Google develops custom AI chip to boost efficiency of Gemini models
Alphabet is reportedly developing a custom server chip, codenamed "Frozen v2," intended to improve the power efficiency of its Gemini models by 2028.
Alphabet is reportedly developing a new server chip, internally codenamed "Frozen v2," aimed at improving the efficiency of its proprietary Gemini models. According to a report by The Information, the hardware is slated for release in 2028. The chip is projected to be between six and 10 times more efficient than Google’s current AI chips, with efficiency measured by the volume of tokens generated per unit of power. While the company did not confirm the specific project, a spokesperson stated that Google is "constantly researching and experimenting with new innovations" as part of its full-stack approach to hardware and software co-design.
This development arrives as major AI firms accelerate efforts to build custom silicon, largely to mitigate dependence on Nvidia, which has historically dominated the AI chip market. Beyond reducing reliance on external suppliers, companies are seeking to address global shortages in computing capacity. The industry shift toward proprietary hardware also reflects a cooling of market euphoria, where efficiency metrics are becoming increasingly critical to justifying the substantial capital expenditures required to maintain competitive AI infrastructure.
Alphabet is under significant scrutiny regarding its AI spending strategy. Earlier this year, the company disclosed plans to spend between $180 billion and $190 billion, a figure that has drawn investor concern. Demonstrating that these investments will yield tangible operational improvements is essential for the company as it navigates high-cost development cycles. Following the report on the "Frozen v2" project, Alphabet’s stock rose approximately 3% on Monday morning, suggesting that investors are viewing potential hardware efficiency gains as a positive indicator for future profitability.
The move to develop custom silicon is part of a broader industry trend. In June, OpenAI announced its own inference processor, "Jalapeño," and reports from earlier this month indicated that Anthropic is exploring a chipmaking partnership with Samsung. These initiatives highlight the intensifying competition among tech giants to control both the software and hardware layers of their AI stacks. By optimizing hardware specifically for the computational demands of their models, these companies hope to secure a performance advantage that off-the-shelf components cannot provide.