Google unveiled two new eighth-generation Tensor Processing Units on Monday: TPU 8tx, built for secure and confidential inference, and TPU 8i Max, built for massive context windows and the highest compute density per rack. Google said this is the first time it has introduced four TPU systems in a single year.
The new systems join TPU 8t and TPU 8i, which Google introduced at Google Cloud Next in April for large-scale training and real-time reasoning. Google says TPU 8t and 8i offer up to 2.7 times better training performance and 80% better inference price performance, respectively, than its previous TPU generation.
TPU 8tx is built on the same silicon as TPU 8t and paired with AMD CPU hosts, which Google's announcement describes as 5th Generation AMD EPYC processors. Security is built into the hardware layer, so sensitive data and model weights stay protected and isolated even while they are being processed. Google is aiming the chip at regulated enterprises and governments, saying some highly regulated customers have asked for specialized encryption capabilities.
Key facts
- 01Google announced TPU 8tx and TPU 8i Max on Nov. 9, 2026, its third and fourth TPU systems of the year.
- 02TPU 8tx uses TPU 8t silicon with AMD CPU hosts and delivers up to 2x lower latency than TPU 8t.
- 03TPU 8i Max packs two TPU 8i chips into one package, doubling compute, SRAM and HBM bandwidth.
- 04Google says its TPU business now has nearly 100 customers, including frontier labs and trading firms.
- 05Google's open-source TPU Sync cuts KV transfer latency and time to first token by more than 90%.
TPU 8tx also borrows the low-latency Boardfly network Google built for TPU 8i, delivering up to 2x lower latency than TPU 8t with what Google calls a near-zero performance penalty. Google says teams can use one TPU 8tx system to fine-tune models on private data, distill large models into smaller ones and train agents with reinforcement learning, without moving sensitive data between separate clusters.
TPU 8i Max takes two low-latency TPU 8i chips and combines them onto a single package. The result has double the compute, double the SRAM and double the HBM bandwidth of TPU 8i, and because it fits twice the silicon into the existing rack footprint, data centers get more compute per square foot and per watt. Google pitched it for workloads like dozens of agents working across a codebase or hours of video.
“This really follows the thesis that we have here at Google, which is that the future is going to be about specialization”— Amin Vahdat, Google's chief technologist for AI and infrastructure
The announcement was authored by Amin Vahdat, Google's SVP and Chief Technologist for AI and Infrastructure. He wrote that TPUs, first designed more than a decade ago for internal needs like voice recognition, are now a rapidly growing business with nearly 100 customers, including frontier labs, trading firms, retailers and universities. TPUs also train and serve Google's own Gemini models, and customers can now deploy TPUs directly in their own data centers.
Google paired the hardware with software updates. TorchTPU, open-sourced last month, lets PyTorch code and Hugging Face models run natively on TPUs with vLLM and SGLang backends. A new open-source TPU agent suite, now in preview, plugs into Claude Code, Cursor and Google Antigravity, and Google's open-sourced TPU Sync cuts KV transfer latency and time to first token by more than 90%.
On the infrastructure side, Cloud Pathways is now generally available on Google Kubernetes Engine, as is GKE dynamic slicing, which Google says makes job startup up to 5x faster and fault recovery up to 4.5x faster. Google also said AlphaChip and Gemini now produce chip layouts in hours rather than weeks.
Google did not disclose pricing, availability dates, pod sizes or memory capacities for either new system.
The launch landed on a volatile day for AI chip stocks. Nvidia shares were down about 4% on Monday after its board opened an independent investigation involving CEO Jensen Huang, while Alphabet shares rose about 3.4%.
Google now positions its eighth-generation lineup as four specialized parts: TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing and TPU 8i Max for massive context windows.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
