Breaking

Google Reportedly Building 'Frozen v2' Chip to Optimize Gemini AI Inference, Lifting Alphabet Shares

July 20, 2026 at 13:31 EDT

Infra & Chips
Foundation Models

Alphabet shares climbed as much as roughly 3% on Monday after a report that Google is developing a new server chip, dubbed "Frozen v2," purpose-built to run its Gemini AI models more efficiently. The design would embed elements of Gemini's model architecture directly into silicon, a step aimed at sharply reducing the compute and energy cost of generating AI responses.

Chip Report · Alphabet

Google's 'Frozen v2' — silicon built to run Gemini cheaper

A report that Google is designing a server chip embedding Gemini's architecture directly into silicon lifted Alphabet shares as much as ~3% — with a striking efficiency goal for AI inference.

+3%

Intraday climb in Alphabet shares on the report

6–10×

Target tokens per unit of power vs latest TPU

2028

Targeted deployment; design still being finalized

Tokens processed per unit of power

Reported efficiency goal for Frozen v2 vs Google's latest TPU (per-watt throughput)

1×

Latest TPU

6–10×

Frozen v2 (goal)

Range shown as 6× floor; upper bound reaches ~10×. Unconfirmed by Google.

HOW IT SAVES COMPUTE

Embed Gemini architecture in silicon

→

Cut compute steps & data movement

→

Cheaper, lower-power inference at scale

A complement to existing TPUs — not a replacement — reported as more of a trial run than a full production line, easing data-center expansion pressure.

The upside

Deeper vertical integration of models, software and hardware — a strategy that could cut costs, ease compute shortages and reduce reliance on Nvidia GPUs.

The catch

Hardwiring a model into fixed silicon sacrifices flexibility. Ambitious efficiency figures and a 2028 timeline leave ample room to shift before production.

Continue reading

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