With Nvidia GPUs in short supply and AI labs hoarding compute behind walled gardens, Google continues to offer its custom TPU chips to outside teams — including Anthropic’s research group.
https://www.youtube.com/watch?v=RgV57kDzcng
Sundar Pichai, when asked why Google doesn’t lock down its compute for a competitive edge, laid out the company’s logic and compute ecology for the first time. Google Cloud operates financially and strategically independent from Google DeepMind. In Pichai’s view, chip supply isn’t a zero-sum game: as long as the foundry can produce enough chips, training Google’s own models and renting out compute can coexist without conflict.
The deeper rationale: pushing cutting-edge models to their limits provides irreplaceable strategic value for evolving Google’s chip ecosystem. “Having the best research institutions use TPUs is exactly how Google can design and manufacture the next generation of the world’s most powerful AI hardware,” Pichai said.
The history of leading-edge chip development shows that top-tier custom chips can’t be designed in a vacuum — they require physical testing and performance feedback from frontier models under extreme load. When Anthropic runs massive models, the network topology, memory throughput, and interconnect bottlenecks it encounters directly guide the tape-out design and compiler optimizations for the next TPU architecture.
The compute arms race reveals that becoming a supplier of chips and underlying ecosystems offers far more strategic depth than simply locking chips in your own server room.