Google announced TPU 8tx and TPU 8i Max on November 9, 2026, positioning them for different AI infrastructure needs: confidential inference and small-scale training for 8tx, and large-context workloads for 8i Max. The announcement expanded the eighth-generation lineup that began with TPU 8t and TPU 8i, introduced at Cloud Next in April 2026. Google’s announcement also covers software and developer tools built around its TPU platform.
Two configurations, two infrastructure priorities
Google describes TPU 8tx as a confidential-computing system built on the same silicon as TPU 8t. It pairs that accelerator with 5th Generation AMD EPYC CPU hosts and the Boardfly network, which Google describes as low-latency. The company says 8tx is designed to protect and isolate sensitive data and model weights while they are in active use. It identifies fine-tuning on private data, distillation and reinforcement-learning agent training as workloads that can run on the system.
Google also claims up to 2× lower latency than TPU 8t, with a near-zero performance penalty. That is a company performance claim. For readers weighing the two configurations, 8tx’s distinguishing aim is the confidential-computing requirement, not a different generation of TPU silicon.
TPU 8i Max takes a different route: Google combines two TPU 8i chips in one package. The company says the configuration doubles compute, on-chip SRAM and HBM bandwidth relative to one TPU 8i, and fits twice the silicon within the existing rack footprint. HBM, or high-bandwidth memory, supplies accelerator memory bandwidth; SRAM is on-chip memory. Google positions 8i Max for large-context workloads and higher compute density per rack.
| Configuration | Design described by Google | Workloads Google identifies |
| TPU 8tx | TPU 8t silicon, 5th Generation AMD EPYC CPU hosts and Boardfly networking | Confidential inference and small-scale training |
| TPU 8i Max | Two TPU 8i chips in one package; twice the compute, SRAM and HBM bandwidth of one TPU 8i | Large-context workloads and higher compute density per rack |
Match the configuration to the constraint: confidential computing points to 8tx, while memory bandwidth and rack density point to 8i Max.
Software tools extend the announcement beyond hardware
Google describes TorchTPU as an open-source route for running PyTorch code and Hugging Face models natively on TPUs. The company also lists out-of-the-box TPU backends for vLLM and SGLang, plus a TPU agent suite in preview for onboarding, diagnostics and kernel tuning. Google says that suite integrates with Claude Code, Cursor and Google Antigravity.
The other tools target serving and system operations. Google describes XLA Serving Engine as a low-latency serving option across TPUs, GPUs and CPUs, and Agent Compute Engine as a way to suspend and resume idle agent sessions. TPU Sync manages model-state movement among HBM, host DRAM, shared storage and the data-center network. Google claims TPU Sync can reduce KV-transfer latency and time to first token by more than 90%.
Google also describes Cloud Pathways and GKE dynamic slicing as generally available. These announcements span the software stack, from model frameworks to serving and cluster management.