Google now has four systems in its eighth-generation TPU line. It has added the TPU 8tx, which targets work on confidential data, and the TPU 8i Max, which is meant for AI agents that need a huge amount of memory close at hand. No earlier year has produced that many TPU systems from Google.
Start with the 8i Max, because the idea is simple: take two TPU 8i chips and put them in one package. Google says that doubles the compute, the on-chip SRAM and the memory bandwidth of a single 8i, while still fitting the same rack space. The pitch is for jobs where many agents read a whole software project, or hours of video, in one go and the chip runs out of memory speed before it runs out of arithmetic.
The 8tx deals with a different worry: who can see data while a model is using it. Disks and networks are routinely encrypted, but the data has to be readable in memory during processing, which is the moment a regulated bank or a government agency would rather not trust a cloud with. Google’s answer is to wall off the data and the model’s weights inside the hardware itself for as long as a job runs. In Google’s announcement the 8tx shares its chip with the 8t, runs beside AMD EPYC processors and uses the 8i’s fast network. Google adds that one system can also customise a model on a company’s own records, so nothing has to be shipped elsewhere.
The two newcomers sit alongside the TPU 8t for training and the TPU 8i for low-latency reasoning, both shown at Google Cloud Next in April. Against the previous generation, Google put the 8t at up to 2.7x the training speed and the 8i at 80% better value for inference. It says close to 100 customers use TPUs, among them AI labs, trading firms, retailers and universities, and that some now install them inside their own data centres.
Most of the rest of the post is about making TPUs easier to adopt. Developers who write in PyTorch get most of the attention: Google open-sourced TorchTPU in October, and it says Hugging Face models and the vLLM and SGLang serving engines now run on TPUs. A preview agent suite hooks into Claude Code, Cursor and Google Antigravity to help with setup and tuning. TPU Sync, now open source, spreads the model’s working memory across the cluster and, in Google’s numbers, cuts time to first token by over 90%. Cloud Pathways and GKE dynamic slicing are generally available, with job start-up up to 5x quicker for the latter.
Every performance claim here is Google’s own, and the post gives no prices. The company pointed to its broadcast and Google Cloud’s website for the full lineup.
