Google is building a chip that hardwires Gemini into silicon

Google has a compute problem. Its Gemini models keep getting bigger, its data centers are running hot, and internal documents suggest the company has been turning away cloud customers because it simply doesn’t have enough AI compute capacity. So Google is taking a radical approach: build a chip that hardwires Gemini’s architecture directly into silicon.

The chip, internally code-named “Frozen v2,” is not another general-purpose GPU or TPU. It is a dedicated server chip designed exclusively for the Gemini family of models, according to sources who spoke with The Information. By etching the model’s underlying computation logic into the hardware itself, Google expects it to process 6 to 10 times more tokens per watt than its existing custom AI accelerators.

The “Frozen” name is deliberate — it describes a design philosophy of freezing part of the model’s operations permanently into the silicon. Where a general-purpose chip like an Nvidia GPU or Google’s own TPU has to execute thousands of logic decisions on the fly for every inference request, Frozen v2 comes pre-wired for Gemini’s math. That means fewer compute steps, less data shuttling between memory and processor, and faster responses.

This isn’t the first time Google has tried this. An earlier effort, also led by Google DeepMind’s chief scientist Jeff Dean, attempted to bake full model weights directly into a chip. That approach had a fatal flaw: the chip could only run a single version of Gemini, making it obsolete as soon as the model updated. The project was shelved for years.

Frozen v2 takes a different approach. Instead of freezing the complete model weights — the trained parameters that determine how the model answers questions — it only hardens the underlying architectural logic into the chip. The model weights themselves remain updatable through software. This gives Google the flexibility to iterate on Gemini’s performance without throwing away the hardware. The tradeoff is that Google is now betting that its future Gemini models will keep the same fundamental architecture; if the company ever decides to fundamentally redesign the model, Frozen v2 would lose its purpose.

The engineering team is still working out exactly how much of the model’s logic to freeze. Too little, and the chip loses its efficiency advantage. Too much, and it becomes inflexible again. The balance will determine whether this is a breakthrough or an expensive experiment.

Frozen v2 will sit alongside Google’s TPU line, not replace it. Google just shipped its eighth-generation TPU this year and has been aggressively marketing it to cloud customers as an alternative to Nvidia’s dominant hardware. The company has already signed a multibillion-dollar deal with Meta to lease TPU compute, and it is actively pitching TPUs to Nvidia’s core cloud client base. But TPUs are still general-purpose enough to run a wide range of models. Frozen v2 is the opposite — narrowly optimized for one architecture, with no ambitions beyond Gemini.

Google has no plans to mass-produce Frozen v2. Sources say production volumes will be far smaller than TPU output, partly because the company can afford to delay bringing in external design and manufacturing partners. Internally, the project is viewed as a technology testbed — a way to build engineering knowledge for future, more specialized designs once large model architectures settle down.

A Google spokesperson declined to comment on Frozen v2 specifically but said the company’s teams “continuously research, develop and test various innovative technologies to deliver extreme performance and energy efficiency for our users and enterprise customers,” adding that “not every project makes it to mass production, but this kind of comprehensive technical exploration is at the core of our full-stack in-house approach.”

For now, the earliest Frozen v2 chips could reach deployment is 2028. That gives Google four years to figure out whether hardwiring a model into silicon is the future of AI infrastructure — or a costly detour.