Google announces TPU 8tx and TPU 8i Max, two new TPU chips
Google announced two new systems in its eighth-generation TPU lineup on November 9, 2026: the TPU 8tx, built for secure and confidential inference, and the TPU 8i Max, which packs two TPU 8i chips into one package. Google says it is the first time it has introduced four TPU systems in a single year. The announcement gives no pricing, general availability date or regional rollout for either system.
What’s new
TPU 8i Max combines two low-latency TPU 8i chips on a single package. Compared with the TPU 8i, Google says it has double the compute, double the SRAM and double the HBM bandwidth. It fits twice the silicon in the same rack footprint, which Google says gives the highest compute density per rack in the lineup. The target workloads are large context windows, such as many agents working across a full codebase or hours of video.
TPU 8tx is aimed at regulated enterprises and governments that run frontier or proprietary models and agents on sensitive data. Google says it is built on the same silicon as the TPU 8t and paired with 5th Generation AMD EPYC CPU hosts. Hardware-level confidential compute keeps data and model weights isolated while they are in use. It uses the Boardfly network from the TPU 8i, and Google claims up to 2x lower latency than the TPU 8t with a “near-zero performance penalty”. Google lists fine-tuning on private data, distilling large models into smaller ones and reinforcement learning for agents as use cases, without moving sensitive data between clusters.
The rest of the eighth generation
According to Google’s post, the TPU 8t handles training, with up to 2.7x better training performance than the previous generation, and the TPU 8i handles real-time reasoning, with up to 80% better inference price performance than the previous generation. Google says TPUs have nearly 100 customers, including frontier labs, trading firms, retailers and universities.
Google also announced software alongside the hardware. TorchTPU was open-sourced last month, so it is not new today; the new items are TPU Sync (open source) and Agent Compute Engine:
- TorchTPU (open-sourced last month) for running PyTorch natively on TPUs, plus vLLM and SGLang backends.
- A TPU agent suite (preview), an XLA Serving Engine, Agent Compute Engine and TPU Sync.
- Cloud Pathways (generally available on GKE) and GKE dynamic slicing (generally available).
What it means for developers
Most developers will not buy a TPU directly, but the software side matters sooner than the hardware:
- PyTorch on TPUs: TorchTPU and the vLLM and SGLang backends mean serving stacks you already use are meant to run on TPUs without a rewrite. Check Google’s documentation for the supported versions before planning a migration.
- Long-context agents: If your agents load a whole repository or long video into context, the 8i Max is the part of the lineup designed for that memory pressure.
- Sensitive data: The 8tx is relevant if compliance blocks you from sending private data to shared inference. Until Google publishes availability and pricing, treat it as a roadmap item, not something to budget for.
How it compares
Within Google’s own lineup, the 8t is for training, the 8i for low-latency inference, the 8i Max for the same at higher density, and the 8tx for confidential inference and small-scale training. Google claims state-of-the-art price-performance on the InferenceX and AgentX benchmarks for its vLLM-TorchTPU stack, but gives no chip-vs-chip numbers in the post, so any comparison with GPUs would be speculation. For another recent move in AI infrastructure and funding, see our post on OpenAI’s reported $1.25 trillion valuation.