Google leases TPUs to Anthropic for mutual benefit

ME AI message, according to monitoring by Beating, amid the scarcity of NVIDIA GPUs and the frantic hoarding of computing power by major labs building high walls, Google continues to provide its proprietary TPU chips to external parties, including Anthropic’s research team. When asked why Google does not restrict access to computing power to secure absolute competitive advantage, Pichai for the first time articulated Google’s business logic and computing ecology. Google Cloud’s commercial operations are financially and strategically independent from Google’s internal model development team (Google DeepMind). In Pichai’s view, chip supply is not a zero-sum game: as long as foundries can produce sufficient quantities of chips, both internal model training and external leasing of computing power can be satisfied simultaneously without conflict. A deeper consideration is that the extreme usage by frontier teams provides irreplaceable strategic value for Google’s iterative development of its proprietary chip hardware and software ecosystem. Pichai stated directly: “Allowing top-tier research institutions to use TPU is the key driver enabling Google to design and manufacture the next generation of the world’s most powerful AI hardware.” The evolution of cutting-edge chips demonstrates that top-tier proprietary chips cannot be developed in isolation within a lab—they must undergo physical testing and performance feedback under extreme conditions from frontier models. Feedback from Anthropic’s operation of ultra-large-scale models—such as network topology, memory throughput, and interconnect bottlenecks—directly informs the tape-out design and compilation optimization of the next-generation TPU architecture. The current state of the computing power arms race shows that becoming a supplier of chips and underlying ecosystems offers a far broader strategic horizon than simply locking chips inside one’s own data centers. (Source: BlockBeats)