Google Is Building a Chip With Gemini Baked Directly Into the Hardware

Google is developing a new server chip that would embed elements of its Gemini model directly into silicon, The Information reported Monday, citing people familiar with the project. Codenamed "Frozen v2," the Google Gemini chip could deliver six to 10 times better efficiency than Google's current custom AI chips, measured by AI tokens served per unit of power. Google did not respond to a request for comment. Alphabet shares climbed 3% on the news.

"Frozen v2" Bakes Gemini Into the Hardware - and Sits Alongside TPUs, Not Above Them

"Frozen" refers to a model whose weights are hardwired into the chip rather than loaded dynamically at runtime. Engineers are still finalizing how much of Gemini gets locked into the design, which matters because baked-in weights cannot be updated without new silicon - a real constraint for a model family Google keeps actively revising.

Google targets deployment as soon as 2028, though the design is not yet final. Critically, Frozen v2 is a separate project from Google's tensor processing units (TPUs) and is not meant to replace them. TPUs remain the general-purpose workhorse for training and flexible inference; Frozen v2 would handle high-volume production serving where the model version is stable and efficiency trumps flexibility.

Google's Compute Crunch Is the Real Reason This Project Exists

The report frames the chip as a direct response to a capacity crisis inside Google. Google has already cut off outside customers from Gemini access as internal demand outstrips supply - a pressure that has caused internal tensions and forced Google Cloud to turn away business. A chip that serves six to 10 times more tokens per watt would relieve that strain without needing to build proportionally more data center infrastructure.

Timing matters here. Gemini 3.5 Pro has already missed multiple internal deadlines, falling short on coding benchmarks in a way that forced the team back to work. A chip optimized around a "frozen" model only makes sense once that model is stable enough to commit to silicon - which may explain why the 2028 target leaves room for the model team to catch up first.

Google Is Not the Only One Going Custom on Inference Silicon

Building model-specific inference chips is now a pattern across the industry. OpenAI and Broadcom unveiled Jalapeño earlier this year, targeting LLM inference efficiency with a custom design rather than leaning on Nvidia's H100s for production serving. Micron has also seen demand quadruple for HBM memory that feeds these inference workloads - a sign that the economics of serving AI at scale are forcing hardware decisions that would have seemed premature two years ago.

For Google, the calculus is straightforward: Gemini is its most expensive product to run and its most important to monetize. A 6-10x efficiency gain on serving costs - if Frozen v2 delivers anywhere near that - would change the unit economics of Google AI Pro, Gemini API calls, and every Workspace feature that runs a model in the background. Whether engineers can finalize a design worth freezing by 2028 is the open question.