Google DeepMind Researchers Quit as Cloud Deals Squeeze Internal Compute Access

The Internal Paradox

Google has assembled what is arguably the most formidable AI infrastructure stack in the industry — custom TPU chips, a booming cloud business, and multi-billion-dollar supply contracts with the very rivals trying to beat it. But that success has produced a strange internal paradox: Google's own researchers reportedly cannot get enough time on the company's machines. According to a Bloomberg report, DeepMind scientists are now competing for compute resources that their employer is simultaneously renting out to outside customers, and a wave of high-profile departures has followed.

The Contractual Squeeze

The root cause is contractual. Google has committed up to $40 billion in infrastructure to Anthropic, including five gigawatts of TPU capacity over five years and access to as many as one million seventh-generation Ironwood chips. A separate arrangement reportedly covers Meta. Those commitments tie up capacity that internal model teams can only access by queueing behind paying customers. DeepMind chief Demis Hassabis has publicly acknowledged the strain, noting that researchers need a lot of chips to experiment on new ideas at sufficient scale. Some of the bottleneck comes from constrained supply of high-bandwidth memory from Samsung, Micron, and SK Hynix, but the internal allocation decisions are entirely Google's own.

High-Profile Departures

The consequences have been visible in the talent market. Andrew Dai, a 14-year Google veteran who co-led data work on Gemini and pre-training for PaLM 2, left in January 2026 to co-found Elorian, a multimodal reasoning lab that raised $55 million. Ioannis Antonoglou, a founding DeepMind engineer behind AlphaGo and MuZero, co-founded ReflectionAI, which has since pulled in $2 billion to build open-weight frontier models. Anna Goldie, co-lead of AlphaChip and a former Gemini researcher, launched Ricursive Intelligence in late 2025, raising $300 million at a $4 billion valuation to use AI for chip design. Each has reportedly found securing compute easier outside Google — even though Google itself spent close to $14 billion on capital expenditure in a single recent quarter.

An Industry-Wide Problem With a Twist

Oren Etzioni, the former head of the Allen Institute for AI, described the situation as the predictable result of an internal market where compute is rationed by managerial seniority rather than unit-cost economics. Google has not disputed the framing, pointing instead to its broader infrastructure spend and to compute scarcity as an industry-wide condition. Every major model provider is constrained, but Google's situation is unique: it has become its main competitors' largest infrastructure supplier while its own researchers wait in line. The departures suggest that, for top talent, that trade-off has become hard to swallow.