What changed

Google announced two eighth-generation TPU systems rather than new base chips. TPU 8tx pairs TPU 8t silicon with 5th Gen AMD EPYC hosts and Boardfly networking for confidential inference; TPU 8i Max combines two TPU 8i chips in one package to double compute, SRAM and HBM bandwidth.

Why it matters

The configurations target two production constraints that generic accelerator comparisons often hide: protecting active model data and fitting very large context workloads within rack and memory limits. They also show system packaging, hosts and networks becoming distinct procurement variables around the same accelerator silicon.

What to do

Ask Google Cloud for regional availability, attestation boundaries, pricing and an exact 8tx/8i Max topology; then benchmark one private fine-tune or long-context workload against the existing TPU configuration with failure and data-residency tests.

Caveat

Google's latency, density and near-zero-overhead claims are vendor results, and the announcement gives no public price, region list or independent confidential-computing evaluation. Treat both systems as announced until the required capacity is confirmed for the target account.