-IPO equity at a company valued in the hundreds of billions offers a payday that even a senior Google salary with vested shares cannot easily match. Shazeer’s $2.7 billion retention deal is the sharpest evidence that Google understood the asymmetry. It still was not enough.
Alphabet’s market reacted immediately. Shares fell more than 5% on June 22 as investors digested the first two departures, wiping $225 billion in market capitalization in a single session. D.A. Davidson analyst Gil Luria described it as a sign that Google was “losing the war for talent at the frontier of AI.” CNBC reported it was the company’s worst trading day in over a year. By the time the Adler and Pritzel departures were confirmed two days later, the total market value erased across the week had climbed past $270 billion.
The product calendar has already slipped. Google pushed the general availability of Gemini 3.5 Pro from June to July, citing quality refinements. The company has not said whether the delay is connected to the personnel changes. What Bloomberg’s reporting reveals is a specific internal concern: that Google lacks a competitive product for companies building AI coding tools, the segment where Adler had been working and where Anthropic has made its most aggressive commercial inroads. As Eastern Herald has reported, AI’s voracious demand on memory and inference infrastructure is now inflating hardware costs across the industry, and the companies that can most efficiently deploy that infrastructure have become the gravitational centers for the people who know how to build it.
Google DeepMind is not emptying out. Hundreds of researchers remain, and Hassabis has spent a decade building one of the deepest scientific AI research organizations in the world. But the departures follow a longer trajectory. David Silver, one of DeepMind’s earliest employees and a leading authority on reinforcement learning, left earlier this year to start a company called Ineffable Intelligence. TechCrunch reported the exits have been ongoing for months, accelerating as the IPO windows at rival companies opened and the equity stakes became concrete rather than speculative.
What remains unclear is what Jumper or his new colleagues will actually work on at Anthropic. The company has not announced their roles. Jumper’s expertise is in structural biology and protein prediction; what that means for a company whose core product is a large language model is not immediately obvious. One reading is that Anthropic is building research depth for a next phase of AI that goes well beyond text and code. Another is that it is hiring Nobel laureates to prevent those researchers from working for anyone else. The two explanations are not mutually exclusive.
The harder question is what the departures mean for Gemini 3.5. Pritzel worked on pretraining, the phase of model development that most directly determines a model’s capability ceiling. A pretraining expert does not leave mid-cycle without someone else picking up that work, and that transition takes time the July release date does not obviously account for. Whether Gemini 3.5 Pro arrives carrying the capabilities Google intended, or a version that reflects an interrupted development process, is not a question the delay announcement answers. Microsoft, facing its own AI investment questions, has seen its stock slump as markets weigh whether capital deployed into AI infrastructure will produce returns before the next model cycle obsoletes the bet.
Shazeer’s arc remains the sharpest illustration of Google’s structural problem. The company paid roughly $2.7 billion to bring him back. He stayed less than two years. Whatever held him at Google long enough to justify that price was worth less, in the end, than the equity upside at OpenAI. If that calculus applied to the person Google valued at $2.7 billion, it applies to everyone below that threshold. And there are a great many people below it.
