Google unveils TPU 8tx and TPU 8i Max to widen its fight with Nvidia

Felix Laurent 9 Nov 2026 5 min read

Big tech

Google unveiled TPU 8tx and TPU 8i Max on November 9, two new systems in its eighth-generation line of AI accelerators. TPU 8tx is built for confidential AI work on sensitive data. TPU 8i Max packs two inference chips into one package for very long context windows.

Neither system uses a new chip. Both are new configurations of the TPU 8t and TPU 8i that Google showed at Cloud Next in April. With them, Google says it has introduced four TPU systems in a single year for the first time.

Amin Vahdat, Google's chief technologist for AI and infrastructure, told Bloomberg that both systems will be available next year. Google has not shared prices or cloud regions.

What TPU 8tx does

TPU 8tx targets governments and regulated companies that want to run frontier or private models on their most sensitive data. Banks, hospitals and public agencies are the obvious buyers.

The problem it solves is simple. Data is usually encrypted when it is stored or sent over a network. But it has to be decrypted in memory while a processor works on it. TPU 8tx keeps data and model weights protected and isolated even during processing, Google says.

The system uses the same silicon as the TPU 8t training chip. Its hosts are 5th Gen AMD EPYC processors. It also borrows Boardfly, the low-latency network Google built for TPU 8i. Google claims up to 2x lower latency than TPU 8t, with a near-zero performance cost from the security features.

Because it combines training-grade compute with a fast inference network, one TPU 8tx system can do several jobs. Teams can fine-tune models on private data, distill large models into smaller ones and train agents with reinforcement learning. The data never has to move between separate clusters.

Nvidia already offers confidential computing on its GPUs, as The Decoder notes. TPU 8tx closes that gap for Google. The second system aims at a different bottleneck.

What TPU 8i Max does

TPU 8i Max is built for what Google calls the memory wall. When dozens of agents work across a whole codebase or hours of video, memory bandwidth runs out before compute does.

Google's answer was to put two TPU 8i chips on a single package. Compared with one TPU 8i, the result has:

double the compute;
double the on-chip SRAM;
double the HBM bandwidth.

It also fits twice the silicon into the same rack footprint. That means more computing power per square foot and per watt. Google ties the design to the million-token context windows of its Gemini models. Vahdat told Bloomberg that long context matters for drug discovery, math and coding.

How the four eighth-generation TPUs compare

Google published full specifications only for the two April chips. For the new systems, it gave relative figures.

System	Announced	Built for	Key specs
TPU 8t	April 2026	Large-scale training	216 GB HBM, 128 MB SRAM, 12.6 PFLOPs FP4, 6,528 GB/s
TPU 8i	April 2026	Low-latency reasoning	288 GB HBM, 384 MB SRAM, 10.1 PFLOPs FP4, 8,601 GB/s
TPU 8tx	November 2026	Confidential inference, small-scale training	TPU 8t silicon with AMD EPYC hosts, up to 2x lower latency than TPU 8t
TPU 8i Max	November 2026	Long context, dense racks	Two TPU 8i chips: 2x compute, SRAM and HBM bandwidth

In April, Google said TPU 8t gives up to 2.7 times better training performance than the previous generation. For TPU 8i, it claimed 80% better inference price performance. Both figures are Google's own and have not been checked independently.

Hardware is only half of the announcement. The rest is about making TPUs easier to adopt.

How Google is making TPUs easier to use

Nvidia's strongest moat is software, since most AI code is written for its CUDA platform. Google is attacking that from several sides:

TorchTPU: open-sourced last month, it runs PyTorch code and Hugging Face models on TPUs without rewrites. It also ships TPU backends for vLLM and SGLang.
TPU agent suite: in preview, it plugs into Claude Code, Cursor and Google Antigravity. It helps with onboarding, finds bottlenecks and tunes custom kernels from plain-language requests.
XLA Serving Engine: an open-source runtime that serves trained models across TPUs, GPUs and CPUs.
TPU Sync: an open-source memory layer that moves KV caches off the chip. Google says it cuts KV transfer latency and time to first token by more than 90%.
Agent Compute Engine: pauses idle agent sessions and frees chip memory while agents wait for tools or people.

Cloud Pathways and dynamic slicing are now generally available on Google Kubernetes Engine. Google says dynamic slicing speeds up job startup by up to 5x and fault recovery by up to 4.5x.

Easier software matters only if customers can get the chips. Supply is the next question.

Who buys Google TPUs and when

Google says TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers and universities. Bloomberg names Apple, Meta and Oracle among them. Anthropic has the right to buy up to $150 billion of TPUs from 2028 through 2031, Bloomberg reported, citing a regulatory filing.

OpenAI has also been reported to be in talks to buy Google chips, though the terms are unconfirmed. In October, Oracle and Google said TPUs will reach Oracle Cloud in 2027. That will be the first big cloud besides Google's own to offer them.

Demand is running ahead of supply. Every TPU Google has is already committed to a customer, the company said in its earnings report, according to Bloomberg. Google faces shortages of memory and manufacturing capacity. It is also spending heavily on new data centers, including its Wyoming campus.

The pace of releases has changed as well. Bloomberg notes that Google's chip unit used to ship one new processor every two years.

Can TPU 8tx and TPU 8i Max challenge Nvidia

They make Google a stronger rival in two niches, but they do not threaten Nvidia's lead on their own. TPU 8tx matches a feature Nvidia already has. TPU 8i Max is a smart repackaging that Google can build quickly, since the silicon already exists.

The bigger shift is strategic. Google now sells TPUs as a broad platform, with open-source software that removes the need to rewrite code. If TorchTPU works as promised and supply catches up in 2027, more labs will have a real alternative to Nvidia GPUs. Until then, the real limit is how many chips Google can make.