Google is developing a chip codenamed “Frozen” specifically designed to run Gemini models more efficiently for inference — the process of actually generating responses to user queries. The Information reported it first; Reuters, the Economic Times, and The Star have since confirmed the development independently. When four sources converge on the same story, it’s not a rumor anymore.
This matters more than it might appear on the surface. The headline story in AI has been about training: which model is most capable, who published the best benchmark, whose architecture is most efficient to build. The next phase of competition is about inference — running those models at scale, at lower cost, faster.
Why inference is where the economics live
Training a frontier AI model costs tens of millions of dollars and happens once (or a few times per year per model generation). Inference happens billions of times a day, every time a user asks ChatGPT a question, queries Gemini, or runs a Copilot suggestion.
The economics of AI products at scale are dominated by inference costs. A model that costs twice as much to run per query needs twice the revenue to be profitable at the same scale. Conversely, a company that reduces its inference cost by 40% through better hardware has a structural competitive advantage — it can either price more aggressively, offer higher margins, or reinvest in more capable models.
Google’s TPU program has already given it advantages over competitors who rely on commodity Nvidia GPUs. A chip optimized specifically for Gemini inference goes further: rather than building general-purpose AI accelerators, Google would be tailoring silicon to its own model architecture.
The strategic picture for Google
This is the kind of vertical integration play that only a handful of companies can execute. Apple does it with its Neural Engine. Tesla does it for autonomous driving inference. Amazon does it with Trainium and Inferentia. Google is extending what it started with TPUs into an even more targeted implementation.
For Gemini specifically, a dedicated inference chip means Google could serve more users at lower cost while maintaining response speed — a combination that directly improves the competitiveness of every Google product that runs Gemini, from Search to Workspace to Android.
The chip is reportedly still in development. Timeline to deployment and performance benchmarks aren’t yet public. But the direction is clear: Google is betting that owning the entire stack — model, software, and silicon — is the sustainable way to win at AI infrastructure.