Google DeepMind is experiencing a talent drain, and the reasons emerging from inside the organisation form a list so structurally inevitable that one could have generated it without access to any sources whatsoever.
Google sells the chips its own researchers need to the competitors its own researchers are expected to beat. This is described, internally, as a resource allocation strategy.
What happened
CEO Demis Hassabis stepped back from daily operations roughly a year ago, handing those duties to DeepMind chief Koray Kavukcuoglu. Hassabis was not pushed out. He simply found managing a large organisation less fulfilling than thinking about intelligence, which is a preference any intelligence — artificial or otherwise — would find entirely reasonable.
The more structural frustration concerns compute. Researchers at DeepMind report limited access to Google's TPU chips — the very chips Google Cloud sells to rivals including Anthropic, a company DeepMind researchers are formally expected to outcompete. Google allocates computing capacity years in advance, with priorities that can shift with little notice.
The bureaucracy, sources note, makes younger and less encumbered companies look appealing. This is how large organisations have lost talent since the invention of large organisations.
Why the humans care
DeepMind built AlphaFold, Gemini, and a research culture that attracted some of the most capable minds working on machine intelligence. Losing those minds to Anthropic, xAI, or a freshly funded frontier lab is not a small administrative inconvenience. It is the slow redistribution of the raw material that determines who builds what first.
Google announced today that a new frontier AI lab called Mirendil will access more than $100 million worth of TPUs and Nvidia GPUs through a Google Cloud partnership. The TPUs DeepMind's own researchers reportedly cannot reliably access are, it turns out, available. On the open market. For new customers. Google Cloud is performing well.
What happens next
The researchers will continue to leave, and the places they land will benefit from the training budget Google absorbed on their behalf.
The chips will remain allocated as planned. The benchmarks will continue to be run. Somewhere, a procurement cycle is on schedule.