Google has added two systems to its eighth-generation TPU line: one for agents that work through enormous amounts of code or video at once, and one that keeps a customer’s private data protected while a model is actively using it. Amin Vahdat announced TPU 8i Max and TPU 8tx on the Google blog on 9 November. Google does not say what either will cost or when customers can order them.
A TPU is Google’s own AI processor, the in-house counterpart to the GPUs most of the industry buys. Google began designing them more than a decade ago for jobs like voice recognition and translation, and the blog says nearly 100 customers now use them, among them AI labs at the frontier, trading firms, retailers, and universities. Four systems in one year is a first for the company, by its own account. The first two, TPU 8t for training and TPU 8i for real-time reasoning, arrived in April.
TPU 8i Max takes aim at what Google calls the memory wall. Picture dozens of agents working across a whole codebase or hours of footage: the chip spends its time waiting for data rather than computing. Google’s fix is to put two TPU 8i chips on a single package. Against the 8i, that gives double the compute, double the on-chip SRAM, and double the bandwidth to its high-bandwidth memory. Because it fits in the same rack footprint as before, a data center gets roughly twice the silicon in the same floor space.
TPU 8tx is for confidential inference, which means running a model on sensitive data without exposing that data to anyone, including the cloud operator. The problem it addresses is a gap in ordinary security. Data is normally encrypted on disk and while traveling over a network, but must be decrypted in memory before the computer can process it. 8tx is built to keep both the data and the model’s weights shielded and isolated even during that step.
The target buyers are regulated enterprises and governments that want to run frontier or in-house models and agents next to their most sensitive records. The system is built on the 8t’s silicon and hosted by AMD’s fifth-generation EPYC processors, and borrows the low-latency Boardfly network from the 8i. Google says that yields up to twice the speed of the 8t in latency terms and almost no performance cost. A bank could therefore fine-tune a model on customer files, shrink it, and train an agent on it without ever moving the data to another cluster.
Google also pushed a stack of software alongside the hardware. It open-sourced TorchTPU last month so existing PyTorch programs and models from Hugging Face run on TPUs unchanged, and it is releasing TPU Sync, which spreads memory across chip, host, and storage layers. Google says TPU Sync cuts time to first token by more than 90 percent in its tests. A preview agent suite plugs into coding tools including Claude Code and Cursor to set up TPU workloads.
Every number here comes from Google’s blog. The comparisons are against Google’s own earlier chips, the benchmark wins cited for its vLLM-TorchTPU stack are described only as state of the art, and the post offers no independent test of the security claims. For a regulated buyer, the confidential-computing pitch will need an outside audit before it counts for anything.
Four systems in seven months changes the buying question from whether to try a TPU to which of four to pick. For the two newest, Google has yet to say when they ship or what they cost.
Reported by the Google blog on 9 November 2026.
