Google Develops New AI Chip "Frozen v2" to Boost Gemini Efficiency

Alphabet’s Google has revealed that it is designing a new server chip to significantly improve the efficiency of its proprietary large language model (LLM), “Gemini.” Internally dubbed “Frozen v2,” this chip is expected to launch in 2028.

According to a report by The Information, citing anonymous sources, the chip could achieve a 6–10x improvement in token generation per watt (tokens/watt) compared to Google’s existing AI chips. This figure highlights Google’s ambition to make a dramatic leap in the critical performance metric of energy efficiency.

When approached by TechCrunch, Google neither confirmed nor denied the report. In a statement, the company said, “Our teams are continuously researching and experimenting with new innovations to deliver the best performance and efficiency for our users and customers. Not all projects transition to products, but this rigorous pursuit is at the core of our full-stack approach. By co-designing hardware and software from the ground up, we ensure systems are integrated and highly optimized for real-world workloads.”

The trend of AI companies developing their own chips has been gaining momentum across the industry. This is driven by the need to operate in-house models more efficiently and address the global shortage of AI computing capacity. Efficiency is becoming an increasingly important selling point for tech companies, especially as concerns about AI-related spending begin to temper market enthusiasm.

At the same time, companies are seeking to reduce their dependence on NVIDIA, which has long dominated the AI chip market. In June this year, OpenAI announced its first custom inference chip, dubbed “Jalapeño.” Earlier this month, Anthropic was reported to be in talks with Samsung about a new chip manufacturing partnership.

Investors, who have previously voiced concerns over Google’s massive planned capital expenditure on its AI strategy, appear to have welcomed the news. Earlier this year, Google disclosed plans to spend between $180 billion and $190 billion, putting pressure on the company to demonstrate returns on these investments.

News of the more efficient Frozen v2 chip seems to have eased investor anxieties, as Google’s stock rose about 3% on Monday morning ahead of its quarterly earnings announcement later this week.

Shifting Power Dynamics in the AI Chip Market

and Google’s Strategy

Google has already developed its “TPU (Tensor Processing Unit)” series for its data centers, leveraging them for AI inference and training. TPUs are particularly effective at handling Transformer models and are integral to Google’s search, advertising, and cloud services. However, NVIDIA’s GPUs, especially next-generation models like the H100 and B200, continue to dominate the AI chip market.

The claimed “6–10x efficiency improvement” of the Frozen v2 chip could have a transformative impact, far beyond a mere generational upgrade. In the realm of inference processing, where existing TPUs already compete with NVIDIA GPUs, such improvements could significantly lower the AI inference costs for Google’s cloud service, “Google Cloud,” enhancing its competitiveness in the market.

Moreover, Google’s focus on developing its own chips reflects its commitment to a “full-stack approach.” By co-designing hardware and software, the company can optimize performance and efficiency specifically for its Gemini model, achieving levels unattainable with general-purpose GPUs.

Impact on Investors and the Market

The 3% rise in Google’s stock price following the report indicates that the market views this development positively. In an era where AI-related capital investments are ballooning, investors are scrutinizing every company’s return on investment (ROI). If Google can achieve efficiency gains through its own chips, it would not only enable the company to offer AI services at lower costs but also help reduce energy expenses and streamline its data center expansion plans.

Google’s $180–190 billion capital expenditure plan, which includes massive investments in AI infrastructure, remains a focal point for investors. The successful commercialization of Frozen v2 could significantly improve the ROI on these investments.

Relations with Competitors

The announcements of OpenAI’s Jalapeño and Anthropic’s partnership with Samsung underscore a broader industry shift to reduce reliance on NVIDIA. While NVIDIA’s GPUs have become the de facto standard for AI training and inference, their rising costs and limited supply present challenges for the industry.

By developing their own chips, companies can not only reduce royalty payments to NVIDIA but also tailor their chips for specific workloads. However, chip development requires substantial R&D investment and long lead times, as evidenced by Frozen v2’s projected 2028 release date.

Integration with Google Research Initiatives

Google has been advancing technology across various fields, such as releasing the time-series forecasting model TimesFM 2.5 and expanding Google Wallet functionalities. The development of Frozen v2 is part of this holistic approach, integrating both hardware and software innovations.

Technical Challenges and Future Outlook

The promised 6–10x efficiency improvement of Frozen v2 raises questions about how well it will perform in real-world workloads. There is often a gap between benchmark efficiency and operational efficiency. Google’s strategy to design Frozen v2 specifically for Gemini models aims to close this gap and achieve near-theoretical maximum efficiency.

The 2028 release timeline for Frozen v2 also raises questions given the rapid pace of AI industry evolution. Changes in Gemini’s model architecture or shifts in overall market demands by then could impact the chip’s design.

Google has cautioned that “not all projects transition to products,” leaving the door open for adjustments. However, the company’s continued pursuit of in-house AI chip development signals a long-term strategy to strengthen its competitive edge.

Editorial Opinion

In the short term, the announcement of Frozen v2 has boosted confidence in Google’s AI strategy, contributing to the rise in its stock price. Over the next three to six months, Google may unveil more concrete results regarding Gemini’s efficiency improvements or provide updates on the next generation of TPUs. Competitors like OpenAI and Anthropic are also advancing their own custom chip initiatives, potentially accelerating a shift in the AI chip market. The apparent cracks in NVIDIA’s dominance mark a significant turning point for the industry as a whole.

From a long-term perspective, Google’s goal of dramatically improving energy efficiency with Frozen v2 could substantially reduce AI operational costs and environmental impact. Over the next one to three years, the industry will likely focus on whether the full-stack approach of co-designing hardware and software becomes a standard development model. Additionally, as more companies transition to custom chips, NVIDIA’s market dominance could weaken, leading to greater diversification and cost reductions across the AI ecosystem.

References

• “Google is working on a new AI chip designed to make Gemini more efficient”, by Lucas Ropek — TechCrunch AI, 2026-07-20T21:21:15.000Z (ARR)
• Source URL: https://techcrunch.com/2026/07/20/google-is-working-on-a-new-ai-chip-designed-to-make-gemini-more-efficient/