Alphabet is quietly developing a new server chip intended to make its Gemini artificial intelligence models far more efficient, underscoring how central custom silicon has become to the AI race.

The processor, known internally as Frozen v2, is being designed as a next-generation accelerator for running large language models in Google’s data centers. According to people familiar with the project cited in multiple reports, Frozen v2 could deliver between six and ten times the efficiency of Google’s current AI chips, measured in tokens generated per unit of power consumed. That metric is crucial for large-scale AI deployments, where electricity and cooling costs can quickly eclipse hardware spending.

Google has not confirmed specific details of Frozen v2 but has acknowledged that such work is underway. The company says its engineers are continually experimenting with new chip designs as part of a broader strategy to co-design hardware and software, allowing Gemini and other models to be tightly tuned to the underlying infrastructure.

Custom AI chips have become a strategic priority across the industry. Building in-house silicon allows companies to reduce dependence on Nvidia, whose GPUs remain the default choice for training and running advanced models but are expensive and often in short supply. By improving efficiency, companies can serve more users with the same data center footprint, a critical advantage as AI workloads surge.

Rivals are pursuing similar paths. OpenAI has announced its own custom inference chip, and Anthropic is exploring partnerships with major semiconductor manufacturers. These efforts reflect a broader shift: the next phase of AI competition is not only about model quality, but also about the economics of running those models at scale.

For Alphabet, the stakes are particularly high. The company has signaled plans for enormous capital expenditures to support its AI ambitions, and investors have pressed for evidence that such spending will translate into sustainable margins. Reports of Frozen v2’s potential efficiency gains have been welcomed on Wall Street, where any sign that Google can lower the cost of serving Gemini is seen as a meaningful catalyst.

If Frozen v2 delivers on its promise, it could become a cornerstone of Google’s AI infrastructure, enabling more powerful versions of Gemini while keeping energy use and operating costs in check.