Google is working on a new AI chip designed to make Gemini more efficient
Alphabet is designing a new server chip, internally dubbed “Frozen v2,” that could make its Gemini AI models substantially more efficient to run. According to a report from The Information, the chip is expected to arrive in 2028 and could deliver between six and ten times the efficiency of Google’s current AI chips when measured by tokens generated per unit of power.
What is Frozen v2 and how does it differ from Google’s current TPUs?
Frozen v2 would hardwire parts of the Gemini model family directly into the chip’s silicon rather than running the model as software on general-purpose AI hardware. The goal is a co-designed system in which the hardware and model are built together, allowing each query to cost less power, less time, and less money.
Google declined to directly confirm or deny the report. In a statement to TechCrunch, the company said, “Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers. While not every project moves into production, this rigorous exploration is central to our full stack approach. By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads.”
How much more efficient could Frozen v2 be?
Engineers working on the project reportedly expect Frozen v2 to be between six and ten times more efficient than Google’s existing Tensor Processing Units when serving AI queries. The efficiency metric cited in the report is tokens generated per unit of power, a common way to gauge how much useful output an accelerator produces for the energy it consumes.
Those gains would matter for Google because running Gemini at scale consumes enormous compute resources, and reducing the energy and hardware cost per response directly improves the economics of offering AI products.
When will Frozen v2 be ready?
The chip is not expected to arrive until 2028, according to The Information’s sources. That timeline positions Frozen v2 as a longer-term infrastructure bet rather than a near-term upgrade to Google’s serving fleet, and it leaves room for the design to change before it reaches production hardware.
Why is Google building its own AI chips?
AI companies have increasingly moved toward custom silicon to make in-house models more efficient and to ease pressure on the global supply of AI compute capacity. Efficiency has also become a key selling point as concerns about AI spending have cooled some of the market enthusiasm that previously surrounded the sector.
At the same time, the major AI labs are working to reduce their reliance on Nvidia, whose GPUs have historically dominated the AI accelerator market and left major model makers dependent on its hardware. In June, OpenAI announced its first custom chip, an inference processor called Jalapeño. Earlier this month, it was reported that Anthropic was discussing a new chipmaking partnership with Samsung.
How does this fit into Google’s AI spending plans?
Investors have previously questioned Alphabet’s large planned expenditures on AI infrastructure. Earlier this year, Google said it plans to spend between $180 billion and $190 billion, a figure that has drawn scrutiny about whether the investments will generate sufficient returns.
News of the more efficient Frozen v2 chip appears to have eased some of those concerns. Following publication of The Information’s report, Alphabet’s stock climbed roughly 3% on Monday morning ahead of the company’s earnings report later that week.
FAQ
What is Google’s Frozen v2 chip?
Frozen v2 is an internally developed server chip that Alphabet is designing to make Gemini more efficient by hardwiring parts of the model into the silicon. It is not expected to be released until 2028.
How much more efficient is Frozen v2 than Google’s current AI chips?
Engineers reportedly expect Frozen v2 to be between six and ten times more efficient than Google’s existing AI chips, measured by tokens generated per unit of power.
Why is Google building custom AI chips instead of buying Nvidia GPUs?
Google is pursuing custom silicon to lower the cost and energy required to serve Gemini at scale, to reduce reliance on Nvidia, and to address ongoing shortages in AI compute capacity. Other AI labs, including OpenAI with its Jalapeño inference chip, are pursuing similar strategies.
Related coverage
- Google Researchers Detail a New System for Detecting Coordinated AI Spam
- Google Gemini 3.5 Pro Delayed as Coding Performance Falls Short
This article summarizes reporting from techcrunch.com.