Breaking

Google Said to Be Building 'Frozen v2' Chip That Bakes Gemini Logic Directly Into Silicon

July 20, 2026 at 11:01 EDT

Infra & Chips
Foundation Models

Google is developing a new AI server chip, codenamed "Frozen v2," that would permanently hardwire parts of its Gemini model into the silicon itself, a design aimed at dramatically improving inference efficiency as demand for AI computing continues to outstrip supply. According to a report from The Information, the effort is a separate track from the company's Tensor Processing Units and is targeted for deployment around 2028.

Targeted for 2028 · Google Silicon

Google's "Frozen v2" Would Bake Gemini Straight Into the Chip

An experimental AI server chip that permanently hardwires parts of a Gemini model into silicon — cutting the data shuttled between memory and compute to make inference far faster and cheaper to run.

6–10×

Power efficiency vs. the latest TPU, on a tokens-per-watt basis

~2028

Targeted deployment for the experimental project

2016

When Google began building its own custom TPUs

Tokens per watt

Inference efficiency: latest TPU vs. Frozen v2

1×

Latest TPU

up to 10×

Frozen v2

Fixing model logic into circuitry slashes memory-to-compute data movement — a primary bottleneck in large-scale AI inference.

How the efficiency compounds

Bake model into silicon

→

Less data movement

→

Lower power & cooling

→

More Cloud capacity

The upside

Faster inference at far lower operating cost
Reduced reliance on external chip suppliers
Complements — not replaces — the TPU lineup

The catch

Frozen logic means model updates may force a costly hardware redesign
Efficiency traded for flexibility
Early, exploratory stage — no benchmarks or pricing disclosed

Continue reading

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