Breaking

Google Reported to Be Building 'Frozen v2' Chip Tailored to Run Gemini Models Far More Efficiently

July 20, 2026 at 21:27 EDT

Infra & Chips
Foundation Models

Google is developing an internal server chip, code-named "Frozen v2," designed specifically to run its Gemini AI models with dramatically lower power consumption, according to a report from The Information. The chip is said to be six to ten times more efficient than the company's current AI silicon, measured in tokens generated per watt, and is aimed at easing the mounting strain on Google's computing capacity.

AI Silicon · Custom Hardware

Google's 'Frozen v2': A Chip That Bakes Gemini Into Silicon

A reported internal server chip would hardwire parts of Gemini's architecture directly into circuitry — trading a general accelerator's flexibility for outsized gains in energy efficiency. Target: around 2028.

6–10×

More efficient than current AI silicon, in tokens per watt

~2028

Reported target; chip in final stages of design

+1.5–3.3%

Alphabet share gain in days after the report

Efficiency, drawn to scale — tokens per watt

1×

Current
AI chips

up to 10×

'Frozen v2'
(reported)

If realized, the same power delivers many times more output — and faster query responses.

The unusual design bet

Gemini model architecture

→

Hardwired into silicon

→

General overhead stripped away

Not a replacement for Google's TPUs — a complementary, purpose-built inference chip focused narrowly on serving Gemini. Not intended for TPU-scale mass production.

The promise

A sign of how aggressively Google is co-designing hardware around its own models — pointing toward tightly integrated chip-and-model stacks at the AI frontier.

The skepticism

A 2028 target does nothing for today's capacity crunch — with Cloud reportedly turning away customers — and internal doubts remain over whether it ever ships in volume.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.