Google is building a chip tailor-made for Gemini — and the stock popped 3% on the news
Google is developing Frozen v2, a chip that etches part of Gemini into silicon and could serve 6 to 10 times more tokens per unit of power than its latest TPUs.
Google is designing a chip with a single purpose in life: run Gemini faster on far less power. It is called Frozen v2, and the idea is bold — instead of a general-purpose processor, part of the model’s architecture gets permanently etched into the silicon, cutting the calculations and data shuffling behind every answer. The report alone was enough to send Alphabet shares up around 3% on Monday.
What is Google’s Frozen v2 chip?
It is a server chip built exclusively to serve Gemini models, pencilled in for 2028. Internal projections suggest it could serve six to ten times more tokens per unit of power than Google’s newest Tensor Processing Units, the in-house chips that have powered its data centres for years. At a time when AI’s electricity bill is growing faster than its revenue, efficiency is the magic word every investor wants to hear.
There is a catch, and it is not small: a chip with the model “frozen” into silicon only works for future models if Google keeps Gemini’s foundations unchanged. It is a bet on the stability of its own technology — risky in an industry that swaps paradigms every eighteen months.
The timing tells its own story too. The company reports earnings this week, just after delaying Gemini 3.5 Pro over failed tests, a stumble that wiped billions off its market value. A miracle chip on the horizon is exactly the kind of narrative that makes people forget a delay.
For Portugal and Europe the stakes are less abstract than they sound: more efficient data centres mean less strain on power grids that are already stretched. If Frozen v2 delivers even half of what it promises, it is AI’s energy consumption — not AI itself — that may finally slow down.
By Oliver Grant
Image: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahu… / Wikimedia Commons (CC BY 4.0)
