Google's four TPUs in one year: what 8tx and 8i Max mean for anyone negotiating with Nvidia

Google announced two more eighth-generation TPU systems on November 9, bringing to four the number it has introduced this year, a cadence it says it has never managed before. TPU 8t for large-scale training and TPU 8i for real-time inference arrived at Cloud Next in April; TPU 8tx and TPU 8i Max complete the set.

The two new systems

TPU 8tx uses the same silicon as 8t but pairs it with fifth-generation AMD EPYC host CPUs to deliver hardware-level confidential computing, keeping both data and model weights isolated while in use. Google positions it for secure inference and for small-scale training, fine-tuning, distillation and reinforcement learning on sensitive data that cannot leave a customer's trust boundary. It borrows the Boardfly interconnect from 8i and Google claims up to half the latency of 8t with a negligible performance penalty.

TPU 8i Max packages two 8i chips together, doubling compute, on-chip SRAM and HBM bandwidth per package while fitting into the existing rack footprint. It is aimed at very long context windows and at squeezing more compute out of each square foot and each watt, the two constraints that now bind data-centre operators more than chip supply does. Google gave no pricing or availability dates for either system.

The software is the story

The hardware announcements came wrapped in a software release that matters more for adoption. TorchTPU, open-sourced in October, lets PyTorch code and Hugging Face models run natively on TPUs, with backends for the vLLM and SGLang serving engines that most inference teams already use. An open-source XLA Serving Engine runs models across TPUs, GPUs and CPUs from one C++ runtime. TPU Sync tiers key-value caches across chip memory, host DRAM and storage, and Google claims it cuts cache transfer latency and time-to-first-token by more than 90%. Cloud Pathways and GKE dynamic slicing, which assemble and partition TPU topologies through optical circuit switching, are now generally available. There is even a preview TPU agent suite that plugs into Claude Code, Cursor and Antigravity to diagnose bottlenecks and autotune kernels.

For years the honest objection to TPUs was that the software assumed JAX and a Google-shaped workflow. That objection is now largely gone, which is why the claim of nearly 100 TPU customers, including frontier labs and trading firms, is more credible than it would have been in 2024.

Why Nvidia buyers should care

Nvidia's pricing power rests on CUDA lock-in and on the absence of a second supplier that can deliver training-class systems at scale. Four TPU systems in a year, with a PyTorch-native stack and confidential compute, is the most direct challenge to both. Buyers rarely need to move workloads to benefit: a credible alternative in the room changes the terms of a GPU reservation or a multi-year cloud commitment. It also arrives in a week when Nvidia's own governance is under an independent board review, which sharpens any procurement committee's interest in a second source.

The counter-argument

TPUs are only available through Google Cloud, so the choice is also a choice of cloud, and that lock-in is arguably tighter than CUDA's. Google's performance claims are its own and the two new systems have no public pricing. Most enterprise inference still runs on models whose vendors optimise first for Nvidia, and Anthropic's new Fable 5.2 fast mode, for example, is available on Google Cloud but was engineered on mixed hardware. The practical gap between 'runs on TPU' and 'runs best on TPU' is where migration projects stall.

What to do

Ask your Google account team for 8tx and 8i Max pricing and dates in writing. Run one production inference service through vLLM on TPU 8i and compare cost per million output tokens against your current GPU fleet. If you handle regulated data, request the confidential-compute attestation flow for 8tx and have your security team review it. And whatever the outcome, bring the numbers to your next Nvidia or cloud renewal.