Google announced TPU 8tx for confidential inference and some small-scale training, while TPU 8i Max combines two TPU 8i chips for long-context workloads. Announced on November 9, 2026, both are system configurations built around existing TPU silicon, rather than entirely new chip designs, Google says.
How the two TPU systems differ
| System | Configuration | Workload Google identifies |
| TPU 8tx | TPU 8t silicon, 5th Generation AMD EPYC CPU hosts and the Boardfly network | Confidential inference; also small-scale training |
| TPU 8i Max | Two TPU 8i chips in one package | Large-context workloads and high compute density |
Google says TPU 8tx protects sensitive data and model weights while they are in active use. It also describes the system for training at smaller scale, including work with private data. The system reuses TPU 8t silicon, pairing it with AMD host processors and Boardfly networking.
TPU 8i Max is aimed at workloads that need to work with very large amounts of context, such as an agent handling a software codebase or hours of video. Google describes the platform as supporting context at million-token scale. Its package combines two TPU 8i chips.
What Google's performance figures describe
Google claims TPU 8tx delivers up to 2x lower latency than TPU 8t, with a near-zero performance penalty. For TPU 8i Max, Google claims double the compute, SRAM and HBM bandwidth of TPU 8i, and twice the silicon within the existing rack footprint. SRAM is on-chip memory; HBM is high-bandwidth memory.
Those claims measure different things: latency for TPU 8tx, and hardware resources for TPU 8i Max. The 2x resource figures describe compute and memory bandwidth, while application performance depends on how a workload uses them, including its model and context size.
Match the configuration to the bottleneck: confidential inference points to TPU 8tx; very long context points to TPU 8i Max.
The software around the hardware
Google also announced an open-source TPU agent suite in preview for onboarding, diagnosing bottlenecks and tuning kernels with coding tools including Claude Code, Cursor and Google Antigravity. Its XLA Serving Engine is an open-source route to train in Python and serve through a low-latency C++ runtime across TPUs, GPUs and CPUs. Google says Agent Compute Engine suspends and resumes idle agent sessions.
Google says its overall TPU business has nearly 100 customers across sectors including frontier AI labs, trading firms, retailers and universities.
