Google's TPU Sales to Anthropic Squeeze Its Own Researchers

There is a particular flavor of internal tension at a big company when the compute you used to run your experiments starts getting sold to your competitors. That is the picture CNBC lays out at Google, where some of the researchers who helped build its AI capabilities are watching TPU capacity go out the door to outside customers, including Anthropic, whose models compete directly with Gemini.

The reported dynamic is straightforward. Google Cloud is monetizing its homegrown AI chips at scale, and some employees have grown frustrated over access to the computing capacity they need to pursue ambitious projects. That is compounded by what CNBC describes as a common source of friction inside Google, the layers of approval required to move research into products. In combination, those two conditions have made emerging companies like OpenAI, Anthropic, and younger startups more appealing to AI researchers who would rather do lab work than manage internal process.

The talent side is the visible symptom. Jonas Adler and Alexander Pritzel, both viewed internally as key contributors to Gemini, are reportedly heading to Anthropic. Adler has been working on Google's AI coding efforts and Pritzel on the process used to train AI systems. That follows earlier high-profile exits, including Nobel laureate John Jumper reportedly moving to Anthropic and star researcher Noam Shazeer going to OpenAI. When the people who wrote the model start walking, the conversation shifts from compensation to whether the work environment still lets ambitious researchers do ambitious work.

Take the specifics as reported, not settled. Much of the piece is sourced to unnamed insiders, so the causal chain from TPU allocation to specific departures is more inference than proof, and Google will have its own view on morale and productivity. What the reporting does not give you is the actual share of TPU capacity going to external customers versus internal research, the current churn rate versus a normal year, or Anthropic's side of the recruiting picture.

The forward question is who benefits. Anthropic appears to get both the chips and, in some cases, the people who understand how to use them best. For anyone tracking the model race, the more interesting signal is that infrastructure spend does not automatically buy internal velocity, and the labs that keep their researchers unblocked may end up shipping faster than the ones with the biggest cloud business.