Jeff Dean, Google's chief scientist and the engineer who built the infrastructure under Google's entire AI stack (MapReduce, TensorFlow, the TPU program), is leaving to start a company with a group of senior researchers from Google DeepMind, Google's AI research lab formed from the 2023 merger of DeepMind and Google Brain. The departures land weeks after Google DeepMind CEO Demis Hassabis restructured the unit's leadership, and they cost Google the person most responsible for the TPU roadmap, the custom chip design that lets Google train large models without buying Nvidia GPUs. That matters because TPU access is the one input Google does not have to fight OpenAI or Anthropic for in a GPU market where Nvidia allocates supply lab by lab.

The new venture reportedly centers on "discovery loop" research, using AI systems to accelerate scientific discovery itself rather than chatbots or coding tools, putting it in different territory from Anthropic's Claude or OpenAI's GPT line. Google now has to defend its TPU program's technical lead (the seventh-generation Ironwood TPU, detailed at Google I/O in May 2026) without the architect who built the six generations before it, while a rival lab launches with a research agenda Google's own infrastructure enabled. The next signal is who Hassabis names to run TPU strategy, and whether Google DeepMind's remaining researchers follow Dean out the door before that seat is filled.