Google expanded its eighth-generation Tensor Processing Unit line on Monday, November 9, adding two systems built for very different jobs. TPU 8tx is designed for secure and confidential inference, while TPU 8i Max is built for massive context windows and the highest compute density per rack. With the TPU 8t and TPU 8i introduced at Google Cloud Next in April, Google has now announced four TPU systems in a single year for the first time.
The announcement also put a rare size on a business Google seldom quantifies. In the post by Amin Vahdat, Senior Vice President and Chief Technologist for AI and Infrastructure, Google said TPUs are now a rapidly growing business with nearly 100 customers, including frontier labs, trading firms, retailers and universities. The April systems set the baseline: TPU 8t for large-scale training, with up to 2.7 times better training performance than the previous generation, and TPU 8i for real-time reasoning, with 80 percent better inference price performance.
TPU 8tx, the third system in the generation, targets regulated enterprises and governments running frontier or proprietary models and autonomous agents alongside their most sensitive data. Data is normally encrypted when stored or in transit, but it has to be unlocked in memory while a computer processes it. Built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC CPU hosts, TPU 8tx builds security into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use. It also adopts the low-latency Boardfly network created for TPU 8i, delivering up to twice lower latency than TPU 8t with what Google calls a near-zero performance penalty. The same system can also fine-tune models on private data, distill large models into smaller ones and train agents through reinforcement learning, without moving sensitive data between separate clusters. Google Cloud said select customers will also be able to deploy TPU 8tx in their own data centers.
TPU 8i Max attacks a different bottleneck: the memory wall. When dozens of AI agents work together across a whole software codebase or hours of video, memory bandwidth quickly becomes the limiting factor. Google combined two low-latency TPU 8i chips onto a single package, doubling the compute, the on-chip SRAM and the high-bandwidth memory bandwidth of TPU 8i. Google Cloud's Mark Lohmeyer said the system delivers 384 GB of high-bandwidth memory. Because it fits twice the silicon within the existing rack footprint, data centers can deploy significantly more computing power per square foot and per watt.
Neither system is a new chip design. Google stressed the difference between a new chip, a multi-year process, and a new system, which recombines packaging, networking and security around existing silicon in months rather than years. The hardware arrived with a broader software push: TorchTPU was open-sourced last month so PyTorch code and Hugging Face models run natively on TPUs, TPU Sync was open-sourced to cut key-value transfer latency and time to first token by more than 90 percent, and Cloud Pathways and GKE dynamic slicing are now generally available.
Google gave no price and no availability date for either system, and named no customer for them, so every performance figure in the launch is Google's own and has not been independently tested. Even so, the direction is clear. As Dave McCarthy of IDC told Data Center Knowledge, Google's pace of specialized releases this year shows a commitment that could help it differentiate from rival cloud providers. The contest for AI infrastructure is increasingly about specialization, and Google now fields a purpose-built accelerator for each stage of the agentic lifecycle.