An ice-cold bet on a chip that runs only one AI model

TL;DR: Google’s reportedly cooking up a new kind of chip with Gemini’s skeleton baked right into the silicon, letting it run the model far more efficiently. The design could be game-changing for compute-strapped AI firms and the devices we use every day—but there’s still a lot we don’t know.

What happened: Let’s get the “Do you want to build an AI chip?” jokes out of the way… yes, the chip’s informal name is Frozen v2—and Google could deploy it as soon as 2028, according to the Information. (Though Google hasn’t yet confirmed the project exists.)

The killer feature: Frozen could be six to 10 times more efficient than Google’s latest line of custom AI chips—which could be a lifesaver for a company deep in a compute crunch.

The cold never bothered it anyway: While only part of Gemini will be permanently frozen into the chip, it’s still a sharp break from general-purpose processors that currently dominate the market, like Nvidia’s GPUs.

Current chips are built to run all sorts of models, which means they have to make more decisions. By hardwiring Gemini’s architecture (the underlying blueprint of how it processes information) into the chip, Frozen can skip those decisions and move less data around (making it a lot faster). The catch: These chips would be stuck with whatever version of Gemini’s architecture they’re built around—but Google could still push updates to the model itself (to a certain degree).

BYO silicon: Google’s been building inference chips for about a decade. Frozen’s just the latest example of the rush to build more efficient ones—which will make AI tasks cheaper to serve (and therefore cheaper for users). Chipmakers are already full steam ahead—and AI labs are starting to dip their toes into bespoke silicon. OpenAI unveiled its first custom chip with Broadcom in June, and Anthropic is working with Samsung, per an Information report earlier this month.

Pick a model and stick to it: If Frozen makes Gemini so much faster, it could eventually reshape how we use AI on our devices (like picking the built-in model versus switching between options). One thing Apple’s reportedly looking into is shrinking AI models so they can live directly on your iPhone—running faster and leaning less on memory.

Bottom line: There’s so much about Frozen and this style of chip design that’s still unproven—but Google is a big enough company that its experiment could shake up the tech industry if it proves feasible. —WK