Google is designing a new server chip that will embed Gemini AI models directly into the hardware, according to reporting from July 20, 2026. The chip, informally dubbed “Frozen v2,” aims to serve AI models more efficiently and address an internal computing capacity crunch that has strained Google’s infrastructure.

The company expects Frozen v2 to be between six and 10 times more efficient than Google’s existing AI chips, measured by tokens generated per unit of power. The new chip is slated for release around 2028. The timeline suggests Google wants to solve long-term capacity issues, not patch quarterly problems.

Capacity Constraints and Internal Tension

Google’s AI infrastructure is under real pressure. The company has faced criticism from its own cloud division for declining deals with outside customers because internal Gemini projects consumed available capacity. That’s a rare problem for a company with Google’s balance sheet and engineering talent, but it’s telling. High-margin AI compute is in shortage.

Designing a chip specifically to reduce that pressure signals conviction that AI workloads won’t plateau. Chip design cycles take years, and if Google is committing engineering resources to Frozen v2 now, management believes the demand will be there when the chip ships in 2028 and beyond.

Custom Silicon as Competitive Moat

Amazon and Microsoft are also designing custom chips for AI inference and training. Building silicon in-house reduces dependency on Nvidia and increases margins. Google’s move is defensive and offensive: it defends against Nvidia supply shocks and lets Google undercut cloud competitors on price.

The future of cloud computing belongs to companies that can design their own silicon. Frozen v2 is Google’s bet that custom chips matter more than general-purpose processors.