Google is developing a specialized chip that could revolutionize the efficiency of its artificial intelligence systems. Known internally as Frozen v2, this processor promises to multiply by up to ten times the energy efficiency of current machines, according to hipertextual.com.

A chip designed specifically for Gemini

Unlike general-purpose processors, Frozen v2 is designed as a specialized tool that would directly integrate components of Gemini's architecture, Google's AI model. This approach would allow reducing both the calculations needed and the data movement required to process user queries.

The result would be spectacular in terms of efficiency: the chip could process between six and ten times more tokens per unit of energy compared to the company's most powerful current TPUs (Tensor Processing Units). In a context where the energy consumption of large AI models has become a global economic and environmental challenge, this figure takes on special relevance.

A tailored solution, not a replacement

It is important to clarify that Google has no intention of replacing its existing TPUs. Instead, Frozen v2 would be added to the catalogue of tools available for specific scenarios where its specialization would be a decisive advantage. This diversified strategy allows the company to optimize resources based on each use case.

However, there is a significant drawback: the chip's own specialization. Frozen v2 would only work with future versions of Gemini if Google maintains the same architecture in its AI models. Any fundamental change could make the processor obsolete or require major updates.

A pilot project with a long-term launch

Due to these risks, Google considers Frozen v2 as a pilot test for now. There are no plans to manufacture it at scale as is the case with its current TPUs. According to available information, the company plans to launch the chip in 2028, meaning there are still several years of development and refinement ahead.

Although the company has not officially denied the project, its spokespeople have been cautious. A Google representative told CNBC that its teams "constantly explore new ways to improve performance and efficiency for users and customers," without explicitly confirming or denying Frozen v2.

Google is not alone in this race

It is important to frame this initiative within a broader industry trend. Practically all major artificial intelligence companies are betting on developing their own chips tailored to their systems:

• OpenAI recently unveiled Jalapeño, its first custom processor for ChatGPT, in collaboration with Broadcom.
• Anthropic would be negotiating with Samsung to create its own chip, although conversations are still in a preliminary phase.
• Google, for its part, has a decade of experience in specialized hardware after designing its TPUs in 2015, when it anticipated that AI demand would exceed its available infrastructure.

Today, Google offers these TPUs to third parties through its data centres, positioning itself as a direct competitor to NVIDIA in the AI processor market.

A project with uncertain but realistic future

Google has a history filled with projects that never reach mass production. However, an AI chip seems to be something different: the company has the experience, resources and motivation necessary to make it happen. The fact that numerous companies in the sector are following the same path suggests that this is not a passing trend, but an inevitable evolution of the market.

The question left unanswered is whether Frozen v2 will manage to maintain its usefulness in 2028, when the architecture of AI models will have evolved significantly. For now, we can only wait and see how Google resolves this and other challenges in its artificial intelligence division.