Summary: The Information reported on July 20 that Google is developing a server chip called Frozen v2 to run Gemini with far less electricity, reportedly targeting six to ten times current AI-chip efficiency by 2028. Google would neither confirm nor deny the project. The article treats the report as unconfirmed and says near-term infrastructure choices should stay on TPU and existing GPU routes.
On July 20, The Information reported that Google is developing a server chip it refers to internally as Frozen v2, and that the project has exactly one purpose: making its Gemini models run on far less electricity. Google would neither confirm nor deny the reporting, offering only a stock statement. The market moved first — Alphabet stock rose about 3% on Monday morning once the story broke, landing just before the company’s earnings report that same week. A story sourced from anonymous insiders rarely gets a welcome that warm, and the reaction says something about where investor attention sits in 2026: not on fresh model demos, but on the power bill behind them.
A Chip Built for Gemini
What matters here isn’t yet another TPU — it’s that the premise the design starts from has changed. The TPUs Google ships today are general-purpose accelerators that carry every workload the company runs; Frozen v2, as reported, is aimed squarely at Gemini, meaning the model’s own characteristics are treated as given inputs to the hardware. The Next Web summed the route up as “a chip with Gemini baked into the silicon.” Google’s answer followed the familiar script: “Our teams are constantly researching and experimenting with new innovations to deliver maximum performance,” followed by the caveat that “not every project moves into production” — an acknowledgment that the project is real, paired with room to walk away from it. That wording also leans on Google’s “full stack approach,” and it may be the most honest framing available: since the company sells the model, the chip and the data center together, designing all three at once is where this was always heading.
The Efficiency Target: Six to Ten Times
As reported, Frozen v2 is due in 2028 and could be six to ten times more efficient than Google’s current AI chips, with the yardstick being tokens generated per unit of power. Choosing that yardstick is a statement in itself — the coming fight is over inference, not training. Training burns a fortune once; inference burns a somewhat smaller fortune forever. As application layers and agent workloads drive token consumption up by orders of magnitude, the electricity cost attached to each token decides serving margins outright — and agents make the problem worse, since one user request can be worth dozens or hundreds of model calls.
The Custom Silicon Arms Race
Frozen v2 is no isolated case. OpenAI showed off its first custom chip in June — Jalapeño, an inference processor built with Broadcom, and notable for targeting inference from day one. Anthropic is said to have been in talks with Samsung about a chipmaking partnership. The motivation is plain on both fronts: lean less on NVIDIA, and take pressure off the compute shortage. NVIDIA remains the supplier everyone defaults to for the industry’s training fleets, and once training and inference demand both outrun supply, custom silicon stops being a cost-optimization option and becomes supply-chain insurance. The constraints differ from lab to lab — OpenAI and Anthropic rent most of what they run, while Google owns its stack — yet all three land in the same place: own the silicon your own workload actually needs.
Capex and the Market Reaction
Set against that, the backdrop numbers are just as striking: Alphabet’s plan calls for $180–190 billion of spending on its AI buildout, and investor unease about the scale of it has not gone away. The efficiency story around Frozen v2 answers that unease head-on — the fix isn’t spending less but getting more tokens from the same power, which is the only version of the argument that holds up over time. The 3% bump in the stock suggests the market is buying the narrative for now, helped along by the timing ahead of earnings. Two real tests lie ahead: this week’s financials, and then 2028, when a six-to-ten-fold claim has to arrive as silicon rather than a slide.
What It Means for Developers
Three points. First, a 2028 timeline means near-term choices can’t wait around for it — TPU and the existing GPU routes stay the mainline for the next two years, and roadmap placeholders seldom survive contact with production planning. Second, an inference-efficiency race ends in cheaper tokens, so the long-run cost curve of Gemini-family APIs deserves a place in any pricing comparison, right next to the current prices everyone quotes. Third, should “custom silicon for one model family” become the norm, models and infrastructure grow more tightly bound, and moving between providers gets harder at the very moment vendors push platform-specific features. A multi-vendor strategy — or at minimum an abstraction layer over your model calls — gains value rather than losing it.