Google adds TPU 8tx and TPU 8i Max, its third and fourth TPUs this year

Data is usually encrypted at rest and in transit but has to be unlocked in memory while a computer works on it. TPU 8tx builds protection into the hardware so sensitive data and model weights stay isolated even while in use[1]. It is built on the same silicon as TPU 8t, paired with 5th Generation AMD EPYC CPU hosts, and uses the low-latency Boardfly network Google designed for real-time reasoning. Google says that gives up to 2x lower latency than TPU 8t with a near-zero performance penalty.

Google is pitching it at regulated enterprises and governments that want to fine-tune on private data, distil large models into smaller ones and train agents with reinforcement learning without moving sensitive data between clusters.

TPU 8i Max puts two TPU 8i chips on a single package, doubling compute, SRAM and HBM bandwidth compared with TPU 8i[1]. Google's argument is that when dozens of agents work across an entire codebase or hours of video, memory bandwidth, not raw compute, becomes the bottleneck. Because it fits twice the silicon in the existing rack footprint, it also packs more compute per square foot and per watt.

Google is careful to call these new systems rather than new chips. Its TPUs were designed as modular building blocks, so packaging, high-speed memory, optical networking and security can be changed around the chip far faster than a new chip can be designed and manufactured[1].

The company says TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers and universities, and that some customers now deploy TPUs directly in their own data centres.

Much of the announcement is about making TPUs easier to adopt for developers who live in PyTorch[1]:

• TorchTPU, open-sourced last month, runs PyTorch code and Hugging Face models natively, alongside TPU backends for vLLM and SGLang.
• A TPU agent suite in preview plugs into Claude Code, Cursor and Google Antigravity to automate onboarding, find bottlenecks and tune kernels.
• TPU Sync, now open source, treats chip HBM, host memory, storage and the network as one memory hierarchy, which Google says cuts KV cache transfer latency and time to first token by more than 90%.
• Agent Compute Engine suspends and resumes idle agent sessions while they wait on tools or people.
• Cloud Pathways and GKE dynamic slicing are now generally available, with Google claiming up to 5x faster job start-up from dynamic slicing.

Google did not publish pricing or availability dates for either new system.

Nvidia's biggest advantage has long been software, not just silicon. By shipping specialised hardware variants every few months and removing the need to rewrite PyTorch code, Google is trying to make TPUs a default option for agent workloads rather than a specialist choice. Confidential computing on TPU 8tx also opens doors in government and regulated industries that have been wary of running frontier models next to sensitive data.