Codenamed Frozen v2, the chip is reported to be six to ten times more efficient than Google’s latest custom AI hardware and is intended to sit alongside rather than replace its existing TPUs
Google is working on a new server chip that would take an unusual approach to AI efficiency: rather than simply building faster or more powerful hardware, it would embed elements of the Gemini AI model directly into the chip itself. The project, codenamed Frozen v2, was reported by The Information and represents one of the more significant hardware bets Google has made in its ongoing effort to reduce the cost and energy consumption of running AI at scale.
What Makes Frozen v2 Different
The proposed chip is expected to be six to ten times more efficient than Google’s latest custom AI chips, measured in AI tokens served per unit of power. That efficiency gain would come from hardwiring certain model parameters into the silicon itself, allowing specific inference tasks to be handled at the hardware level rather than through general-purpose computation. Engineers are still finalising the design and the extent of information that will be embedded, and a deployment target of 2028 has been reported, though that timeline remains subject to the outcome of ongoing development work.
How It Fits With Google’s Existing Chip Strategy
Frozen v2 is characterized as a distinct line of chips and not a successor to Google’s current generation of TPUs that currently run its internal AI operations and its services across Google Cloud. The lines would do different work, Frozen being built to fill the role to assist with specific types of inference workloads
Why This Matters Now
The development comes at a moment when Google Cloud has faced capacity constraints serious enough to prevent it from taking on certain outside customers. A chip that dramatically improves tokens delivered per watt would ease that bottleneck while reducing the enormous energy costs associated with running large language models at Google’s scale.
A Google Cloud spokesperson confirmed the company’s ongoing co-design approach across hardware and software without commenting on specific projects.
