Google Developing Specialized ‘Frozen v2’ Chip to Supercharge Gemini AI Efficiency

Google is engineering a new class of specialized silicon, internally dubbed “Frozen v2,” designed to drastically reduce the energy and computational costs of running its Gemini artificial intelligence models. The project represents a strategic shift toward hardware that is hard-wired for specific software architectures, aiming to solve the persistent “compute crunch” currently hampering the tech giant’s AI ambitions.

Unlike Google’s existing Tensor Processing Units (TPUs), which serve as general-purpose accelerators for various machine learning tasks, Frozen v2 is being built to integrate specific elements of the Gemini architecture directly into the chip’s circuitry. According to reports from The Information, this deep integration could allow the processor to handle between six and ten times more tokens per unit of energy than the company’s most advanced current hardware. By minimizing the distance data must travel and reducing the total number of calculations required for a single query, Google hopes to make its AI services significantly more sustainable and faster.

The development comes at a critical juncture for Google’s DeepMind and Cloud divisions. The company has reportedly faced internal friction due to a shortage of available computing power, which has delayed the rollout of newer iterations of Gemini Pro. Furthermore, the competitive landscape is shifting rapidly; while Google pioneered custom AI silicon in 2015 to handle its massive search and translation workloads, its rivals are now catching up. OpenAI is currently collaborating with Broadcom on its own custom chip, codenamed “Jalapeño,” and Anthropic has reportedly entered preliminary discussions with Samsung to develop bespoke hardware.

However, the high level of specialization in Frozen v2 carries significant risks. Because the chip is tailored to the specific mathematical structure of Gemini, it could become obsolete if Google’s researchers decide to pivot to a different model architecture. This dependency has led the company to treat the project as a pilot program rather than a full-scale replacement for its versatile TPU lineup. If the project moves forward successfully, the specialized chips are expected to be deployed within Google’s data centers by 2028.

The push for custom silicon is also a defensive move against the rising dominance of NVIDIA, whose H100 and Blackwell GPUs currently command the lion’s share of the AI hardware market. By designing its own chips, Google not only reduces its reliance on external vendors but also optimizes its cloud infrastructure for third-party developers who may eventually use these specialized processors to run Gemini-based applications.

While Google has a long history of experimental hardware projects that never reach mass production, a spokesperson confirmed to CNBC that the company is “constantly exploring new ways to improve performance and efficiency.” This exploration is increasingly vital as Google faces pressure from both Western competitors and emerging Chinese AI models, such as Kimi, which are beginning to gain traction in global markets.