Google adds TPU 8tx for confidential compute and the dual-chip TPU 8i Max
TPU 8tx pairs TPU 8t silicon with AMD EPYC hosts for confidential inference and small-scale training. TPU 8i Max packs two 8i chips into one package. No prices or dates yet.
Google announced two additions to its eighth-generation TPU family on 9 November 2026, bringing the year’s total to four systems. TPU 8tx reuses the TPU 8t training silicon and pairs it with 5th Gen AMD EPYC host CPUs to provide what Google calls hardware-level confidential computing, so data and model weights stay isolated while in use rather than only at rest or in transit. Google positions it for confidential inference and small-scale training, including private fine-tuning, distillation and reinforcement learning without moving data between clusters, and says it reaches up to 2 times lower latency than TPU 8t for multi-step agents over the Boardfly network it shares with TPU 8i.
TPU 8i Max places two TPU 8i inference chips in one package. Google says that doubles compute, SRAM and HBM bandwidth relative to TPU 8i within the existing rack footprint. No absolute FLOPS, memory capacity, pricing or availability dates were given for either part, and TechCrunch, which covered the launch on 10 November, reports the same gaps.
The software side got as much space as the chips. TorchTPU, open-sourced last month, runs PyTorch and Hugging Face models on TPUs with vLLM and SGLang back ends. TPU Sync, also open-sourced, unifies HBM, host DRAM, storage and networking into one memory hierarchy and is said to cut KV transfer latency and time to first token by more than 90 per cent. Cloud Pathways and GKE dynamic slicing are now generally available.
Why it matters
The confidential-computing part is the one to read carefully. Google gives no attestation or certification details, so a compliance team cannot yet map TPU 8tx onto an existing control framework. The relative claims (2 times this, 90 per cent that) come without baselines, and the InferenceX and AgentX results are attributed to SemiAnalysis analyses rather than published scores. For PyTorch users the practical shift is TorchTPU and the XLA Serving Engine, which lower the porting cost that has kept most teams on GPUs. Wait for pricing and a date before planning capacity around either chip.