Breaking

Google Said to Be Developing 'Frozen v2' Chip That Etches Gemini Into Silicon for Efficient Inference

July 20, 2026 at 10:49 EDT

Infra & Chips
Foundation Models

Google is developing a specialized inference chip, dubbed "Frozen v2," that would hardwire the architecture and weights of its Gemini models directly into silicon, according to a report from The Information published July 20, 2026. The effort, said to be led by Google DeepMind, aims to slash the cost and energy footprint of running Gemini at scale as demand for AI compute strains the company's infrastructure.

July 20, 2026 · Google DeepMind

Google's "Frozen v2": Baking Gemini Straight Into Silicon

A reported DeepMind effort would hardwire Gemini's architecture and weights into a purpose-built inference chip — trading flexibility for a dramatic cut in the cost and energy of running the model at scale.

Efficiency claim

Up to 6–10× more tokens per watt vs. Google's latest TPUs

1×

Latest TPU
(tokens per watt)

6–10×

Frozen v2
(claimed, unverified)

Freezing the architecture strips programmable flexibility to slash data movement.

~2028

Reported deployment target

~$1B

Per-month capacity Google reportedly added externally

2016

Year Google began in-house TPU design

The bet

DeepMind is said to believe the transformer architecture has matured enough to bake permanently into silicon — a fixed-function ASIC tuned to one largely frozen model.

The risk

A frozen design risks obsolescence if model architectures shift. Frozen v2 is framed as an addition — not a replacement — for general-purpose TPUs and GPUs.

Still unconfirmed

• Pricing

• Concrete benchmarks

• External Cloud availability

• Large-scale clustering plans

Google has not publicly confirmed the effort; no benchmarks or hands-on evaluations exist yet.

Continue reading

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