Google's TPU 8tx and 8i Max expand enterprise AI hardware

By Tom Whitfield

Google unveiled two new Tensor Processing Unit variants on Monday: the AMD-designed TPU 8tx and the memory-heavy TPU 8i Max.

Both extend Google's eighth-generation chip line, introduced in April, which already added inference capacity and support for larger training requirements. Google develops the TPUs in-house as it seeks to reduce its cloud computing business's reliance on chips made by other companies.

Google presented the launch as two separate additions rather than one replacement chip. One focuses on confidential AI work; the other supplies more memory for demanding inference.

What does this change for enterprise AI?

TPU 8tx targets organisations that must train and run models while keeping sensitive data and model weights protected. Designed with AMD and based on the TPU 8t, the chip places that protection at the hardware layer.

Google wrote in a blog post that it “builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use.” The chip also carries a custom inferencing network, allowing reasoning to run alongside training.

That combination supports distillation and training of smaller models on the fly. A company can train on proprietary data without moving the data between devices, which narrows the exposure of information that would otherwise travel with the training process.

For buyers, the change is specific rather than sweeping. The chip puts isolation around data and weights during active use, then lets inference share the same system with training. It does not remove governance, procurement or model-management duties. It does give those duties a hardware boundary to work against.

TPU 8i Max addresses a different constraint. Google combined two inference-focused TPU 8i chips, launched in April, into a single package.

The design doubles memory capacity, processing power and bandwidth. AI models can therefore handle a larger range of information during queries, while the paired resources arrive as one unit.

Google described the two launches together as broadening its portfolio into confidential inference and greater memory capacity for large workloads. Three doubled resources make a neat specification sheet. Workload economics will decide whether the package earns its place.

The launch also fits a wider push to loosen Nvidia's grip. Microsoft, Amazon and AI labs OpenAI and Anthropic are developing AI-specific semiconductors and accompanying software for their own cloud and AI operations.

These efforts share a common rationale: proprietary hardware reduces reliance on Nvidia while giving each operator more control over the software running on its chips. Google's two additions serve that same aim, with confidentiality for one product and capacity for the other. The competition now covers both the silicon and the software attached to it.

The common thread is workload-specific hardware, not one universal accelerator. Google has split this launch by security and capacity requirements, while rivals are pursuing proprietary chips and their supporting software. Cloud and AI teams gain more architectural choice, with less dependence on a single supplier.

The timing gives Nvidia a less comfortable backdrop. Over the weekend, CEO Jensen Huang admitted to an affair with someone alleged to be a Chinese spy.

Google's chip announcement therefore arrives during a strategic contest and a personal scandal at the industry's leading supplier. The hardware comparison is straightforward: confidential training in one chip, doubled inference resources in the other. The reputational comparison is less tidy.