Google broadens its AI hardware portfolio with two new specialized chips

Google added two variants to its eighth-generation tensor processing units. TPU 8tx, built with AMD on top of the TPU 8t, is designed to train and run models confidentially, with protection for data and model weights built into the hardware while they're in use. TPU 8i Max packs two of April's inference-focused TPU 8i chips into one package, doubling memory, compute and bandwidth so models can hold more information per query.

This is about who owns the cloud AI stack. Microsoft, Amazon, OpenAI and Anthropic are all building their own silicon to lean less on Nvidia. Google is now slicing its lineup by use case, not just by generation, which is what a mature chip business does.

We think the confidential-computing chip is the more interesting move. The biggest drag on enterprise AI adoption isn't model quality anymore. It's that legal and compliance teams won't let sensitive data leave the building, or sit unprotected on someone else's hardware. A chip that keeps data and weights isolated during active use, and that can distill and train small models on proprietary data without shuttling it between devices, goes straight at that objection.

If it works as described, that's useful AI: banks, hospitals and governments could fine-tune on the data that actually matters to them. But hardware security claims deserve skepticism until outside researchers have tried to break them. Secure enclaves have a long history of side-channel surprises, and a chip marketed as safe for the most sensitive workloads raises the cost of being wrong.

The AMD partnership is worth a beat too. Google designing a TPU variant with AMD tells us the anti-Nvidia coalition is getting practical. Nobody wants to be dependent on one supplier, and the hyperscalers have both the money and the motive to keep chipping away.

TPU 8i Max is the less glamorous play, but inference is where the recurring bill lives. Doubling memory in one package is a direct answer to longer contexts and agent workloads that keep more in play. I'm watching whether Google prices it to win share or to protect margin.

The timing is pointed. This lands while Nvidia's CEO is under scrutiny after admitting an affair with someone alleged to be a Chinese spy. Customers don't swap chip platforms over a news cycle, but uncertainty at the top of your main supplier is exactly when a procurement team takes the second meeting.

Google introduced its eighth TPU generation in April with an emphasis on more inference capacity and support for large training jobs. TPUs are sold through Google Cloud rather than as standalone parts, so each new variant is really a new cloud product. Google described the new chips in a company blog post.