Google TPU 8tx and 8i Max Target Private AI and Huge Context

Google TPU 8tx and TPU 8i Max are two new AI systems Google announced on November 9, 2026, built from its existing eighth generation chips rather than new silicon. TPU 8tx keeps sensitive data encrypted even while it is being processed, and TPU 8i Max packs two chips into one package for AI agents that need huge amounts of memory.

A TPU, or Tensor Processing Unit, is Google's own AI accelerator. The new systems join TPU 8t for training and TPU 8i for low latency reasoning, which Google introduced at Cloud Next in April, making this the first year Google has launched four TPU systems, as Amin Vahdat, Google's SVP and Chief Technologist for AI and Infrastructure, wrote. Google says TPUs now have nearly 100 customers, from frontier AI labs to trading firms, retailers and universities.

What is Google TPU 8tx for?

TPU 8tx is aimed at banks, governments and other regulated customers. Normally, data is encrypted on disk and on the network but has to be unlocked in memory while a computer works on it. TPU 8tx uses the same silicon as TPU 8t, runs with 5th Generation AMD EPYC host processors and builds protection into the hardware, so data and model weights stay isolated while in use. This is known as confidential computing.

Google says TPU 8tx borrows the Boardfly network from TPU 8i and delivers up to 2x lower latency than TPU 8t with a near zero performance penalty from the security layer. Teams can fine tune models on private data, distill large models into smaller ones and train agents with reinforcement learning on one system, without moving data between clusters. If you are new to the hardware side, our TPU vs GPU explainer covers the basics.

How much more can TPU 8i Max handle?

TPU 8i Max targets the "memory wall," the point where memory bandwidth, not raw compute, slows a job down. That happens when dozens of AI agents work across an entire codebase or hours of video. Google says the dual chip package doubles the compute, SRAM and HBM bandwidth of TPU 8i. According to Network World, that means 768 MB of on chip SRAM, 17.2 TB/s of memory bandwidth and 384 GB of HBM, 33% more memory than TPU 8i.

Because it fits twice the silicon in the same rack space, data centers can get more compute per square foot and per watt. Power and floor space are the scarce resources in the current AI data center buildout, so density is a real selling point.

System	Built from	Main job	Google's key claim
TPU 8t	Eighth generation training chip	Large scale training	Up to 2.7x better training performance than the last generation
TPU 8i	Eighth generation inference chip	Low latency reasoning	80% better inference price performance than the last generation
TPU 8tx	TPU 8t silicon plus AMD EPYC hosts	Confidential inference, small scale training	Up to 2x lower latency than TPU 8t
TPU 8i Max	Two TPU 8i chips in one package	Huge context windows, rack density	Double compute, SRAM and HBM bandwidth of TPU 8i

When can customers use them?

Google's post gives no price and no availability date. Bloomberg reported that both systems are due in 2027 and framed the launch as Google stepping up its rivalry with Nvidia. On the software side, Google pointed to TorchTPU, which it open sourced last month so PyTorch code and Hugging Face models run natively on TPUs, plus TPU backends for vLLM and SGLang and a TPU agent suite in preview for coding tools such as Claude Code and Cursor.

Our take

The interesting part is not a speed record but the speed of product cycles. Google is reusing the same silicon in new packages, memory setups and security layers, which takes months instead of the years a new chip needs. Confidential computing is often what stands between a hospital or a bank and putting sensitive records into an AI model, so TPU 8tx answers a real buying objection. For everyday users the effect is indirect: more private AI services from cloud providers and more price pressure on Nvidia. All performance figures here are Google's own and have not been independently tested.