Google Adds TPU 8tx and TPU 8i Max to Its Eighth-Gen AI Chips

TPU 8tx: confidential inference

TPU 8tx is designed to run inference in Confidential Computing environments. It protects sensitive data and model weights while in use, in addition to protecting them at rest and in transit. Google is targeting government bodies and companies in tightly regulated sectors that need to run private models or applications built on sensitive information.

The system uses the same chips as TPU 8t and pairs them with fifth-generation AMD EPYC server processors, which coordinate tasks, prepare data and manage communication with the accelerators. According to AMD, the design combines Secure Encrypted Virtualization (SEV) with mechanisms that verify the identity and security settings of the accelerators.

Google says its low-latency Boardfly network cuts latency by up to half compared with TPU 8t, with limited performance impact, based on the company's own measurements. It adds that models can be run and customized on a single architecture, including fine-tuning and reinforcement learning.

TPU 8i Max: long context

TPU 8i Max targets AI workloads that need long context and large memory, such as autonomous agents working with huge code repositories or long video content. It combines two TPU 8i units in one package. Google says this doubles compute, SRAM and HBM memory bandwidth compared with the base TPU 8i, increasing compute density inside data center racks.

A four-system family

Google introduced TPU 8t, for large-scale training, and TPU 8i, for low-latency inference, in April 2026. With the two additions, the family now has four systems covering training, fast inference, confidential computing and long-context processing. Google is also easing development through open-source frameworks such as PyTorch and the TorchTPU project.

The announcement gave no pricing, commercial availability dates or supported regions for the new systems.