Google adds two more TPUs: one for confidential AI, one for huge context windows
Quick answer: On Nov 9, 2026, Google announced two new systems in its eighth-generation TPU line: TPU 8tx, built for secure, confidential inference, and TPU 8i Max, built for very large context windows and higher compute density per rack. Google says it is the first time it has introduced four TPU systems in a single year. (Google blog)
What’s new
- TPU 8tx: aimed at regulated enterprises and governments running frontier or proprietary models and AI agents next to sensitive data. Google says it builds security into the hardware layer so data and model weights stay protected and isolated while in use, not just in storage or on the network. It is paired with 5th Gen AMD EPYC CPU hosts, and Google says it delivers up to 2x lower latency than TPU 8t with a “near-zero performance penalty.” It is also pitched for small-scale training, such as fine-tuning on private data.
- TPU 8i Max: two TPU 8i chips combined on one package. Google says that gives double the compute, double the SRAM and double the HBM bandwidth of TPU 8i, in the same rack footprint, aimed at the “memory wall” when many agents work across a whole codebase or hours of video.
- They join two chips announced in April: TPU 8t (training) and TPU 8i (real-time reasoning), which Google said offer up to 2.7x better training performance and 80% better inference price performance than the previous generation.
All performance figures above are Google’s own claims; the post doesn’t include independent benchmarks.
The software side
Google also says it is making TPUs easier to use: PyTorch and Hugging Face models run natively through the open-source TorchTPU, with backends for vLLM and SGLang; a TPU agent suite (preview) plugs into coding tools including Claude Code, Cursor and Google Antigravity; and Cloud Pathways and GKE dynamic slicing are now generally available. Google says TPUs have nearly 100 customers. (Google blog)
Why it matters
Most readers will never buy a TPU, so this won’t change your laptop. It matters because it shows where AI hardware is heading: chips tuned to one job (training, reasoning, secure inference, long context) instead of one chip for everything, and a privacy angle: running AI on sensitive data without the data sitting readable in memory. Whether TPU 8tx’s protections hold up in practice will need independent testing, and the post gives no availability date or pricing.
Sources
- Google, “Introducing TPU 8tx and TPU 8i Max” (Amin Vahdat, Nov 9, 2026): https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
— AltTab Dad