Google announced TPU 8tx on November 9, 2026, as a system for confidential inference and small-scale training. Google says it uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC host CPUs and the Boardfly network. That makes 8tx a system configuration around existing TPU silicon, rather than a newly designed TPU chip. Google’s announcement describes its intended workloads and configuration.

The useful distinction: TPU 8tx is a system configuration built around TPU 8t silicon, with a different stated role and security-focused setup.

How TPU 8tx differs from TPU 8t

Google positions TPU 8tx for confidential inference and small-scale training, while TPU 8t is intended for large-scale training. The two systems share silicon, according to Google; the stated difference lies in the system configuration and its intended work.

SystemGoogle-stated roleSilicon
TPU 8txConfidential inference and small-scale trainingSame silicon as TPU 8t
TPU 8tLarge-scale trainingTPU 8t silicon

What the AMD EPYC hosts do

Google says TPU 8tx pairs its accelerator with fifth-generation AMD EPYC host CPUs and uses the Boardfly network associated with TPU 8i. In ChosunBiz’s account, the EPYC CPUs prepare data, distribute workloads, communicate with the TPU, supply it with data, and manage compute jobs; the TPU carries out the AI computation. ChosunBiz reports that the system uses AMD Secure Encrypted Virtualization (SEV) for its confidential-computing implementation.

Google describes the security goal as protecting and isolating sensitive data and model weights while they are in use. ChosunBiz’s report names prompts, model weights, intermediate computation data, information held in CPU memory, and data moving between the CPU and TPU as protected. It also reports that the accelerator checks its identity and security configuration before accessing protected memory.

Workloads and performance claims

Google says TPU 8tx is intended to support confidential inference and small-scale training, including fine-tuning with private data, model distillation, and reinforcement learning for agents. Google’s description says these workloads can keep sensitive data within a system rather than moving it between separate clusters.

Google also claims that TPU 8tx offers up to 2× lower latency than TPU 8t and describes the performance penalty from its security features as “near-zero.” The latency comparison is against TPU 8t. Google’s announcement presents those figures as system-performance claims.