Breaking

Google Said to Be Developing 'Frozen V2' Chip to Run Gemini Far More Efficiently Than Current TPUs

July 20, 2026 at 10:01 EDT

Infra & Chips
Foundation Models

Google is reportedly developing a new custom chip, referred to as "Frozen V2," that could run its Gemini models 6 to 10 times more efficiently than the company's existing Tensor Processing Units, according to a report from The Information.

Custom Silicon · Google

A New Chip Built to Run Gemini 6–10× More Efficiently

Google is reportedly developing a custom accelerator, "Frozen V2," said to run its Gemini models far more efficiently than its current Tensor Processing Units — a purpose-built bid to cut the cost of serving large models at scale.

6–10×

Reported efficiency gain over existing TPUs

Gemini

Model family the chip is tuned to accelerate

Codenamed "Frozen V2" — unconfirmed, in development

Efficiency Per Chip

How much work the same silicon could do

1×

Current TPU
(baseline)

up to 10×

"Frozen V2"
(reported)

Range reflects the reported 6–10× span; upper bound shown. Early claims may rest on narrow benchmarks and idealized conditions.

Why it matters

A chip designed around Gemini's exact computations could cut energy use and operating cost, boost throughput, and strengthen Google against rivals leaning on merchant GPUs.

The caveats

No confirmed timeline, no disclosed benchmarks, and no official word from Google. Figures are preliminary and may not hold across every real-world workload.

Control over custom silicon is now a strategic advantage.

Lowering the cost of running Gemini would give Google more room to scale and compete on price and performance — but the real impact depends on results once the hardware moves beyond development.

Continue reading

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