NVIDIA's (NVDA) immense market value is currently built largely on specialized graphics processing units (GPUs) for artificial intelligence servers. However, CEO Jensen Huang is increasingly demonstrating a preference for general-purpose central processing units (CPUs). For decades, CPUs have traditionally been considered the "brain" of computers, a product category almost synonymous with Intel (INTC) and sometimes associated with Advanced Micro Devices (AMD). Huang previously noted that historically, 90% of computing tasks were handled by CPUs, with only 10% performed by GPUs, but this ratio has reversed in recent years.

Nevertheless, as AI companies shift their focus from model training to model deployment, CPUs are regaining importance and are even being viewed as a choice not inferior to GPUs. NVIDIA plans to play a significant role in this transition. During the company's fourth fiscal quarter earnings call, Jensen Huang stated, "We love CPUs, just as we love GPUs." He assured analysts that NVIDIA is not only prepared for the resurgence of CPUs but also that its data center CPU products, first launched in 2023, will possess capabilities surpassing those of competitors.

At the Consumer Electronics Show in Las Vegas last month, Huang further predicted an explosive growth in the usage of high-performance NVIDIA CPUs within data centers, adding, "I wouldn't be surprised if NVIDIA becomes one of the world's largest CPU manufacturers."

For decades, CPUs and GPUs have served distinct functions. CPUs, as general-purpose chips, are designed to handle various mathematical tasks assigned by software programmers at a reasonable speed. In contrast, GPUs specialize in executing a set of relatively simpler mathematical operations but can perform them thousands of times simultaneously in parallel. In video games, this translates to calculating the values of thousands of pixels on the screen multiple times per second; in AI, it involves multiplication and addition operations on large numerical matrices used by developers to represent real-world data like text and images.

AI companies are increasingly deploying "agents" capable of autonomously performing tasks such as writing code, filtering documents, and composing research reports. Ben Bajarin, an analyst at Creative Strategies, pointed out that this type of computing "is increasingly, and sometimes primarily, running on CPUs." He believes that the configuration of NVIDIA's current flagship AI server, the NVL72—which contains 36 CPUs and 72 GPUs—might evolve toward a 1:1 ratio when handling so-called "agent" workloads, and could potentially operate without GPUs entirely.

To demonstrate its ambitions in the CPU arena, NVIDIA recently announced an agreement with Meta Platforms, which will adopt NVIDIA's Grace and Vera CPU chips in large quantities as standalone products. This marks a relatively new direction compared to the model in NVIDIA's existing AI servers, where each CPU is paired with multiple GPUs. However, this does not necessarily mean Meta is replacing its CPU supplier; it is more likely expanding its supplier portfolio. A few days later, AMD also announced a significant deal with Meta involving its CPUs, and Meta has a history of purchasing AMD CPUs spanning many years.

During the analyst call, Jensen Huang outlined the fundamentally different path NVIDIA is taking in CPU design. He explained why NVIDIA has minimized the approach used by Intel and AMD, which involves breaking chips down into smaller units, emphasizing that NVIDIA's CPUs can continuously process numerous simple tasks efficiently by leveraging strong access to large amounts of memory. "It is designed to focus on extremely high data throughput," Huang stated, "because most of the computing problems we focus on are data-driven—artificial intelligence is one of them."

Dave Altavilla, principal analyst at HotTech Vision and Analysis, believes NVIDIA aims to prove that the traditional CPU, once primarily supplied by Intel, "is no longer the default assumed foundation for modern computing infrastructure, but merely one architectural choice among many." Huang revealed that NVIDIA will disclose more information about its CPUs at its annual developer conference in Silicon Valley next month.