Google Is Reportedly Building a Chip With Gemini's Design Baked Into the Silicon
A server chip code-named Frozen v2 would hardwire the shape of Gemini into hardware. Engineers reportedly project six to ten times more output per unit of power. Google has not confirmed any of it.
Google is reportedly working on a new server chip, internally code-named Frozen v2, that would build parts of the Gemini model’s design directly into the hardware. According to reporting picked up this week, engineers project it could deliver six to ten times more AI output per unit of power than Google’s current TPUs. Deployment is said to be targeted for 2028. Google has not confirmed that the project exists, and the efficiency figure is an internal projection rather than a measured result, so treat the number as a hope rather than a spec.
The interesting part is the word “architecture.” A model has two things: its architecture, meaning the blueprint of how the layers and connections are arranged, and its weights, meaning the millions of tuned numbers that make it know things. Frozen v2 reportedly hardwires the blueprint but not the numbers. That matters, because it means Google could still load a newer, smarter Gemini onto the same chip later. Freeze the weights too and the chip becomes obsolete the day a better model ships. The reported design is a compromise between speed and staying useful.
The efficiency comes from doing less shuffling. In a normal chip, running a model means constantly moving data back and forth between memory and the processing units, and that movement, rather than the maths itself, eats a surprising share of the power. If the chip already knows the shape of the computation it will be asked to do, a lot of that shuffling disappears.
What’s behind it: This is a bet that model architectures are starting to settle down. For years the design of frontier models changed too fast for anyone to commit it to silicon, because chips take roughly three to four years from design to deployment. Betting on 2028 hardware means betting that a Gemini-shaped model will still look broadly Gemini-shaped in two years. The business reason is more immediate: Google is reportedly short on internal compute, to the point where its cloud arm cannot serve every enterprise customer that wants in. Cheaper serving is worth as much as a better model right now, and possibly more.
What this means for you: Nothing today, and probably nothing in 2027 either. But the direction is worth knowing about, because the cost of running AI is what determines what you get for free. Every big drop in serving cost over the past two years has shown up as either a price cut or a more generous free tier within months. If specialised chips like this work, the practical result for most people is that the good models stop being rationed. The caveat worth holding onto: unconfirmed projections from unnamed engineers about hardware two years out are the least reliable category of AI news there is. A six to ten times range is wide enough to include “transformative” and “modest” in the same sentence.
Sources
Anthropic Is Giving Away $50,000 in Credits to Rare Disease Researchers, and the Deadline Is 2 August
A focused call for applications under Anthropic's AI for Science programme, aimed at the roughly 400 million people living with one of more than 7,000 rare diseases. Two tracks, open until 2 August.
