Google's 'Frozen v2' chip aims for 10x Gemini efficiency

Running cutting-edge AI models usually means buying more generic GPUs and tweaking software. Google is taking a much more radical route: literally etching its model architecture directly into silicon.

Alphabet is developing a new in-house server chip dubbed "Frozen v2," aimed for release by 2028, according to a report from The Information. The chip could deliver a massive 6x to 10x efficiency leap over Google's current TPUs, measured in tokens generated per unit of electricity.

Hardcoding Gemini into silicon

The "Frozen" moniker is quite literal. By permanently baking parts of the Gemini model architecture into the silicon, the processor skips time-consuming computational decisions and minimizes internal data transfer bottlenecks.

Essentially, Google is sacrificing flexibility for raw, dedicated efficiency.

When asked for comment, Google neither confirmed nor denied the project directly. The company told TechCrunch that while not every experimental project enters production, hardware-software co-design remains central to its strategy to optimize real-world AI workloads.

Driven by compute shortages

This aggressive push comes amid severe infrastructure constraints. Google's AI compute shortage has grown acute enough that Google Cloud has been forced to turn away prospective customers.

Cost is another massive driver. Google previously announced planned expenditures between $180 billion and $190 billion this year, sparking market anxiety over whether massive AI spending will ever pay off. Word of the Frozen v2 project helped calm nervous investors, pushing Alphabet's stock up roughly 3% on Monday morning.

Google isn't alone in racing away from Nvidia dependence. OpenAI unveiled its custom "Jalapeño" inference chip in June, and Anthropic was recently reported to be exploring a silicon partnership with Samsung.

Why it matters

We see Frozen v2 as an enormous gamble on the architectural longevity of Google's flagship model.

On paper, the upside is huge. A 6x to 10x leap in energy efficiency would dramatically slash unit inference costs, giving Google a formidable pricing and margin buffer as it integrates Gemini deeper into consumer and enterprise software. If compute power remains scarce, extreme efficiency is king.

However, baking neural architecture into custom silicon is notoriously risky. AI models evolve at breakneck speed, and silicon design cycles take years. If Transformer architectures shift drastically before 2028, these chips could end up obsolete before they even deploy. It looks like Google is betting everything that Gemini's core design will remain relevant for years to come.