Google Is Building an AI Chip Just for Gemini—And Investors Already Moved On It
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google is developing a server chip codenamed Frozen v2 that hardwires part of Gemini's architecture directly into silicon, with engineers projecting a 6-10x improvement in tokens generated per watt of electricity versus current TPUs.
- Alphabet shares climbed roughly 3% to $356 intraday Monday on the report, with Q2 2026 earnings due Wednesday, July 22.
- Google told Meta in March it couldn't fill the volume of Gemini compute Meta wanted, forcing Meta to instruct employees to ration their AI usage.
- OpenAI, Anthropic, and Chinese labs already account for up to 45% of U.S. company AI token usage, running 60-90% cheaper than Google.
- Nvidia controls roughly 85% of the AI GPU market, and Meta, Amazon, Microsoft, and OpenAI all have custom silicon programs to escape that dependency.
- Frozen v2 won't be offered to outside Cloud customers because it's hardwired for Gemini only; deployment is targeted for 2028 at the earliest and Google hasn't confirmed the project exists.
- Google is paying SpaceX $920 million per month to rent 110,000 Nvidia GPUs from xAI's data centers as a bridge until its own custom silicon arrives.
Why it matters: The chip targets a capacity crunch Google couldn't solve with $190B in AI infrastructure spend — Meta already had to ration Gemini. A projected 6-10x efficiency gain would let Google undercut OpenAI, Anthropic, and Chinese labs running 60-90% cheaper, which already eat 45% of US enterprise AI token usage.
