Meta buys millions of Nvidia chips

Mark Zuckerberg, CEO of Meta, announced the new agreement on February 17, saying the expansion of this partnership continues to advance the company’s efforts to “bring personal superintelligence to everyone in the world.” The deal includes new standalone CPUs and Nvidia’s next-generation Vera Rubin system.

Meta has used Nvidia chips for at least a decade, but today’s deal marks a much deeper technology partnership between the two Silicon Valley giants. A notable point is that Meta became the first company to deploy Nvidia’s Grace CPU as a standalone chip in the data center instead of integrating it with the GPU in the server.

“They are really designed to run inference tasks and agent jobs as a complement to Grace Blackwell/Vera Rubin server racks. Meta’s deployment at scale is an affirmation of the ‘A to Z’ strategy that Nvidia is pursuing across both the CPU and GPU infrastructure groups,” said chip analyst Ben Bajarin from research firm Creative Strategies.

According to Ian Buck, Nvidia’s head of high-performance and hyperscale computing, Grace CPUs have proven they can reduce power consumption by 50% for some common tasks like running databases, and the next generation, Vera, achieves even more improvements. Meta plans to deploy next-generation Vera CPUs in 2027.

The financial terms of the new deal were not disclosed, but Bajarin said it was worth tens of billions of dollars. “We anticipate a significant portion of Meta’s capital spending will be on building infrastructure with Nvidia chips,” he said.

The new deal is part of Meta’s overall commitment to spend $600 billion in the US through 2028 on data centers and necessary infrastructure. In January, Meta announced plans to spend up to $135 billion on AI this year. The technology giant is planning to build 30 data centers, of which 26 centers will be located in the US. The two largest AI data centers under construction include the 1 GW Prometheus in New Albany, Ohio, and the 5 GW Hyperion in Richland Parish, Louisiana.