Google Develops Frozen v2 Chip to Boost Gemini AI Efficiency

Google is developing a specialized custom server chip, internally named Frozen v2, designed to embed elements of its Gemini AI models directly into the hardware. This hardware-level integration aims to deliver six to ten times more AI tokens per unit of power compared to the company's existing Tensor Processing Units (TPUs), which the new chips are intended to complement rather than replace.

The project seeks to mitigate an internal computing capacity crunch that has forced Google Cloud to reject certain customer deals and led Google to pay Space Exploration Technologies Corp. nearly $1 billion monthly for additional compute capacity. While Alphabet Inc. describes the initiative as a rigorous exploration of a full-stack approach, reports indicate the chip is currently a trial run that lacks the flexibility to support future Gemini architectures. Engineers expect to deploy the hardware as early as 2028.

These developments come as Google faces operational headwinds, including delayed Gemini Pro releases after failing internal performance goals in coding, researcher attrition, and intensifying competition from Chinese firms such as Alibaba and Moonshot AI. Following the report of the Frozen v2 chip, Alphabet Inc. shares rose 3% on Monday.