Google Develops New AI Chip to Improve Gemini Efficiency
Google's Frozen v2 chip targets a 6-to-10x efficiency gain over its current AI chips, a leap that could lower operating costs for Gemini at a time when AI spending scrutiny is intensifying across the industry.
Key Facts
- Google is developing a new AI chip internally called Frozen v2, expected for release in 2028.
- The Frozen v2 chip could be 6 to 10 times more efficient than Google's existing AI chips in tokens generated per unit of power.
- OpenAI announced its own inference chip, Jalapeño, co-developed with Broadcom, in June 2026.
- Anthropic is reportedly in discussions with Samsung to develop its own custom AI chip.
- Google's existing AI chip lineup includes the 8th generation TPU, with training-focused TPU 8t and inference-focused TPU 8i.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
Google is developing a new proprietary server chip, internally called "Frozen v2," designed to improve the operational efficiency of its Gemini AI servers. According to a report from The Information, the chip is expected to be released sometime in 2028 and could be 6 to 10 times more efficient than Google's existing AI chips when measuring tokens generated per unit of power. TechCrunch asked Google about the report, and a Google spokesperson responded that the company is always researching new innovations but noted that not every project becomes a product. The development comes as AI companies race to build custom chips to address a global shortage of AI computing capacity and rising spending concerns. OpenAI announced its own "Jalapeño" inference chip in June 2026, and Anthropic is reportedly in discussions with Samsung to develop a custom AI chip. Google's existing AI chip lineup includes the 8th generation TPU, with training-focused TPU 8t and inference-focused TPU 8i. The Frozen v2 chip is part of a broader trend where efficiency has become a key selling point for technology companies as market enthusiasm for AI cools.
Google's full-stack approach means it designs hardware and software together, a strategy the company says ensures systems are "highly optimized for real-world workloads." The Frozen v2 chip is not the only custom silicon project among AI companies. OpenAI announced its "Jalapeño" inference chip, developed with Broadcom, at the end of June 2026. Anthropic is in discussions with Samsung to develop its own custom AI chip, a report from early July 2026 noted.
TechCrunch asked Google about the Frozen v2 report. A Google spokesperson responded: "Our teams are always researching and experimenting with new innovations to deliver the best performance and efficiency to our users and customers. Not every project becomes a product, but this thorough exploration is core to our full-stack approach."
The Information's sources say Frozen v2 is expected to launch in 2028. Google's existing AI chip lineup includes the 8th generation TPU, with the training-focused TPU 8t and the inference-focused TPU 8i. The industry-wide push for efficiency comes as market enthusiasm for AI cools and concerns over AI-related spending grow, TechCrunch noted.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.