Story perspectives

Google Unveils Frozen v2 Chip, Boosts Gemini Efficiency 10×

7/20/2026

Summary

Google Launches Chip

Google is developing a server chip called Frozen v2 for its Gemini AI model.
Frozen v2 aims for six-to-ten-fold higher token-per-power efficiency than Google's AI processors and will ship to data centers in 2028.
The chip hardwires Gemini's architecture while allowing weight updates, addressing a compute shortage that forced Google Cloud to turn away external customers, and builds on Jeff Dean's earlier “Frozen” concept that would have baked weights into silicon.

Summary

Alphabet Shares Surge

Alphabet's shares rose up to 3.7% to $359.68 in early New York trading after the chip news.
OpenAI recently launched a custom AI chip with Broadcom, highlighting fierce hardware competition.
Amazon is negotiating to sell custom chips for other data centers.
Anthropic is discussing a processor with Samsung.
Google postponed the Gemini 3.5 Pro launch, originally slated for June, after missing internal coding goals.

History

Historical Context

The AI-chip race traces back to NVIDIA’s 2016 launch of the DGX-1 server, which paired eight Pascal GPUs and introduced NVLink’s 80 GB/s links, later upgraded to Volta V100 GPUs delivering 960 TFLOPS. Building on that momentum, Google now unveils the Frozen v2 chip for Gemini, while rivals OpenAI, Amazon and Anthropic pursue custom silicon.

1 of 3

Google Launches Chip

Google is developing a server chip called Frozen v2 for its Gemini AI model.
Frozen v2 aims for six-to-ten-fold higher token-per-power efficiency than Google's AI processors and will ship to data centers in 2028.
The chip hardwires Gemini's architecture while allowing weight updates, addressing a compute shortage that forced Google Cloud to turn away external customers, and builds on Jeff Dean's earlier “Frozen” concept that would have baked weights into silicon.

1 / 3