Google developing Frozen v2 chip for Gemini AI
In brief
- Frozen v2 is Google's custom chip designed to embed Gemini's architecture directly into hardware for optimized performance.
- The chip projects 6–10x improvement in tokens generated per watt compared to standard TPUs.
- Deployment targeted for 2028; Google has not officially confirmed the project.
- Custom silicon development reflects industry-wide push by Meta, Amazon, Microsoft, and OpenAI to reduce Nvidia dependency.
How Frozen v2 Works
The chip bakes Gemini's architecture into hardware rather than treating the model as software running on general-purpose silicon. What gets frozen is the architecture—the structural blueprint that defines how Gemini processes information—while the model's weights (the actual knowledge Gemini picks up through training) remain updatable. This approach lets the chip skip redundant calculations and stops shuttling data across memory on every query.
Engineers project a six to ten times improvement in tokens—the small text chunks that make up each AI response—generated per watt of electricity consumed. These figures are based on preliminary engineering projections and assume successful completion of the design; actual performance will depend on silicon testing and may differ significantly.
Why Google Needs Custom Silicon
Google told Meta in March it couldn't fill the volume of Gemini compute Meta wanted to purchase. Meta had to instruct employees to ration their AI usage. Meanwhile, Google is spending up to $190 billion on AI infrastructure this year and paying SpaceX $920 million a month to rent 110,000 Nvidia GPUs from xAI's data centers. The company's appetite for compute is outpacing supply.
Nvidia controls roughly 85% of the GPU market for AI. That concentration creates both cost and availability risk. OpenAI, Anthropic, and Chinese labs already account for up to 45% of U.S. company AI token usage, largely because they run 60–90% cheaper. Custom silicon that cuts power consumption by an order of magnitude would help Google compete on price and availability.
Strategic Limits and Timeline
Frozen v2 won't be offered to outside Cloud customers because hardware hardwired for one model can't run anyone else's. This is the core tradeoff: efficiency for specialization. Deployment is targeted for 2028 at the earliest.
Reality Check: Exploratory Status
Frozen v2 is still exploratory, with key design decisions not yet finalized. Custom silicon projects routinely miss timelines—delays in tape-out, yield issues in manufacturing, or unforeseen design flaws can push deployment by years. Meta, Amazon, Microsoft, and OpenAI all have custom silicon programs, but translating lab prototypes into production silicon is notoriously difficult. AWS committed to deploying 1 million Nvidia GPUs through 2027 while simultaneously building its own chips—a hedge that suggests even these companies aren't betting entirely on custom silicon yet.
Market Reaction
Alphabet shares climbed roughly 3% during Monday's session on the news, touching $356 intraday. Investors appear to have already priced in the efficiency gains, at least on the surface. Whether the project delivers on its 2028 timeline—or delivers at all—remains an open question.
