.jpg?w=1024)
Google is developing a new server chip designed specifically to make its Gemini AI models run more efficiently, according to reports from The Information cited by TechCrunch.
The chip, internally known as Frozen v2, is expected to arrive as early as 2028, although engineers are still finalizing its design.
Rather than replacing Google's existing tensor processing units (TPUs), the new processor would serve as a specialized addition to the company's growing portfolio of custom AI chips.
According to the reports, Frozen v2 would embed parts of Gemini's architecture directly into the chip's hardware. That approach could reduce the amount of data movement and computation needed to generate responses, improving speed while lowering power consumption.
Google has not confirmed the project but has also not disputed the reports.
"Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers," a Google spokesperson told TechCrunch. "While not every project moves into production, this rigorous exploration is central to our full stack approach. By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads."
Aiming for a major efficiency jump
Internal projections reportedly suggest Frozen v2 could deliver six to 10 times more AI tokens per unit of power than Google's latest AI chips. That figure remains a design target rather than a tested performance result, and Google has not committed to bringing the chip into production.
The project is also intended to coexist with TPUs rather than replace them, allowing Google to continue using general-purpose AI accelerators while deploying highly optimized hardware for Gemini workloads.
Part of a broader AI chip race
Google is far from alone in pursuing custom AI hardware.
Major AI developers are increasingly designing their own chips to improve efficiency, reduce infrastructure costs, and lessen dependence on Nvidia, whose processors still dominate the AI market.
OpenAI recently unveiled its first custom inference chip, while reports indicate Anthropic has explored a chip partnership with Samsung. Amazon is also expanding its custom silicon efforts for AI infrastructure. For Google, the effort comes as demand for AI computing continues to strain capacity. Reuters reported that computing shortages have contributed to internal resource pressures and have even forced Google Cloud to turn away some outside business.
Lower power costs could reshape Gemini infrastructure
Frozen v2 highlights how the AI competition is shifting beyond building larger models to making them cheaper and faster to operate.
If Google can significantly lower the energy needed to serve Gemini across products such as Search, Workspace, Cloud, and other services, it could reduce long-term operating costs while supporting more users without proportionally expanding its infrastructure.
The strategy also carries tradeoffs. A chip built around Gemini's architecture may be less flexible if Google's future AI models change significantly, limiting its usefulness compared with more general-purpose processors. The project also remains several years away, meaning its final design and performance could still change before launch.
Looking ahead
The reported project arrives as investors closely watch Google's massive AI spending plans, with the company expecting capital expenditures of $180 billion to $190 billion this year. News of Frozen v2 appeared to reassure the market, helping Alphabet shares climb about 3% on Monday following the report.
Even if Frozen v2 never reaches production in its current form, the effort signals Google's willingness to tailor hardware specifically for Gemini as competition in AI infrastructure continues to intensify.
Also read: Google’s AI Mode connected apps update shows how Gemini is becoming more deeply tied to Search, Gmail, Drive, and Workspace.