Google announces two TPU configurations, 8tx and 8i Max

Google announced on 9 November 2026 two configurations of its TPU systems, processors designed for artificial intelligence. The 8tx is designed for confidential processing; the 8i Max targets, among other things, applications that use long context windows, meaning large volumes of information supplied to the model to generate a response.

Both systems reuse existing components. In April, Google introduced the TPU 8t for training large models and the TPU 8i for low-latency reasoning. The 8tx uses the 8t’s silicon, while the 8i Max combines two 8i chips in a single package.

Reuters reports that Google is gradually making its TPUs available to its cloud customers, after using them mainly for its own services. The company is thus offering its accelerators to business customers in a market dominated by Nvidia.

The 8tx combines Google’s chips with fifth-generation AMD EPYC host processors and the Boardfly network. Google says its hardware protection isolates sensitive data and model weights while they are in use. It is designed to protect information during processing, complementing encryption for data stored or moving across the network.

Google says Boardfly delivers up to half the latency of the 8t and cites an almost zero performance penalty in that comparison. The company does not clearly specify whether that penalty refers to the overhead of the hardware protection.

For the 8i Max, Google claims twice the compute capacity and SRAM memory, as well as double the HBM bandwidth, compared with the 8i. Memory holds information used by the model; its bandwidth determines how quickly that information reaches the model. Higher throughput can help process long contexts, without guaranteeing the quality of responses or a given level of real-world performance.