- Published on
Google Is Building an AI Chip Just for Gemini—And Investors Already Moved On It
- Authors
- Name
- DeepHorus
Google is developing a specialized artificial intelligence chip, codenamed Frozen v2, designed exclusively to run its Gemini model with significantly greater efficiency. The chip, reported by The Information on Monday, represents a strategic move to address Google’s growing capacity constraints as demand for AI computing outstrips its server supply. By hardwiring part of Gemini's architecture into silicon, the chip could enable six to ten times the efficiency of current Tensor Processing Units (TPUs), according to engineering projections.
Google has been struggling to meet AI compute demand despite massive infrastructure investments. In March, the company informed Meta that it could not fulfill the volume of Gemini compute Meta wanted to purchase, forcing Meta to ration AI usage among its employees. Google is spending up to $190 billion on AI infrastructure this year, yet it still had to turn away customers due to insufficient server capacity. Frozen v2 aims to alleviate this bottleneck by making each query dramatically cheaper to run.
Unlike Google’s existing TPUs, which can handle any AI model, Frozen v2 bakes part of Gemini’s architecture—its structural blueprint—directly into the hardware. This design, referred to as “freezing” in machine learning, locks the architecture in place while leaving the model’s updatable weights untouched. By skipping redundant calculations and reducing data shuffling across memory, the chip is projected to improve tokens generated per watt by six to ten times. That efficiency gain could allow Google to serve ten queries for the power cost of one.
The implications for Google’s competitive position are significant. A cheaper-to-run Gemini could help the company compete against OpenAI, Anthropic, and Chinese AI labs, which already account for up to 45% of US companies’ AI token usage, largely due to being 60–90% cheaper. While users may not notice a difference in Gemini’s performance, Google stands to become more profitable per query. The chip also reduces Google’s reliance on Nvidia, which controls roughly 85% of the GPU market for AI—a dependence that other tech giants like Meta, Amazon, and Microsoft are also seeking to break.
Investors reacted positively to the news, with Alphabet shares climbing about 3% during Monday’s trading session to reach $356 intraday. However, the gain partially receded as the market awaits Google’s Q2 2026 earnings report, due Wednesday, July 22. The chip represents another step in the industry-wide effort to develop custom silicon that bypasses Nvidia’s dominant but overhead-heavy GPUs, originally designed for video games rather than language models.
Frozen v2 remains in an exploratory phase, with key design decisions not yet finalized and Google declining to confirm the project’s existence. The chip is not intended for external cloud customers, as its specialized design limits it to running only Gemini. Deployment is targeted for 2028 at the earliest, according to reports. In the interim, Google is paying SpaceX $920 million per month to rent 110,000 Nvidia GPUs from xAI’s data centers as a bridge solution, underscoring the urgency behind its quest for more efficient hardware.
