Google reportedly building chip with Gemini baked into silicon

What happened: The Information reported on 20 July 2026 that Google is internally developing a server chip called "Frozen v2", built to run Gemini models more efficiently. Google has not confirmed the report.

The details: The chip permanently embeds parts of Gemini's architecture into the silicon itself, cutting the calculations and data movement needed to answer a query. Per The Information, it "embeds model architecture instead of weights". The name reflects the idea of "freezing" model parameters.

The context: The concept reportedly came from Google's Jeff Dean. His original "Frozen" design baked the model weights themselves into the chip, but Google dropped it because it would only work with one version of Gemini and go stale too quickly.

The numbers: Google's engineering estimate puts the gain at six to ten times more tokens served per unit of power than its newest TPUs. The design is still being worked on, so the figure could shift.

Why it matters: A compute shortage has strained Google's internal teams and forced Google Cloud to turn away some business. Cheaper serving could let Google undercut OpenAI and Anthropic on price.

The catch: Frozen v2 only keeps its efficiency edge if future Gemini models share the same underlying architecture. A major architecture change could make the chip obsolete.

Who's affected: Alphabet shares closed 1.51% higher the Monday the report landed, according to CNBC.

What's next: Google plans to trial Frozen v2 from 2028, in smaller volumes than its general-purpose TPU line, which now splits into training-focused 8t and inference-focused 8i chips announced at Cloud Next. It is meant to complement, not replace, those TPUs.

Sources: CNBC, The Decoder