Google's 'Frozen v2' Chip Joins a Crowded Custom Silicon Race
Google is developing an internal server chip called Frozen v2, targeting a 2028 release. The chip is designed specifically to run Gemini models more efficiently, with projected efficiency gains of 6 to 10 times over Google's current AI chips, measured in tokens generated per watt.
The Numbers
6 to 10x efficiency is a wide range. That spread suggests the figure is still a projection, not a measured result. Even at the low end, it would be a meaningful improvement given Google's planned AI infrastructure spend of $180 to $190 billion.
Tokens per watt is the right metric to watch. As inference scales, power costs become the dominant variable. A chip that generates more tokens per unit of power is worth more than a chip that's simply faster.
Not Alone
Google isn't the only lab going custom. OpenAI announced its first in-house chip in June 2026, an inference processor called Jalapeño. Anthropic was reported to be in discussions with Samsung about a chipmaking partnership earlier this month.
Three years ago, the major labs were effectively NVIDIA customers. Now all three are pursuing some form of custom silicon. The pace of that shift is notable.
What This Might Mean
Vertical integration in chips typically trades flexibility for cost and optimization. A chip designed exclusively for Gemini can make assumptions a general-purpose GPU cannot. This could mean better efficiency at the cost of being useless for anything else.
One possibility is that Frozen v2 is designed around Gemini's specific attention patterns or quantization schemes. That would explain the efficiency projection. It would also mean the chip becomes a liability if the model architecture changes significantly before 2028.
Timeline
2028 is two years out. A lot can change. OpenAI's Jalapeño is only recently announced. Anthropic's Samsung partnership is still in discussion. The chip race among AI labs is real, but it's early.
What's confirmed: Google has a 2028 target, a 6-10x efficiency claim, and a nine-figure infrastructure budget behind it. Everything else is still speculative.
Source: Techcrunch