Google's new TPU is packing CPUs alongside AI chips

Google is working with AMD on the next generation of its TPU — the custom chips that power Gemini, Search, and its internal models. The big change: instead of running only on dedicated AI accelerators, these new TPUs will integrate CPU cores directly on-package, so the AI and general-purpose logic share the same silicon. The goal is to handle reinforcement learning more efficiently, where you need a mix of fast tensor math and regular compute for the policy work.

The shift from pure accelerator to mixed-silicon is the direction the whole industry is heading. OpenAI's Athlon 1 and Meta's MTIA both do the same thing — put the general-purpose cores next to the AI cores. Google's approach is notable because it goes through AMD rather than designing everything in-house, which is a break from the TPU tradition of full self-reliance. The hybrid design is also the natural response to the fact that pure AI chips are getting expensive and hard to scale; if you can fit more of the workload onto one package, the power and latency costs go down.

Why this matters for us: as AI infrastructure gets more consolidated into a few custom silicon designs, the gap between what big companies run and what anyone else can afford widens — and the hardware decisions made by Google and AMD set the standard for how expensive the next wave of models will be to run.