Google adds TPU 8tx for confidential AI inference

Google now has a chip for the AI work you can't send to an ordinary cloud.

Google announced two more eighth-generation TPU systems on November 9: TPU 8tx, built for confidential inference, and TPU 8i Max, built for very long contexts. Neither has a price or a ship date.

Two new TPUs make four this year

In a post by Amin Vahdat, Google's SVP and Chief Technologist for AI and Infrastructure, Google says the additions bring it to four TPU systems in a single year, a first for the company. TPU 8t, for training, and TPU 8i, for inference, were announced in April at Google Cloud Next.

TPU 8tx uses the same silicon as TPU 8t, paired with 5th Generation AMD EPYC host CPUs. Google says it provides hardware-level confidential computing that keeps data and model weights isolated while in use. Over its Boardfly network, Google says, it has up to 2x lower latency than TPU 8t.

Google says TPU 8tx also handles private fine-tuning, distillation and reinforcement learning training on one system, so sensitive data doesn't move between clusters.

TPU 8i Max puts two TPU 8i chips on one package. Google says that doubles the compute, SRAM and high-bandwidth memory bandwidth of TPU 8i and fits twice the silicon in the same rack footprint.

Google gave no pricing and no availability dates for either system.

Confidential inference is a pitch to the buyers who said no

Google aims TPU 8tx at regulated enterprises and governments running sensitive data alongside frontier or proprietary models and agents.

That is the buyer who has held back because the data can't leave a controlled environment. A chip that keeps weights and data sealed while they run answers that objection directly, if it works as described.

TPU 8i Max is for long context work, which Google describes as many agents working across a codebase, or hours of video.

Google says TPUs have nearly 100 customers, including frontier labs, trading firms, retailers and universities. It names none. Google's chips have been finding outside buyers this fall: Oracle said in October it will rent Google TPUs.

The software is what you can use now

The chips have no date. Google's post lists software alongside them, and some of it is open or generally available:

• TorchTPU is open source, with native PyTorch and Hugging Face model support on TPUs.
• Google open-sourced its XLA Serving Engine and TPU Sync. It says TPU Sync cuts KV cache transfer latency and time to first token by more than 90%.
• Cloud Pathways and GKE dynamic slicing are generally available. Google says dynamic slicing starts jobs up to 5x faster.
• A TPU agent suite, in preview, works with Claude Code, Cursor and Google Antigravity.

If you have wanted to try TPUs without rewriting your PyTorch code, TorchTPU is where to start.

Watch for a price and a test

Google says TPU 8tx runs confidential workloads with a "near-zero performance penalty." That is Google's claim, and it needs an outside test.

A price will tell you more than a spec sheet. Ask your Google Cloud rep for both.