Google adds TPU 8tx and 8i Max for confidential inference

Google said on 10 November 2026 that it is expanding its eighth-generation TPU line with two new systems: TPU 8tx, aimed at secure and confidential inference, and TPU 8i Max, aimed at very large context windows. It says this is the first time it has introduced four TPU systems in a single year, and that TPUs now have nearly 100 customers.

TPU 8tx: data protected while in use

Google describes TPU 8tx as built for regulated enterprises and governments running frontier or proprietary models and autonomous agents next to sensitive data. Data is normally encrypted in storage and on the network but unlocked in memory during processing. Google says TPU 8tx builds protection into the hardware so data and model weights stay isolated while in active use.

It uses the same silicon as TPU 8t, paired with 5th Generation AMD EPYC CPU hosts, and the low-latency Boardfly network from TPU 8i. Google says that gives up to 2x lower latency than TPU 8t with a near-zero performance penalty. It also says teams can fine-tune, distill and run reinforcement learning on private data on one system, without moving that data between clusters.

TPU 8i Max: more memory per rack

For agents working across a whole codebase or hours of video, Google says memory bandwidth is the bottleneck. TPU 8i Max combines two TPU 8i chips in one package. Google says that doubles compute, SRAM and HBM bandwidth against TPU 8i, and because it fits twice the silicon in the existing rack footprint, a data centre can deploy more compute per square foot and per watt.

For context, Google says the earlier TPU 8t and 8i, introduced at Google Cloud Next in April, offer up to 2.7 times better training performance and 80% better inference price performance than the previous generation.

Software changes in the same post

Google also listed software updates. It says it open-sourced TorchTPU last month so PyTorch code and Hugging Face models run natively on TPUs, with backends for vLLM and SGLang. A new open-source TPU agent suite, in preview, plugs into Claude Code, Cursor and Google Antigravity to help with onboarding and kernel tuning.

Google says its Agent Compute Engine software suspends and resumes idle agent sessions to cut infrastructure costs, and that GKE dynamic slicing, now generally available, cuts job startup times by up to 5x and speeds fault recovery by up to 4.5x. It also says its vLLM-TorchTPU stack reached state-of-the-art price-performance on the InferenceX and AgentX benchmarks. These are Google-reported results; the post does not name an independent evaluator.

What the post leaves out

There are no prices, regions, availability dates or sizes of cloud offering for TPU 8tx or 8i Max. Google also does not say which attestation or certification evidence customers will get for the confidential-computing claim. A security or compliance team should treat the claim as a vendor statement until Google Cloud documentation says what is attested and who verifies it.