Alphabet is designing a new server chip, internally called “Frozen v2,” that could dramatically cut the cost of running its Gemini AI models. The Information first reported the project, citing anonymous sources who claim the chip will be six to ten times more efficient than Google’s current silicon when measured by tokens generated per unit of power. A 2028 release is the target. Google’s response to TechCrunch was a classic non-denial: “Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency,” a spokesperson said, adding that “not every project moves into production.”
That careful language makes sense. $180 to $190 billion in planned capital expenditure has investors watching Alphabet’s every move, and the pressure to show returns is real. Monday morning, the stock jumped roughly 3% on the Frozen v2 news alone — a tidy boost ahead of this week’s earnings call. If Google can shrink its per-token power consumption by an order of magnitude, the math on those massive AI infrastructure bets starts looking a lot friendlier.
The efficiency race isn’t happening in a vacuum. OpenAI shipped its own inference chip, Jalapeño, last June. Anthropic is reportedly talking to Samsung about a custom silicon partnership. Everyone wants to loosen Nvidia’s grip on the AI hardware supply chain, and in-house chips are the obvious escape hatch. Google has a head start with its TPU program, but a 6-10x leap in efficiency suggests something more radical than incremental iteration.
What’s still unclear is whether Frozen v2 is a traditional ASIC, some kind of optical or analog design, or something else entirely. The internal name could be a placeholder, or it could hint at the kind of cooling required. Either way, 2028 is a long way off — and a lot can change in the chip landscape before then. If the efficiency claims hold up, though, Google won’t just be competing with Nvidia. It’ll be rewriting the economics of serving frontier models.
