Google adds two more eighth-generation TPU systems, aimed at confidential AI and long contexts

Google on November 9, 2026, announced two new systems built on its eighth-generation tensor processing units (TPUs), the custom AI chips it uses for its own Gemini models and rents to Google Cloud customers. TPU 8tx is meant for confidential AI work, and TPU 8i Max for models that must hold very long inputs in memory. They join TPU 8t and TPU 8i, which Google introduced on April 22. Google says it is the first time it has introduced four TPU systems in a single year.

Google draws a line between a new chip and a new system. A chip, it says, takes years to design and manufacture, while a system is the way compute, networking, storage and packaging are put together around it, which can change in months. Both new products reuse existing silicon: TPU 8tx is built on the same silicon as TPU 8t, and TPU 8i Max combines two TPU 8i chips.

Four systems, four jobs

TPU 8tx targets what the industry calls confidential computing. In most systems, as Google notes, encryption protects data on disk and on the network, but not in memory while a processor is working on it. Google says TPU 8tx, paired with fifth-generation AMD EPYC host processors, extends that protection to data and model weights while they are being processed. It also claims up to 2x lower latency than TPU 8t, “with a near-zero performance penalty.” It is pitching the system to regulated businesses and governments that want to fine-tune or run models on sensitive data without moving that data between separate clusters.

TPU 8i Max is aimed at a different limit. Google says memory bandwidth runs short when dozens of agents tackle a whole codebase or a long video together, and the new package is its response. Because the dual-chip package fits twice the silicon in the existing rack footprint, Google says data centers can install more computing power per square foot and per watt.

What the posts leave out

Google's announcement, written by Amin Vahdat, its senior vice president and chief technologist for AI and infrastructure, and repeated on the Google Cloud blog, gives no prices, regions or dates for when customers can use TPU 8tx or TPU 8i Max. In April, Google said TPU 8t and TPU 8i would be generally available later this year; the November posts do not say whether that has happened.

The performance figures are Google's own. Against the previous generation, which its April post identified as Ironwood, Google now cites up to 2.7 times better training performance for TPU 8t and 80% better inference price-performance for TPU 8i. Google also points to results on outside benchmarks called InferenceX and AgentX, which we have not reviewed.

The company makes business claims too. It says TPUs are now a fast-growing business with nearly 100 customers, among them frontier AI labs, trading firms, retailers and universities, and that customers can now install TPUs in their own data centers rather than only renting them in Google's cloud. Google does not give revenue figures for the business in either post.

Software around the chips

Google paired the hardware with software that it says will make TPUs easier for any developer to use. It says it open-sourced TorchTPU in October, so code written for the PyTorch framework runs on TPUs without rewrites, and it released a TPU agent suite in preview that works with AI coding tools. It is also open-sourcing TPU Sync, a memory layer that it says cuts the delay before a model's first output by more than 90%, and has made two cluster tools, Cloud Pathways and GKE dynamic slicing, generally available.

Google says it now uses its own AI to design TPUs, from Google DeepMind's AlphaChip system to Gemini, which it says produces chip layouts in hours rather than weeks.