Google's secret 'Frozen v2' AI chip could make Gemini up to 10 times more efficient, report says

Google may be preparing one of its biggest hardware upgrades in the artificial intelligence race. According to a report by The Information, the company is developing a new custom AI server chip internally known as Frozen v2. Unlike Google's existing Tensor Processing Units (TPUs), the new processor is reportedly designed specifically to run Gemini models faster while using far less electricity. The report excited investors, sending Alphabet shares up about 3 per cent on Monday, as markets viewed the project as a potential answer to Google's growing demand for AI computing power. Google has not confirmed the existence of Frozen v2, but said its teams are constantly researching new hardware and software technologies to improve AI performance and efficiency.

What makes Frozen v2 different?

Today's AI chips are built to support many different types of machine learning workloads.

Frozen v2 reportedly takes a much more specialised approach. Instead of relying solely on software, parts of Gemini's architecture would be permanently embedded into the silicon itself. This reduces unnecessary calculations and data movement every time the AI responds to a query. According to engineers cited by The Information, Frozen v2 could deliver six to ten times more AI tokens per unit of power than Google's newest TPUs.

In simple terms, that means Gemini could answer more questions while consuming significantly less energy one of the biggest challenges facing the AI industry today.

Why Google needs a new AI chip

Running advanced AI models has become incredibly expensive. Every chatbot response requires massive computing resources, and demand continues to grow as more businesses adopt generative AI. The report says Google has been struggling with an internal shortage of AI computing capacity, forcing difficult decisions over where those resources are allocated. The company has even reportedly relied on external infrastructure to meet enterprise demand. Frozen v2 is designed to ease that pressure by making Gemini much more efficient without requiring the same increase in computing hardware.

However, there is a trade-off.

Because the chip is optimised for Gemini's current architecture, major changes to future AI models could reduce its usefulness. That is reportedly why Google views Frozen v2 as an experimental project rather than a full replacement for its TPUs.

Hardware is becoming the next AI battleground

The reported chip comes at a crucial time for Google.

The company is facing intense competition from OpenAI, Anthropic, Moonshot AI and Alibaba, while reports suggest the launch of Gemini 3.5 Pro has been delayed. At the same time, demand for AI infrastructure continues to surge as companies look for faster, cheaper and more energy-efficient ways to deploy large language models.

Increasingly, success in artificial intelligence is no longer determined only by better models. The companies that build the most efficient chips may also gain a significant competitive advantage. If Frozen v2 reaches production around the reported 2028 target, it could become one of Google's most important AI hardware projects, helping the company reduce costs while scaling Gemini to billions of users worldwide.