Compute Shortages Drive AI Researchers from Google to Startups

A shortage of computing power at Google is driving veteran AI researchers to leave the company and establish their own startups. Former employees report that Google's internal allocation of tensor-processing units (TPUs) prioritizes high-revenue projects and the development of the Gemini model over experimental or academic research. This system has created a competitive internal market where researchers must align their work with corporate priorities to secure necessary resources.

Sundar Pichai, CEO of Alphabet Inc., acknowledged that the company is "compute constrained in the near term," though he maintains that resources are allocated to the most important priorities. The environment has led to the departure of key talent from Google and DeepMind, including researchers who felt unable to secure compute for visual reasoning or reinforcement learning projects.

Departing researchers have launched several new ventures, including Elorian, ReflectionAI, and Ricursive Intelligence. These founders cite the need for greater autonomy and more reliable access to computing power from diverse external sources as the primary drivers for their exits. One former DeepMind researcher rejected a compute-based incentive to stay, prioritizing the independence found in the startup ecosystem over Google's internal resource constraints.