Google is building a new AI chip that hardwires its Gemini model directly into silicon — making the company that already decides what billions of people see online faster, cheaper, and harder to compete with.
The chip, internally called "Frozen v2," would bake Gemini's neural-network architecture into the circuitry itself, according to The Information. Instead of a model onto general-purpose hardware, the chip becomes the model. The reported payoff: six to 10 times more efficiency than Google's current custom AI chips, measured by tokens generated per unit of power. Deployment is targeted for as early as 2028.
For ordinary Americans, the stake is straightforward. A company with a documented track record of suppressing search results, demonetizing dissent, and funneling users toward approved narratives is building dedicated hardware to run its AI cheaper and faster. When the infrastructure of speech control gets more efficient, dissent gets more expensive. TNW reported that Frozen v2 would lock Gemini's architecture into the circuitry while keeping the model's "weights" updatable — meaning Google can still refresh what the AI outputs, but the underlying structure stays fixed. The project is partly a response to an AI capacity crunch inside Google severe enough that Google Cloud has turned away outside customers, stirring internal tensions.
Follow the money. Forbes reported that Alphabet shares jumped as much as 3.7% on the news, adding a combined $15 billion to the net worths of cofounders Larry Page and Sergey Brin. Alphabet plans to spend between $180 billion and $190 billion building out its AI strategy. Wall Street anticipates quarterly spending of roughly $45 billion — a 26% increase over the previous quarter. Someone is paying for all of this, and it isn't just advertisers.
Google gave TechCrunch a classic non-denial: "Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers. While not every project moves into production, this rigorous exploration is central to our full stack approach." Translation: they didn't deny it.
There's a strategic angle beyond raw efficiency. Google already designs its own TPUs to cut dependence on Nvidia, whose dominance has left AI makers dependent on its hardware. A Gemini-specific chip deepens that self-reliance. OpenAI announced its own custom inference chip, Jalapeño, in June. Anthropic is reportedly discussing a chipmaking partnership with Samsung. The big AI players are racing to own the full stack — hardware, software, and the models that shape what you're allowed to see.
The catch, as TNW noted, is rigidity. AI moves fast, and a chip built around today's Gemini architecture could look dated by 2028. But Google is betting that serving one model to billions makes the trade-off worth it. A startup called Taalas is already selling the same concept, claiming its chip serves up to 17,000 tokens per second versus roughly 150 per user on a top Nvidia GPU — no expensive high-bandwidth memory required.
TechCrunch framed the story as an efficiency play amid investor worries about AI spending. Forbes framed it as a stock rally that made two billionaires $15 billion richer. TNW alone explained the technical significance — that the model gets etched into the hardware itself. What none of them asked: what happens when the infrastructure of AI-mediated speech gets cheaper, faster, and more centralized in the hands of companies that have already proven willing to silence disfavored voices?
The chip is years away. The consolidation of power is not.