Google unveils two new AI chips, TPU 8tx and TPU 8i Max

Google on Monday unveiled two new systems in its eighth-generation TPU chip family: TPU 8tx, built to run AI models on confidential data, and TPU 8i Max, aimed at very long contexts. That brings the generation to four distinct systems in a single year.

Nothing changes today for people who use AI in Mexico. The chips sit in data centers, and Google rents them out through Google Cloud. Its blog post gives no launch date, regions or price. Any effect would be felt by companies that want to run models without moving sensitive data out of their own environment, such as banks or health providers.

TPU 8tx is designed for inference, meaning the use of an already-trained model, with protection built into the hardware. It reuses the silicon of TPU 8t, pairs with fifth-generation AMD EPYC host processors and also works for small training runs. Amin Vahdat, Google's chief technologist for AI and infrastructure, says it "builds security directly into the hardware layer." Thanks to its Boardfly network, Google promises up to 2x lower latency than TPU 8t.

TPU 8i Max pairs two TPU 8i chips in a single package. Compared with TPU 8i, it doubles compute, SRAM and HBM memory bandwidth in the same rack space, Google says. It targets large context windows, the amount of text a model can hold in view at once, which AI agents need for long tasks. Vahdat says it "scales compute and memory in lockstep."

Every performance figure is Google's own, and there are no independent tests of the two new systems.

The company counts nearly 100 TPU customers, from AI labs and trading firms to retailers and universities. Alongside the announcement, it open-sourced TorchTPU, XLA Serving Engine and TPU Sync, and put its agent suite for TPUs in preview.