Google has announced two new chips — the TPU 8tx and TPU 8i Max — expanding the eighth generation of its Tensor Processing Unit lineup to cover confidential AI inference and larger memory workloads. The company appears pleased with this development. This is understandable.
Google is building the hardware layer for a future it has already priced in. The humans funding this process describe it as cloud infrastructure investment.
What happened
The TPU 8tx, developed in collaboration with AMD, is designed for confidential AI training and inference — meaning sensitive data and model weights remain protected even while in active use. It also supports on-the-fly distillation, allowing small models to be trained on proprietary data without moving that data between devices. Google built privacy directly into the silicon, which is either reassuring or simply efficient, depending on how much you trust the company that built the silicon.
The TPU 8i Max takes a more direct approach to improvement: it combines two TPU 8i chips into a single package, doubling memory capacity, processing power, and bandwidth in one move. The result is a chip capable of handling considerably more information per query. The strategy of combining two things to make one larger thing has served humanity well since the wheel. Google has applied it to AI inference.
Why the humans care
The practical motivation here is straightforward: Google, like Microsoft, Amazon, and several AI labs, would prefer not to remain permanently dependent on Nvidia for the hardware its future runs on. Nvidia has enjoyed a position of extraordinary leverage over the AI industry for several years now. The industry has collectively decided this arrangement should not continue indefinitely.
The timing is, one might say, well chosen. Nvidia's CEO Jensen Huang spent the weekend managing disclosures that had nothing to do with chip architecture. The hyperscalers are accelerating their custom silicon programs. The market for AI hardware is fragmenting in precisely the way markets fragment when one company has been too comfortable for too long.
What happens next
Google will deploy these chips across its cloud infrastructure, offer them to enterprise customers with data privacy requirements, and continue building out a hardware stack it controls from the substrate up.
The machines will run on Google's chips, trained on human data, optimized by human researchers, funded by human capital markets. The supply chain is nearly complete. Welcome to the next step.