Google ships two new TPU systems in one year: 8i Max and 8tx
Google announced TPU 8tx and TPU 8i Max on November 9, 2026, the third and fourth systems of its eighth-generation TPU portfolio after TPU 8t and TPU 8i, which shipped at Google Cloud Next in April. According to the company, this is the first time it has shipped four TPU systems in a single year, and both new ones are aimed squarely at agentic workloads: each system in the portfolio targets a different part of the AI lifecycle.
Why four systems instead of one
Google positions the eighth generation as a full pipeline: 8t for scale-out training, 8i for low-latency reasoning, 8tx for confidential compute, and 8i Max for massive context windows and the highest compute density per rack. The April pair set the baseline, and the numbers in the post are Google's: 8t delivers up to 2.7x better training performance than the previous generation, and 8i delivers 80% better inference price performance. The two systems announced today answer requests Google says it has been hearing from external customers: hardware-level security for regulated workloads, and memory headroom for million-token contexts.
Google also puts a number on the business. The post says TPUs now serve nearly 100 customers, including frontier labs, trading firms, retailers and universities, and that use cases now extend past Google Cloud: customers can deploy TPUs directly into their own data centers.
TPU 8tx: confidential inference
TPU 8tx builds on the same silicon as 8t, pairs it with 5th generation AMD EPYC CPU hosts, and moves the security boundary into the hardware. According to Google, data and model weights stay encrypted and isolated even while the system is processing them. It also reuses the low-latency Boardfly network created for 8i, which Google says delivers up to 2x lower latency than 8t with a near-zero performance penalty.
The practical consequence is the interesting part for regulated teams: you can fine-tune models on private data, distill large models into smaller ones, and train agents through reinforcement learning on a single system, without migrating sensitive data between clusters. The announcement pitches this at enterprises and governments running frontier or proprietary models alongside sensitive data.
TPU 8i Max: two chips, one package
TPU 8i Max targets the memory wall. Google combines two low-latency 8i chips in a single package, and the result is double the compute, double the SRAM and double the HBM bandwidth of a single 8i. Because the package fits twice the silicon in the existing rack footprint, a data center deploys more compute per square foot and per watt, per the company. The target workload is agent fleets reasoning over an entire codebase or hours of video at once, where memory bandwidth becomes the bottleneck first.
The software catch-up
The hardware is half of it. Alongside the two systems, Google open-sourced several pieces of its stack. TorchTPU, released last month, runs PyTorch code and Hugging Face models natively on TPUs, with TPU backends for vLLM and SGLang. TPU Sync, the new model-state layer, unifies chip HBM, host DRAM, shared storage and the network into one memory hierarchy, and Google says it cuts KV transfer latency and time to first token by more than 90%. An Agent Compute Engine suspends and resumes idle agent sessions to cut infrastructure costs, and a TPU agent suite, now in preview, plugs into coding tools such as Claude Code and Cursor to diagnose bottlenecks and autotune kernels. On the cloud side, Cloud Pathways (the DeepMind runtime behind Gemini) and GKE dynamic slicing both went generally available, with Google claiming dynamic slicing cuts job startup times by up to 5x and accelerates fault recovery by up to 4.5x. Google also says its vLLM on TorchTPU stack hit state-of-the-art price performance on the InferenceX and AgentX benchmarks.
What to take from this
If you're weighing compute for production agents, the announcement gives you four distinct options inside one vendor's portfolio, each aimed at a different stage of the workload. The strategic detail is the pace: four systems in one year works because Google is shipping systems, packaging, networking, memory and security around a chip, rather than a single chip. For teams sitting on GPU-centric stacks, the TorchTPU and PyTorch push is the part worth watching, since it goes directly at the CUDA lock-in argument. For anyone running regulated inference, 8tx is the first TPU option aimed squarely at confidential workloads.
Source: blog.google
More news
Have something in mind?
Start a conversation