Google is developing a new AI chip, Frozen v2, projected to significantly enhance Gemini's efficiency, leading to a 3% stock increase.
Summary
- Google is developing the Frozen v2 chip to enhance its Gemini AI system's performance.
- The new chip could deliver six to ten times the efficiency of current TPUs, with a launch expected in 2028.
- Alphabet's stock saw a 3% increase following the announcement, ahead of the Q2 2026 earnings report.
Google is actively working on a new chip, named Frozen v2, designed to optimize the functionality and cost-effectiveness of its Gemini AI system. This development comes as the company faces challenges in meeting the growing demand for its AI services.
Reported by The Information, Frozen v2 aims to address a critical shortage of computing capacity. Earlier this year, Google informed Meta that it could not provide the necessary computing resources for its Gemini project, leading to Meta advising employees to limit their AI usage. Despite investing approximately $190 billion in AI infrastructure this year, Google has had to turn away clients due to inadequate server capacity.
Frozen v2 is not merely an upgrade of Google's existing Tensor Processing Units (TPUs), which have been utilized since 2015. This new chip integrates aspects of Gemini's architecture directly into its design, optimizing how the model processes information.
The term "freezing" in this context refers to permanently securing a structure—in this case, the architecture itself—rather than the model's training weights, which can still be updated. By embedding this architecture into the chip, Google expects to eliminate unnecessary calculations and reduce data transfer needs, potentially increasing output efficiency by six to ten times per watt of energy consumed.
This enhancement means Google could potentially handle ten queries with the energy typically required for just one. While users of Gemini may not notice any changes, operational costs could decrease, giving Google a competitive advantage against rivals like OpenAI, Anthropic, and Chinese labs that dominate a significant portion of the U.S. AI market due to their lower operational costs.
Following the announcement, Alphabet's stock rose around 3%, reaching an intraday peak of $356, although this gain has stabilized as investors await the upcoming earnings report on July 22.
The development of Frozen v2 signifies a strategic effort by Google and other tech giants to reduce their reliance on Nvidia hardware, which currently commands about 85% of the AI GPU market. As Nvidia's products were originally designed for gaming rather than language models, companies like Google are pursuing custom chips that promise greater efficiency, translating into substantial cost savings.
Despite the potential benefits, Frozen v2 is still in the early stages of development, with key design decisions pending. Google has yet to publicly acknowledge the project, and the chip will be specifically tailored for its own models rather than external cloud customers, with a projected launch in 2028. In the meantime, Google is leasing 110,000 Nvidia GPUs from SpaceX at a cost of $920 million per month as a stopgap measure.
