Google's TPU shortage and bureaucracy are driving DeepMind researchers to rivals

The steady stream of researchers leaving Google DeepMind isn’t just about better pay—it’s about being able to do their jobs. CNBC reports that limited access to Google’s own TPU chips is a core frustration, made worse by the fact that Google Cloud happily sells that same compute to direct competitors like Anthropic. Imagine being told to beat a rival while they’re running on hardware you can’t get your hands on.

That tension reflects a deeper structural problem. Google allocates computing capacity years in advance across research, product, and cloud divisions, but priorities shift with little warning, according to a source speaking to CNBC. When your next experiment depends on chip access that can evaporate overnight, the appeal of a younger, more nimble company becomes obvious. The bureaucratic machinery designed to optimize resources across a trillion-dollar enterprise ends up repelling the very talent needed to stay ahead.

Semafor adds another layer: CEO Demis Hassabis stepped back from daily operations about a year ago, handing the reins to Koray Kavukcuoglu. Hassabis wasn’t pushed—he reportedly found management unfulfilling and sees himself as a visionary scientist, not an executive. That leadership vacuum, combined with chip rationing, creates an environment where top researchers question whether DeepMind is still the best place to do cutting-edge work.

The timing of Google’s announcement today that “exciting frontier AI lab” Mirendil will tap over $100 million in TPUs and Nvidia GPUs through a Google Cloud partnership only underscores the irony. Google can fund, equip, and empower external labs with immense compute, yet its own scientists are fighting over scraps. This isn’t a hiring problem—it’s a priority problem. And until Google reconciles its dual identity as both a research powerhouse and a cloud vendor, the exodus will likely continue.