Google Lost Two of Its Greatest AI Researchers in 48 Hours — And Alphabet Paid $250 Billion for It

Forty-eight hours. That's how long it took Google to lose the co-author of the Transformer architecture and a Nobel Prize winner to rival AI labs — and shed approximately $250 billion in market capitalization in the process.

On June 18, 2026, Noam Shazeer — who co-authored "Attention Is All You Need," the 2017 paper that underpins every frontier AI model in production today — announced he was leaving Google for OpenAI. The following day, John Jumper, 2024 Nobel laureate and co-creator of AlphaFold, posted his departure from Google DeepMind for Anthropic. Alphabet's stock fell 6.8% on Monday, June 22, as the market opened and processed both announcements together. The Nasdaq dropped 1% in sympathy.

Neither move was a surprise to anyone paying attention to Google's internal climate. Both had circulated as rumors for several weeks. What was surprising was the compression — two departures of that magnitude in 48 hours, back-to-back, with no rebuttal hiring announcement from Google to absorb the optics. The market wasn't pricing in the individual departures so much as the implied signal: that Google's ability to retain the humans who build frontier AI is structurally weakening.

Shazeer: The $2.7 Billion Man Who Stayed 22 Months

Google paid approximately $2.7 billion in September 2024 to acquire Character.AI — and specifically to bring Noam Shazeer back. He had left Google in 2021 to co-found Character.AI with Daniel De Freitas. The acquisition wasn't primarily about Character.AI's product. It was about re-securing Shazeer before OpenAI, Anthropic, or any other lab could. Google's co-founder Sergey Brin reportedly returned to the office in 2022 to help Google respond to the ChatGPT moment. Shazeer's return was part of the same response.

That decision made sense in 2024. Shazeer isn't just a transformer contributor in the citation sense — he's the architect of multi-head attention as it actually runs at scale. He's credited with Mixture-of-Experts innovations, efficient attention variants, and deep intuition about how architectural choices at the design stage compound into training efficiency across weeks-long runs. That class of knowledge is extremely rare. It doesn't live in papers. It lives in the decisions that weren't made — in the shape of experiments that failed quietly before a working configuration was locked in.

He lasted less than 22 months.

Sam Altman's X post arrived the same morning Shazeer's departure was confirmed: "noam is one of the people I have most wanted to work with since the very beginning of openai. only took 10 years. i think it will be worth the wait." The understatement is characteristic. Shazeer had been on OpenAI's wishlist since 2015. The message landed publicly before Alphabet's market opened, which is part of why the June 22 session started as badly as it did.

Beyond research output, what OpenAI is acquiring is Shazeer's institutional knowledge of Gemini's architecture. He spent 18 months making decisions about what Gemini's training runs look like, where the efficiency tradeoffs sit, and which architectural directions were tried and abandoned. None of that is in any published paper. It lives in his head. Competitor intelligence at that depth is a second, less discussed dimension of what the hire means for OpenAI's GPT-6 trajectory.

Jumper: What DeepMind Gave Away

John Jumper spent nearly nine years at DeepMind. In that time, he and his team built AlphaFold — the protein structure prediction model that solved a problem biologists had worked on for fifty years. He and Demis Hassabis shared the 2024 Nobel Prize in Chemistry for it. The prize was awarded for a genuine scientific contribution at a level AI models had never previously achieved. Not "assisted researchers in identifying protein structures." Predicted them with accuracy sufficient to replace laboratory methods across most use cases.

His June 19 departure post was brief. He thanked colleagues, mentioned "taking time to recharge," and didn't describe a role or product context at Anthropic. Anthropic has not released a job title. Early speculation about a "ClaudeFold" — some AlphaFold-equivalent layer built into Claude for scientific reasoning — is hype until Anthropic confirms scope at the June 30 event, where several new product directions are expected to be announced.

What is confirmed: Anthropic now has a Nobel Prize winner on staff. That is not primarily a technical asset — it's a scientific legitimacy asset that no benchmark score can replicate. Nobel recognition signals that AI can make genuine scientific contributions, not merely assist with existing workflows. Jumper's presence positions Anthropic to compete for scientific institution partnerships, government research contracts, and pharmaceutical industry deals in ways that require that kind of credentialing. A "Claude for Drug Discovery" conversation lands differently with Merck if Jumper is in the room.

The concrete loss for Google: Jumper had been working on AI-native coding tools at DeepMind — a domain where DeepMind has the research depth but has struggled to turn that depth into commercially deployable products. His departure removes DeepMind's highest-profile name from that team two years into what should have been the execution phase.