Google on November 9 expanded its eighth-generation Tensor Processing Unit portfolio with two new systems, TPU 8tx for secure and confidential inference and TPU 8i Max for massive context windows. The additions bring the eighth-generation family to four systems in a single year, which Google said is the first time it has introduced that many TPU systems in one year in more than a decade of designing the chips.
The new systems follow TPU 8t for large-scale training and TPU 8i for real-time reasoning, which Google introduced at Google Cloud Next in April. Google said the 8t and 8i offer up to 2.7 times better training performance and 80 percent better inference price performance than the previous generation. Rather than laying out entirely new silicon, Google described the latest announcements as new systems built around existing chips, with different packaging, memory, networking, and security configured for specific jobs.
TPU 8tx is aimed at regulated enterprises and governments that want to run frontier or proprietary models and autonomous agents alongside sensitive data. Data is normally encrypted while it is stored or sent over a network, but it has to be unlocked in memory while a computer processes it. Google said 8tx, built on the same silicon as TPU 8t and paired with fifth-generation AMD EPYC CPU hosts, keeps sensitive data and model weights protected and isolated even while they are in active use, with security built into the hardware layer.
Google also said 8tx adopts the low-latency Boardfly network created for TPU 8i, delivering up to twice lower latency than TPU 8t with a near-zero performance penalty. Because the system pairs training-grade compute with that inference network, Google said teams can fine-tune models on private data, distill large models into smaller ones, and train agents through reinforcement learning on a single system, without moving sensitive data between separate clusters. Those performance and security claims are Google's own, and the announcement gave no price or general-availability date for 8tx.
TPU 8i Max addresses a different bottleneck that Google calls the memory wall. When dozens of agents work together across an entire codebase or hours of video, memory bandwidth can stall the processors. Google's answer is to combine two TPU 8i chips onto a single package. The company said 8i Max delivers double the compute, double the SRAM, and double the high-bandwidth memory bandwidth of TPU 8i, while fitting twice the silicon into the existing rack footprint, for more computing power per square foot and per watt and the highest compute density per rack in the family.
Google paired the hardware with a broader software push. It said it open-sourced TorchTPU last month so PyTorch code and Hugging Face models run natively on TPUs, alongside backends for the vLLM and SGLang inference engines. Cloud Pathways on Google Kubernetes Engine and GKE dynamic slicing are now generally available, and Google said it is open-sourcing TPU Sync, a model-state layer that tiers memory across chip, host, storage, and network. Google said TPUs now serve nearly 100 customers, including frontier labs, trading firms, retailers, and universities.
For enterprise IT buyers, the significance is specialization. Training, reasoning, confidential inference, and long-context work are diverging into separate infrastructure choices, and Google is trying to meet each one with a configured system rather than a single general accelerator. Whether 8tx and 8i Max persuade regulated customers to move sensitive agent workloads into Google Cloud will depend on the details Google has not yet published, including pricing, availability, and independent verification of the confidential-computing design.