Google Ships a TPU for Buyers Who Cannot Hand Over Data

TPU 8tx pairs existing silicon with AMD EPYC hosts and confidential computing, while 8i Max fuses two chips per package.

Google added two systems to its eighth-generation TPU line on 9 November 2026. TPU 8tx reuses the 8t silicon but pairs it with custom fifth-generation AMD EPYC Turin host processors and confidential computing, keeping data encrypted while in use inside a trusted execution environment. TPU 8i Max places two 8i chips in a single package to double memory bandwidth for long-context inference. Google also confirmed a seven-hop network diameter for the inference fabric against sixteen for the training torus. The commercial read is that confidentiality is being sold as a specification rather than a feature: it shifts the buying criterion from throughput per dollar to whether the workload is legally permitted to run on that hardware at all. That is a question benchmarks cannot answer.