Google AI Researchers Compete for Computing Power

Google's own AI researchers are struggling to access the company's powerful computing resources, leading to competition with revenue-generating customers and flagship projects like Gemini [1, 2]. This scarcity of Tensor Processing Units (TPUs) is influencing research directions, promotion decisions, and the pace of work within the merged Google DeepMind lab [1].

The intense demand for computing power means that even internal AI researchers must contend for access, sometimes feeling they are outranked by paying clients [1, 2]. Andrew Dai, a former Google AI researcher, left the company last summer after concluding he could not secure enough computing power to pursue his idea for AI that could better understand images, a project he felt was crucial but unlikely to get priority within Google [1, 2]. He is now a founder of Elorian, an AI startup focused on visual reasoning [1].

This crunch for resources is pushing some top researchers to leave Google and found their own startups, citing better access to compute, fewer bureaucratic obstacles, and more freedom to explore their own AI frontiers [1]. Former DeepMind researcher Ioannis Antonoglou, who left in 2024 to co-found ReflectionAI, believed reinforcement learning was the next frontier but found it unclear if Google would prioritize that path [1]. Another former researcher, Anna Goldie, was offered more computing power to stay but ultimately left to found Ricursive Intelligence, stating she now has the freedom to make decisions with her co-founder without seeking permission from multiple layers of management [1].

Alphabet CEO Sundar Pichai acknowledged that the company is "compute constrained in the near term," with Google Cloud's backlog nearly doubling to over $460 billion [1, 2]. Google stated it has a rigorous process for allocating resources to balance customer needs with long-term investments in research and innovation [1, 2]. However, the allure of compute power and the prestige of contributing to large training runs, like those for coding models, acts as both a carrot and a stick for researchers [1].

The competition for TPUs is shaping which AI projects are pursued and how quickly they advance, potentially prioritizing short-term gains over riskier, experimental ideas [1].