Google Unveils TPU 8tx and 8i Max: Secure Training and Twice the Memory

If you run models in the cloud, you probably have two long-standing pain points. First: you have to hand sensitive data and model weights over to someone else's infrastructure, and your lawyers and security teams are not thrilled about it. Second: long requests run into the memory limits of the accelerator. On Monday, Google showed two new chips aimed squarely at these problems: TPU 8tx and TPU 8i Max.

What Exactly Was Released

Both chips belong to the eighth generation of TPU, which Google introduced in April. At that time the lineup included TPU 8t for large-scale training and TPU 8i for inference. Now it is expanding to cover two specific scenarios.

TPU 8tx is built on the 8t and was developed jointly with AMD. It is designed for confidential training and serving of models. According to Google, the protection is built directly into the hardware layer: sensitive data and model weights remain isolated and protected even during active operation. Data is usually encrypted at rest and in transit, but during computation it is exposed. Here, protection is claimed specifically for that stage.

TPU 8i Max is simpler: it is essentially two 8i chips in one package. Memory, compute and bandwidth are doubled. For a developer, this means the model can hold a larger amount of information in play within a single request.

Training and Inference on the Same Hardware

The 8tx has one more feature. It includes a dedicated inference network that lets you run a model in parallel with training. Google suggests using it for distillation and for training small models on the fly.

Imagine a bank or a clinic. They have private data on which a compact model needs to be fine-tuned for their task. Usually that means shuttling data between different devices: training on some, validation and inference on others. Each such transfer enlarges the attack surface and the amount of approvals required. Google says that on the 8tx the data does not need to be moved: the model learns and responds on the same device inside a protected perimeter.

Now, about the 8i Max. If your service works with long documents, large codebases or multi-step agent scenarios, memory often becomes the main constraint. Instead of splitting the model or the context across several chips, you can take one chip with doubled resources. Fewer inter-chip transfers usually means a simpler deployment architecture, although Google does not give specific latency figures in the announcement.

Why Google Is Doing This and What You Should Do

This is part of the larger race for independence from Nvidia. Google builds its own TPUs to depend less on other companies' chips in its cloud business. Microsoft and Amazon, as well as labs like OpenAI and Anthropic, are also working on their own AI chips and accompanying software. Google's bet is clear: if confidential data processing and large memory become arguments in favor of TPU, that will help sell the cloud to companies for which confidentiality matters.

There are also limitations to consider. The announcement has no availability dates, prices or independent measurements. Claims about protecting weights and data remain Google's claims for now, and auditor verification is still ahead. And as always with TPU, you are tying yourself to Google's stack, and moving to GPUs will take work.

The practical takeaway is this.

• If you work in a regulated industry and train models on private data, keep an eye on TPU 8tx availability and ask Google for details on its threat model and certification.
• If memory is choking you on long context, try the 8i Max on your own workloads as soon as pricing and access appear.
• If you are choosing a platform for the next year or two, factor in that the big clouds are gaining real alternatives to Nvidia, which means room to negotiate.

The announcement is so far more of a pitch than a result. But the direction is clear: cloud providers are starting to compete not only on power, but also on how comfortable you can feel handing them your data.