Google Reportedly Building ‘Frozen v2’ AI Chip to Supercharge Gemini
Quick Read
- Google is developing a new AI chip, internally called “Frozen v2,” built to run Gemini models far more efficiently than today’s hardware.
- The chip could deliver 6 to 10 times more AI tokens per unit of power than Google’s latest TPUs, according to The Information.
- Frozen v2 hardwires parts of Gemini’s architecture directly into the silicon, cutting down on data movement and processing steps during inference.
- Google is targeting deployment as early as 2028, positioning the chip as a complement to its TPUs rather than a replacement.
Google appears to be taking a bold new approach to solving its AI compute crunch. According to a report from The Information, the company is developing a specialized AI chip, internally known as Frozen v2, designed to run its Gemini models with dramatically improved efficiency.
Unlike Google’s general-purpose Tensor Processing Units (TPUs), which are built to handle a wide range of AI models, Frozen v2 works by permanently etching part of the Gemini large model’s computing logic onto the silicon hardware itself. Think of it like the difference between a Swiss Army knife and a tool built for exactly one job the specialized tool does that one job faster and with far less wasted effort. By baking Gemini’s architecture straight into the chip, Google can cut down on the amount of data the chip has to shuffle around, making it faster at responding to queries.
The payoff, if it works, is significant. The chip could deliver six to ten times more AI tokens per unit of power than Google’s latest TPU chips. In plain terms, Frozen v2 could let Google serve far more AI responses using the same amount of electricity a big deal at a time when power and compute capacity, not just chip supply, have become the industry’s biggest bottleneck.
That bottleneck is reportedly a major driver behind the project.
The initiative aims to ease a severe internal compute shortage that has forced Google Cloud to turn away business, and even led to a nearly $1 billion monthly deal with SpaceX to help manage capacity. Google is said to be feeling the squeeze on multiple fronts, including pressure from rivals like OpenAI and Anthropic, while its own next-generation Gemini model reportedly runs behind schedule.
Google isn’t planning to replace its TPUs with Frozen v2. Instead, the chip intended to complement its Tensor Processing Units as the company works to expand AI computing capacity overall. Engineers are reportedly still deciding exactly how much of Gemini to freeze into the hardware, balancing efficiency against flexibility.
This isn’t unfamiliar territory for Google. The company has been building custom AI silicon since it launched its first-generation TPU back in 2016, well ahead of most of its Big Tech peers, who have relied more heavily on Nvidia GPUs. Frozen v2 signals Google pushing further into an “Apple-style” vertically integrated approach, tightly coupling its own software and hardware in pursuit of cheaper, faster AI inference.
For now, Frozen v2 remains in development, with deployment targeted for around 2028. Markets reacted quickly to the news regardless Alphabet shares climbed 3% on Monday following the report, with investors betting on the long-term payoff of cheaper, more efficient AI infrastructure even though the chip is still years away from powering real workloads.
