Demis Hassabis has stepped back from daily operations at Google DeepMind, handing leadership to Koray Kavukcuoglu because he views himself as a scientist rather than an executive. While Hassabis was not forced out, reports indicate many researchers are frustrated by restricted access to Google’s TPU chips. This limitation is compounded by a conflict of interest, as Google Cloud sells the same hardware to rivals like Anthropic. Computing capacity is allocated years in advance among research teams, product operations, and cloud customers, yet priorities can shift with little notice. This bureaucratic rigidity reportedly makes younger companies more appealing to talent. Today, Google announced that Mirendil will use more than $100 million worth of TPUs and Nvidia GPUs through a Google Cloud partnership.
The situation highlights a structural friction where internal research competes with commercial cloud revenue. Researchers require consistent compute power to train models, but sales teams prioritise external clients. This dynamic creates an environment where top talent seeks stability elsewhere. The loss of key figures could slow progress on frontier AI projects within the organisation.
- Computing priorities shift without notice
- Internal teams compete with external cloud sales
- Researchers seek stability in smaller firms
