← Back

Wire · technology

Google plans new chip to run Gemini models more efficiently, the Information reports

Published

20 July 2026

Topic

technology

Sectors

AI InfrastructureSemiconductors

Geography

United States

Source

Read at reuters.com

Verified

Fusion42 · 20 July 2026 · Fusion42 review

Google is developing a custom server chip called 'Frozen v2' that embeds Gemini model elements directly into hardware, targeting 2028 deployment with 6-10x better efficiency than current TPUs to address internal AI capacity constraints.

This Wire brief sits within Fusion42's coverage of AI Infrastructure and Semiconductors, and 5 sources have reported it between 20 Jul 2026 and 22 Jul 2026.

◆ The Wire takeaway

If you're selling inference optimisation, cooling, power delivery, or tensor software to AI labs, Google just signalled it's building these capabilities in-house by 2028—your buyer is becoming your competitor. Start mapping what Google can't bake into silicon and sell into their supply chain before they do.

Coverage

5 sources · first reported 20 Jul 2026 · latest 22 Jul 2026

Related on Wire

Topics

AI InfrastructureSemiconductorscustom-siliconinference-optimizationgemini-hardwiringcapacity-crunch2028-deployment