Google’s new TPU 8tx keeps data and model weights sealed while a chip is working

Google has added two systems to its eighth-generation tensor processing units: TPU 8tx, built for confidential inference and training of small models, and TPU 8i Max, which puts two TPU 8i chips in one package. The launch comes with no ship dates or prices.

Confidential compute built into the silicon

TPU 8tx is based on TPU 8t, the training chip Google introduced in April, and is paired with AMD host processors. Google says the hardware keeps sensitive data and model weights “protected and isolated even while in active use.”

The target customers are regulated enterprises and governments that want to run proprietary models, and autonomous agents, close to their most sensitive data. Google says one system can fine-tune models on private data, distill large models into smaller ones and train agents with reinforcement learning, without moving the data between devices.

TPU 8tx also borrows the Boardfly network that Google designed for TPU 8i, which lets it run reasoning alongside training. Google claims lower latency than TPU 8t with a “near-zero performance penalty.” These are the company’s own figures, and the post does not say how they were measured.

Two inference chips in one package

TPU 8i Max combines two TPU 8i chips, the inference part Google launched in April. Google says that doubles the compute, memory and bandwidth of a single TPU 8i. Because the package fits twice the silicon in the existing rack footprint, Google argues that data centers get more computing power per square foot and per watt. The system is aimed at very large context windows and the highest compute density per rack.

What the announcement leaves out

The post gives only relative figures. It has no absolute numbers for FLOPS, memory capacity, pod size, power or pricing, and no general availability date for either chip. The only schedule detail comes from Bloomberg, which according to a Techmeme summary reported that both systems launch in 2027. Google’s own post does not confirm that.

The coverage also differs on purpose. TechCrunch describes TPU 8tx as geared to confidentially training and running models. Google’s post puts secure inference first, with small-scale training as a secondary use.

Signals that will confirm Google’s claims

The first thing to watch is dates. A 2027 launch would mean neither chip reaches cloud customers for months, and pricing will show whether the April generation’s economics carry over.

The second is the confidentiality claim. Hardware isolation of data in use is what regulated buyers need, but it matters only if outside security reviewers can validate it. Google’s post offers no audit or certification details.

The third is the competitive context. According to TechCrunch, Microsoft, Amazon, OpenAI and Anthropic are all working on their own AI chips and software to reduce their reliance on Nvidia. Bloomberg, through Techmeme, said the 8tx pairs with AMD CPUs, so AMD is part of the design even though the TPU silicon is Google’s.

For readers who build on Google Cloud today, little changes immediately. The two chips widen the range Google can offer later: a sealed environment for sensitive data and a denser package for long-context inference. Whether that helps depends on prices and dates that Google has not yet published.