Nvidia’s Partnership with Meta Marks a Transformation in Computational Power

Inquire with anyone about what Nvidia produces, and they will probably respond with “GPUs.” For years, this chip manufacturer has been synonymous with advanced parallel computing, and the rise of generative AI, along with heightened GPU demand, has greatly benefited the company.

However, Nvidia’s latest initiatives indicate that it aims to secure more clients at the lower compute-intensive segment of the AI market—clients who may not require the most robust, high-performance GPUs for training AI models but are instead seeking efficient methods to operate agentic AI software. Recently, Nvidia invested billions to obtain technology from a chip startup specializing in low-latency AI computing, and it has also begun offering stand-alone CPUs as part of its new superchip system.

Just yesterday, Nvidia and Meta announced that the social media behemoth had agreed to purchase billions of dollars’ worth of Nvidia chips to support its vast infrastructure projects—Nvidia’s CPUs included in the package.

This multiyear agreement marks an extension of a strong ongoing collaboration between the two firms. Meta had previously estimated that by the close of 2024, it would have acquired 350,000 H100 chips from Nvidia, and by the end of 2025, the company would have access to a total of 1.3 million GPUs (although it was unclear if all of these would be from Nvidia).

In the latest announcement, Nvidia stated that Meta would “construct hyperscale data centers optimized for both training and inference, in alignment with the company’s long-term AI infrastructure strategy.” This entails a “large-scale deployment” of Nvidia’s CPUs and “millions of Nvidia Blackwell and Rubin GPUs.”

Importantly, Meta is the first major tech player to disclose a significant bulk purchase of Nvidia’s Grace CPU as a stand-alone chip, something Nvidia indicated would be feasible when it introduced the full specifications of its new Vera Rubin superchip in January. Nvidia has also been highlighting its technology that connects various chips, as part of its comprehensive approach to compute power, as described by one industry analyst.

Ben Bajarin, CEO and principal analyst at the tech market research firm Creative Strategies, believes this move reflects Nvidia’s awareness that a broader range of AI software now requires operation on CPUs, similar to traditional cloud applications. “The reason why the industry is so optimistic about CPUs in data centers right now is agentic AI, which places new requirements on general-purpose CPU architectures,” he notes.

A recent report from the chip newsletter Semianalysis reinforces this perspective. Analysts observed that CPU utilization is rising to facilitate AI training and inference, referring to one of Microsoft’s data centers for OpenAI, where “tens of thousands of CPUs are required to process and manage the petabytes of data produced by the GPUs, a scenario that wouldn’t have existed without AI.”

However, Bajarin points out that CPUs are merely one part of the most sophisticated AI hardware setups. The quantity of GPUs Meta is acquiring from Nvidia remains greater than that of CPUs.

“If you’re among the hyperscalers, you won’t be relying solely on all your inference computing on CPUs,” Bajarin explains. “You simply need whatever software you’re using to be sufficiently fast on the CPU to interact with the GPU architecture that’s truly driving that computation. Otherwise, the CPU can become a bottleneck.”

Meta opted not to comment on its extended agreement with Nvidia. During its recent earnings call, the social media giant stated that it intends to significantly enhance its fund allocation for AI infrastructure this year to between $115 billion and $135 billion, an increase from $72.2 billion last year.