Google Unveils TPU 8tx and TPU 8i Max, Expanding Its Eighth-Generation AI Chips

What Google announced

Google introduced its eighth generation of TPUs in April 2026, splitting the line into a training-oriented chip, TPU 8t, and an inference chip, TPU 8i. The two new chips build on that family, according to a report by TechCrunch.

TPU 8tx: confidential training

TPU 8tx is based on TPU 8t and was designed together with AMD. It targets confidential computing: Google says it "builds security directly into the hardware layer," keeping sensitive data and model weights protected and isolated during active use.

The chip also includes a custom inference network, which lets it run reasoning workloads alongside training. Google says this allows it to distil and train smaller models on the fly. The pitch is aimed at companies that want to train models on proprietary data without moving that data between devices.

TPU 8i Max: two inference chips in one

TPU 8i Max combines two TPU 8i chips in a single package. Compared with one TPU 8i, it doubles memory capacity, processing power and bandwidth. More memory per package lets models handle larger amounts of information during a query, such as longer contexts, which matters for reasoning models and AI agents that read and generate many tokens.

Putting two dies in one package mirrors a wider trend in AI hardware. NVIDIA's Blackwell GPUs also join two large dies into a single processor to get past the size limit of one chip.

What we do not know yet

Google has not shared:

• Performance figures (TFLOPS) or memory sizes for either chip
• Pricing on Google Cloud
• Availability dates
• Details of AMD's role beyond co-design of TPU 8tx

We will update this article when Google publishes specifications.

Why it matters

Confidential computing for AI

Many organisations, including banks, hospitals and governments, want to train and fine-tune models on data they cannot expose. Confidential computing protects data while it is being processed, not only when it is stored or in transit. Building it into an AI accelerator could make cloud training acceptable for more regulated workloads. NVIDIA added confidential computing support to its H100 GPUs, so this is becoming an important feature for AI chips.

Inference is the growth market

As we explained in our article on Google's Ironwood TPU, the cost of serving AI models to users now rivals or exceeds the cost of training them. Inference speed and capacity are limited mainly by memory capacity and bandwidth, which is exactly what TPU 8i Max doubles.

Big tech is building its own chips

Google's move is part of a broad push to reduce reliance on NVIDIA. Microsoft, Amazon, OpenAI and other companies are developing custom AI accelerators and the software to run them, and several have announced new chips or chip partnerships in 2026. Google has the longest track record: it has used TPUs since 2015 and rents them to customers through Google Cloud. For background on how TPUs differ from GPUs, see TPU vs GPU and NPU vs GPU vs TPU.

Key facts at a glance

Sources

• TechCrunch, Google broadens its AI hardware portfolio with two new specialized chips (November 10, 2026)
• Google Cloud, TPU documentation