"Missed the GPU Boom, But Now It's CPU? NVIDIA Takes the Lead [Tech Talk]"

Vera CPU Takes Center Stage at GTC

As AI Tasks Grow More Complex, Focus Shifts to CPU Bottlenecks

Fierce Competition Emerges Among Intel, AMD, and ARM

Editor's NoteThis series explains the essential, yet often unfamiliar, technologies shaping our lives—from AI and semiconductors to communications and biotechnology—in simple, accessible language.

At the NVIDIA GTC (GPU Technology Conference) held in San Jose, USA, on March 16 (local time), the spotlight was on none other than the central processing unit (CPU). NVIDIA's next-generation CPU, Vera, will not be sold merely as a component for AI computers, but as a standalone product. What led NVIDIA, a company renowned for its GPU designs, to enter the CPU business?

NVIDIA's First Standalone CPU: Vera

NVIDIA's Vera Central Processing Unit (CPU). NVIDIA

The newly unveiled Vera CPU is a processor equipped with 88 Olympus cores, designed in-house by NVIDIA. These cores are based on the ARM v9.2 architecture from ARM Holdings, a company with which NVIDIA maintains a close partnership.

NVIDIA has previously designed ARM-based CPUs, but starting with Vera, it will offer CPUs as standalone products. CEO Jensen Huang stated, “There is significant market interest in Vera, so we will sell this product separately,” adding, “We never imagined we would sell CPUs independently, but we expect to generate billions of dollars in revenue from the CPU market soon.”

Until now, CPUs have been treated as mere accessories in AI data centers. While graphics processing units (GPUs) and memory handled AI training workloads, CPUs were relegated to various auxiliary tasks. This is also why CPUs have been relatively overlooked as the prices of GPUs and memory soared.

However, the landscape is gradually changing. Since last year, leading AI companies have identified CPUs as the next bottleneck for AI. Sam Altman, CEO of OpenAI, emphasized at an AMD event in mid-2025 that “tremendous CPUs are needed.”

'Amdahl's Law': No Matter How Much GPUs Improve, Overall Computer Performance Depends on the CPU

AI data centers require high-performance CPUs capable of directing an astronomical number of compute cores. Yonhap News

Today’s hyperscale AI data centers coordinate tens of thousands of GPUs at once. Since each GPU contains tens of thousands of cores, an astronomical number of compute units are in operation. This gives rise to a new bottleneck: after completing an AI computation, some GPU cores are left idle until the next task is assigned. No matter how much parallel processing is optimized, certain tasks within the computer system must be performed sequentially, resulting in bottlenecks known in computer science as “Amdahl’s Law.”

To ensure that GPUs are utilized efficiently and operate without idle time, a central orchestrator is needed to direct tasks sequentially to millions of cores. This is a role only the CPU can fulfill. Supplying data stored in memory to GPUs and carefully scheduling tasks to prevent GPUs from waiting too long for input is called “scheduling.” The future performance of AI data centers will depend on the scheduling and orchestration capabilities of CPUs.

Competition Heats Up Among Intel, AMD, ARM, and Others

Intel semiconductor factory sign. Photo by Reuters Yonhap News Agency

Another advantage of CPUs is their versatility. While GPUs are accelerators designed for AI training and inference tasks, CPUs are fundamentally capable of handling a broad range of work. In other words, CPUs can also serve as engines for large-scale inference operations that deploy trained AI models.

This is why NVIDIA plans to sell the Vera CPU as a standalone product. The new product, called the “Vera CPU Rack,” will be available in a large-scale hardware configuration integrating 256 Vera CPUs. The Vera CPU Rack serves as an “AI factory,” capable of running countless AI agents simultaneously.

NVIDIA is not alone in addressing CPU bottlenecks. Longtime CPU powerhouse Intel participated in GTC for the first time this year, providing its Xeon 6 server processor as the host CPU for the Rubin AI server. Meanwhile, AMD responded by announcing the development of its next-generation EPYC CPU on its corporate blog. ARM Holdings, which supplies key CPU design technologies to NVIDIA, is also expanding the ARM CPU ecosystem through collaborations with tech giants such as Google, Microsoft, and Amazon.

This content was produced with the assistance of AI translation services.