Six prominent AI researchers have left Google DeepMind since early 2026 for Anthropic, OpenAI, and Meta, wiping roughly $270 billion from Alphabet's market cap and raising hard questions about whether Google's infrastructure advantages can outlast a talent crisis at the top.
The week of June 18 was not a good one to hold Alphabet shares. Noam Shazeer, a co-author of the 2017 "Attention Is All You Need" paper and co-lead of the Gemini project, announced he was heading to OpenAI. Two days later, John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, said he was joining Anthropic. On the Monday that followed, as Fortune and Axios reported the scale of the exodus, Alphabet shed $225 billion in market value in a single session, its largest one-day capitalization loss on record. By the end of that week, two more Gemini researchers, Jonas Adler and Alexander Pritzel, were confirmed as Anthropic-bound, according to Bloomberg. Denny Zhou, who founded Google Brain's reasoning research team, had already been at Meta's Superintelligence Lab for roughly four months. The total, per TechTimes, is at least six high-profile departures since February 2026.
The proximate cause, by most accounts, was a strategic pivot DeepMind made in April. Sergey Brin, newly hands-on at Google again, pushed the formation of an AI coding strike team under Sebastian Borgeaud, a research engineer who had previously run Gemini's pretraining. The mandate was blunt: close the gap with Anthropic's Claude on agentic coding tasks, where Gemini had fallen behind. The problem is that closing that gap required pulling resources away from longer-horizon frontier research, including world model work and the kind of foundational science that attracted many of DeepMind's best people in the first place. The researchers who disagreed with that reordering voted with their feet.
There's also a financial arithmetic that Google simply can't match right now. Anthropic and OpenAI are pre-IPO companies, and the equity they can offer researchers is still, in the right circumstances, enormously valuable. Shazeer's situation makes the cost plain: Google spent approximately $2.7 billion to reacquire him through a licensing deal with Character.AI in 2024, and he was gone in under two years. That's not a retention strategy. That's a revolving door with very expensive hinges.
Don't write off Google's infrastructure edge too quickly. Anthropic, for all the talent it's absorbing from DeepMind, committed to purchasing up to one million TPU chips from Google in a deal worth tens of billions of dollars, which is a remarkable arrangement: Google's most aggressive talent competitor is also one of its largest infrastructure customers. Google confirmed general availability of its seventh-generation Ironwood TPU earlier this year and has previewed an eighth-generation architecture with two purpose-built chips, one for large-scale training and one for inference, both targeting TSMC 2nm manufacturing. Sundar Pichai cited a 78% reduction in Gemini serving unit costs across 2025 as evidence that the vertical integration is actually working. No independent AI lab can match that cost curve.
The data moat is real too, even if harder to quantify. Search query history, YouTube watch patterns, Gmail context: the longitudinal proprietary data that Google has accumulated over two decades is the one asset competitors genuinely cannot compress in time. A researcher who leaves DeepMind for Anthropic takes their knowledge with them, but not the data. That's a durable advantage, provided Google can build models capable of exploiting it.
That's the catch. Gemini's models now rank outside the top five on major benchmarks, holding around 13.5% of the global AI chatbot market against ChatGPT's roughly 60%, according to independent assessments cited by Epoch AI. The people who built Gemini's strongest capabilities are leaving. TPUs and YouTube data are inputs; you still need world-class researchers to turn them into competitive outputs.
Frankly, the more interesting question isn't whether Google can survive this, it clearly can, but whether DeepMind can remain the kind of place that attracts the researchers who do frontier science rather than product engineering. Those are different jobs requiring different cultures, and the coding strike team pivot made DeepMind's priorities legible in a way that previous management ambiguity had obscured. The people who left weren't confused about what Google wanted. They understood it perfectly, and chose differently.
What comes next probably hinges on whether the strike team actually delivers. If Gemini's coding performance closes the gap with Claude by the end of 2026, Brin and Kavukcuoglu will have validated the pivot. If it doesn't, Google will have traded its best frontier researchers for a commercial product that still trails the labs now staffed by those same researchers. That's a bet the company is already committed to, whether it intended to make it or not.
Also read: Suno's Spark artist grants come with a permanent silence clause that says more about the lawsuit than the music • AI coding agents are turning code review into the next startup risk • Micron Technology has become the AI infrastructure bet that most investors are sleeping on