# Google's silicon bet could cut Gemini serving costs six-fold by 2028

> Google is reportedly building "Frozen v2", a specialised chip for cheaper Gemini inference by 2028.

*A chip reportedly baking Gemini's architecture into silicon aims to ease Google's compute crunch and undercut OpenAI and Anthropic on price.*

By Bez · InsidersFeed
Canonical: https://insidersfeed.com/news/google-s-silicon-bet-could-cut-gemini-serving-costs-six-fold-by-2028

Alphabet shares rose 1.51% on Monday after The Information reported, citing two people familiar with the matter, that Google is developing a server chip internally called "Frozen v2" built to run Gemini models far more cheaply than its current fleet of TPUs. Google has not confirmed the project.

The chip reportedly embeds parts of Gemini's architecture permanently into the silicon rather than the model's weights, cutting the calculations and data movement needed to answer a query. The Information described it as a design that "embeds model architecture instead of weights". Google's engineers estimate it could serve six to ten times more tokens per unit of power than the newest TPUs, though the design is still being worked on and the figure is an engineering estimate, not a confirmed spec.

The project traces to an earlier idea from Jeff Dean, whose first version tried to bake the model's actual weights into the chip. Google scrapped that approach because it would only work with one version of Gemini and go out of date too fast. Frozen v2 is built to survive across Gemini versions as long as the underlying architecture does not change substantially; a major architectural shift would risk making the chip obsolete.

Google plans to deploy 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. Frozen v2 is a complement to that line, not a replacement.

The timing points to a capacity problem. A compute shortage has strained internal teams and pushed Google Cloud to turn away some business. Cheaper inference could relieve that pressure while giving Google room to undercut [OpenAI and Anthropic on price](https://www.cnbc.com/2026/07/20/alphabet-googl-stock-ai-chip-report.html).

## The Insider view

If the efficiency gains hold, Google gets a lever few rivals can match: a chip tuned to its own model family rather than general-purpose silicon, which could let it price inference below what OpenAI and Anthropic pay for equivalent compute on merchant GPUs. That would sharpen the margin gap in a market already fighting on price per token.

The exposure is architectural lock-in. Baking Gemini's structure into silicon pays off only if Google holds that architecture steady through 2028 and beyond; any redesign of Gemini's core could strand the investment, as happened with Dean's [first version of the chip](https://the-decoder.com/googles-frozen-v2-chip-reportedly-bakes-geminis-architecture-directly-into-silicon-for-efficiency-gains/). Watch whether Google commits publicly to the project once it clears internal review, and whether Cloud's turned-away business returns once Frozen v2 ships.

## Key takeaways

- Frozen v2 could serve 6-10x more tokens per unit of power than Google's newest TPUs
- Alphabet shares rose 1.51% after the report, though Google has not confirmed the project
- Cheaper serving could let Google undercut OpenAI and Anthropic on price from 2028

## Sources

- [Alphabet stock pops on report it's developing a more efficient AI chip](https://www.cnbc.com/2026/07/20/alphabet-googl-stock-ai-chip-report.html) — CNBC, 2026-07-20
- [Google's "Frozen v2" chip reportedly bakes Gemini's architecture directly into silicon for efficiency gains](https://the-decoder.com/googles-frozen-v2-chip-reportedly-bakes-geminis-architecture-directly-into-silicon-for-efficiency-gains/) — The Decoder, 2026-07-21