Google is designing a new server chip, internally dubbed Frozen v2, that incorporates parts of its Gemini model architecture directly into silicon. The strategic goal is to make running these models drastically more efficient, reducing operational costs and energy consumption at datacenter scale.
Alphabet, Google's parent company, already has an extensive and mature portfolio of custom chips — the TPUs (Tensor Processing Units), now in their eighth generation and used for both training and inference of models. Frozen v2, however, represents a fundamentally different approach. Instead of being a general-purpose accelerator for diverse AI workloads, it is specifically designed to optimize the Gemini model architecture and its unique computational patterns.
The rationale behind Frozen v2 is simple in concept but extremely complex in execution: if the same company manufactures both the AI model and the chip that runs it, why not design the silicon to exactly fit the mathematical operations the model needs to perform? It is the same reasoning that led Apple to develop the Neural Engine for its on-device AI models, but now applied at datacenter scale.
The potential impact of this approach is significant. Gemini models consume enormous amounts of compute power — and consequently, electrical energy — during inference at scale. A chip custom-designed for them could reduce per-query costs by orders of magnitude, allowing Google to offer more competitive pricing in the AI API market or, alternatively, run larger and more capable models within the same energy budget.
The trend that Frozen v2 crystallizes is clear: as competition in the AI market intensifies, competitive differentiation is shifting from software to hardware. Companies that control the full stack — from specialized chip to optimized model — gain structural cost advantages that competitors reliant on third-party hardware (such as NVIDIA) simply cannot replicate. Frozen v2 is Google's latest bet in this direction, and its success or failure will say much about the future of AI architecture in the company's datacenters.
This vertical integration strategy mirrors what other tech giants are doing. Amazon builds its own Trainium and Inferentia chips. Microsoft is reportedly exploring custom silicon. The race is no longer just about who has the best model — it is about who can run that model at the lowest cost per token.
Source: TechCrunch
