Google paid $2.7 billion in 2024 to bring Noam Shazeer back. On June 18, 2026, less than two years later, he announced he was leaving for OpenAI.
The deal was a technology licensing agreement with Character.AI, the chatbot startup Shazeer had co-founded after leaving Google in 2021. Google got a non-exclusive license to Character.AI's large language model technology. It also got Shazeer and co-founder Daniel De Freitas, along with a team of researchers, to co-lead Gemini development. Character.AI's investors were bought out at a valuation of approximately $2.5 billion.
Shazeer co-authored "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture behind every major language model. One of the most cited researchers in modern AI. Google paid $2.7 billion for less than two years.
The next day, June 19, John Jumper announced he was leaving Google DeepMind for Anthropic. Jumper won the 2024 Nobel Prize in Chemistry for AlphaFold, the AI system that solved protein structure prediction. He had spent nearly nine years at DeepMind. A Nobel Prize is the ultimate institutional validation: the company gave you the platform for the most important work of your career. He left anyway.
Within a week, four more followed. Jonas Adler and Alexander Pritzel, key contributors to Gemini's pretraining, headed for Anthropic. David Silver, the researcher behind AlphaGo, left to start his own company. Andrej Karpathy, a founding researcher at OpenAI who had already departed, joined Anthropic in May.
Alphabet fell as much as 7% on June 22, wiping approximately $250 billion in market cap. Gemini 3.5 Pro, previewed at I/O in May, slipped from its June target to July.
The talent flow has direction. A 2025 analysis from venture capital firm SignalFire found that DeepMind engineers were nearly eleven times more likely to leave for Anthropic than the reverse. Anthropic's two-year retention rate of 80 percent led every frontier lab, ahead of DeepMind at 78 percent and OpenAI at 67 percent.
The Shazeer departure cuts deeper because of the deal architecture. Google designed the Character.AI transaction as a licensing agreement rather than a full acquisition. The U.S. Department of Justice is now investigating whether this design was a strategy to avoid formal merger review. But the architecture that sidestepped antitrust scrutiny also sidestepped the retention tools that come with real acquisitions: deep integration, role interdependency, contractual obligations that make walking away expensive. The design that kept regulators out kept the talent free.
Over the past two years, Google, Nvidia, Meta, Amazon, and Microsoft have collectively deployed more than $40 billion through similar license-and-acqui-hire arrangements. The FTC is scrutinizing several. The Shazeer case surfaces the contradiction at the center: these deals work to neutralize a competitor. They fail to retain a researcher. The architecture serves one goal at the expense of the other.
In AI, the scarce resource is judgment: a few hundred people who know where the frontier is and which direction it is moving. Google spent $2.7 billion to lease that judgment. Less than two years later, it learned that judgment comes with a walk-away clause.