Google developing “Frozen v2” chip to boost Gemini AI efficiency

Google’s internal “Frozen v2” server chip, slated for a 2028 deployment, is designed to embed parts of Gemini’s architecture in silicon and could serve six to ten times more tokens per unit of power than the company’s latest TPUs [1].

Chip design and expected gains

The “Frozen v2” concept hard‑wires elements of Google’s Gemini models into the chip, cutting the number of calculations and data movements needed for inference [1]. Google engineers say this integration could deliver six to ten times the token‑per‑power efficiency of the newest TPU generation [1]. By contrast, existing TPUs remain general‑purpose accelerators that must repeatedly fetch model data from memory, a step the new silicon aims to eliminate [2]. The trade‑off is reduced flexibility: the chip would only support future Gemini models that retain the same underlying architecture [1].

Why Google needs a new silicon family

Google’s AI compute shortage has forced the cloud unit to reject external business and to pay SpaceX roughly $1 billion a month for additional capacity [1]. More efficient hardware could alleviate the need for massive data‑center expansion, lowering electricity and infrastructure costs while keeping Gemini services responsive [2]. The move also reflects a broader industry shift toward custom silicon that reduces reliance on third‑party chipmakers, a strategy Google hopes will give it an edge as generative‑AI demand surges [2].

Competitive context

Google’s rivals, including Moonshot AI and Alibaba, have recently launched new models that narrow the capability gap with Gemini [1]. Meanwhile, Chinese AI models now account for 45 % of U.S. company token usage, intensifying pressure on Google to both improve model performance and cut operating costs [1]. The “Frozen v2” chip, if it meets its efficiency targets, could help Google sustain its market share without the scale‑up costs that competitors are incurring.

What to watch

• 2028 rollout timeline – monitor any updates on design finalisation or early‑stage production.
• Gemini model updates – see whether future Gemini releases remain compatible with the hard‑wired architecture.
• Google Cloud capacity – watch for changes in the volume of external deals the cloud unit can accept as the chip matures.

If “Frozen v2” delivers the claimed efficiency gains, Google could substantially expand its AI serving capacity while curbing the capital and energy expenditures that currently limit its cloud growth. The open question remains whether the reduced flexibility will lock Google into a single model architecture, potentially hampering future AI innovation.