On November 9, 2026, Google announced two additions to its eighth-generation TPU lineup on its official blog: TPU 8tx for confidential inference and TPU 8i Max for ultra-long context. Together with the 8t and 8i released in April, Google has now introduced four TPU systems within a single year — a first, according to the company. Notably, neither new product is a brand-new chip; both repackage existing silicon with new packaging and networking, and Google gave no pricing or launch date.
8tx: Compute Without Decrypting the Data
TPU 8tx is built on the same silicon as the 8t and paired with fifth-generation AMD EPYC processors, delivering confidential computing at the hardware layer: data and model weights stay isolated and encrypted while being processed, not only when stored or transmitted. Google says it suits regulated enterprises and government agencies running frontier models and autonomous agents alongside sensitive data. Using the 8i's low-latency Boardfly network, the 8tx cuts latency by up to half compared with the 8t, with a near-zero performance penalty; the same system can fine-tune models on private data, distill large models, and run reinforcement learning without moving sensitive data between clusters.
8i Max: Two Chips in One Package
TPU 8i Max targets a different wall: the memory wall. When dozens of agents work across an entire codebase or hours of video at once, memory bandwidth becomes the bottleneck first. Google combined two low-latency 8i chips into a single package, doubling compute, on-chip SRAM, and HBM bandwidth while keeping the same rack footprint, raising compute per square foot and per watt. It is aimed at million-token context windows and the highest compute density per rack.
The Lineup Is Complete, but Pricing and Benchmarks Are Still Missing
The four systems now cover distinct roles: the 8t for large-scale training, the 8i for low-latency inference, the 8tx for confidential computing, and the 8i Max for ultra-long context. Google is also lowering adoption barriers: TorchTPU, open-sourced last month, lets PyTorch code run on TPUs without rewrites, and the Pathways runtime that trains Gemini is now generally available on GKE. Google is also pushing TPUs to outside customers faster — TPUs have landed on Oracle Cloud, and OpenAI is reportedly planning to buy them. But the announcement gives no pricing or full specifications, the performance figures are Google's own claims with no independent benchmarks yet, and Bloomberg reports both systems will only ship in 2027. What will decide their real standing is next year's pricing, delivery pace, and independent test results.
