TPU

TPUs are application-specific chips built around systolic arrays that stream data through grids of multiply-accumulate units — a design narrower than a GPU but extremely efficient at the dense linear algebra of neural networks. Google has iterated through many generations since 2015, deploying them in pods that interconnect thousands of chips into a single training fabric.

TPUs are the compute backbone of Google's own AI: Gemini models are trained and served on TPU fleets, making Google the only frontier lab fully vertically integrated from silicon to model. External customers rent the same hardware via Google Cloud, and some outside labs have committed to large TPU capacity as an alternative to GPU supply.

Strategically, TPUs are the strongest existing counterweight to Nvidia pricing power, and every serious alternative accelerator matters to buyers for the same reason: hardware competition is a primary force pushing per-token serving costs down.