A massive new agreement announced on Tuesday local time shows that Meta will use millions of Nvidia chips in its artificial intelligence data center, including Nvidia's new independent CPUs and next-generation Vera Rubin systems. Meta CEO Mark Zuckerberg said in a statement that this expanded cooperation will continue to advance the company's goal of "providing personal superintelligence to everyone in the world," a vision formally proposed in July last year.
Financial terms of the agreement were not disclosed.
In January this year, Meta announced that AI-related investment in 2026 will reach a maximum of $135 billion. Ben Bajarin, a chip industry analyst at Creative Strategy Consulting, said: "The size of this agreement is undoubtedly in the tens of billions of dollars, and it is expected that a considerable part of Meta's capital expenditure will be invested in this construction cooperation with Nvidia."
This is not the first time that the two parties have collaborated, as Meta has been using Nvidia GPUs for at least a decade, but the agreement marks a significant broadening of the scope of technology cooperation between the two Silicon Valley giants.
Independent CPUs are the biggest highlight of this cooperation: Meta will be the first company to deploy NVIDIA Grace CPUs as independent chips in data centers, rather than integrating them with GPUs for use in servers. Nvidia says this is the first large-scale independent deployment of Grace CPU.
Bajarin said: "These chips are designed for running inference tasks and agent tasks as supporting components of the Grace Blackwell/Vera Rubin cabinet. Meta's large-scale adoption of this move confirms Nvidia's full-stack layout strategy on both CPU and GPU infrastructure."
The next generation Vera CPU is scheduled to be deployed by Meta in 2027.
The multi-year agreement is part of Meta's overall plan: the company has committed to investing $600 billion in data centers and supporting infrastructure in the United States by 2028.
Meta plans to build 30 data centers, 26 of which will be located in the United States. Two of the largest AI data centers are currently under construction: the 1-gigawatt Prometheus project in New Albany, Ohio, and the 5-gigawatt Hyperion project in Richland Parish, Louisiana.
The agreement also includes NVIDIA networking technology Spectrum-X Ethernet switches for high-speed GPU interconnect in large-scale AI data centers. Meta will also use Nvidia’s security capabilities to support WhatsApp’s AI capabilities.
The social media giant is not solely dependent on the leading chip manufacturer. In November last year, there were reports that Meta was considering using Google's Tensor Processing Unit (TPU) in data centers in 2027. Nvidia's stock price fell 4% that day as soon as the news came out.
Meta also develops its own chips and uses AMD products. In October last year, AMD reached a major cooperation with OpenAI. AI giants are looking for second supply channels when Nvidia's supply is tight.
Nvidia's current Blackwell GPU has been out of stock for several months, and its next-generation Rubin GPU has just recently entered mass production. Through this agreement, Meta has locked in sufficient supply.
NVIDIA and Meta engineering teams will conduct in-depth joint design to optimize and accelerate cutting-edge AI model training and inference for Meta.
Meta is developing a new cutting-edge large model, codenamed Avocado, as a successor to its Llama series of AI technologies. According to previous reports, the latest version released last spring failed to satisfy developers.
Meta's stock price has been volatile in recent months, and its AI strategy has particularly confused Wall Street.
In October last year, after Meta announced its ambitious AI investment plan, its stock price suffered its largest one-day drop in three years; in January this year, after the company issued revenue guidance that exceeded expectations, its stock price rose 10% in a single day.