NVIDIA GTC Conference Embraces AI Chip Shift? U.S. Media: CPUs Will Return to the Center Stage

Nvidia’s Latest Vera CPU

TechCrunch reports that on March 14, Beijing time, according to CNBC, for years, Nvidia’s graphics processors (GPU) have been among the hottest chips on the market, but the sudden rise of AI agents has revived its more low-profile host chip central processing unit (CPU).

Now, Nvidia is preparing to announce new details about its AI-optimized CPU at the annual GTC conference opening on Monday, likely unveiling a rack of pure CPUs at the booth.

“CPUs are becoming the bottleneck for scaling AI and agent workflows,” said Dion Harris, head of Nvidia AI infrastructure, to CNBC this week. He sees this as an “exciting opportunity.”

Nvidia released its first data center CPU, Grace, in 2021, and the next-generation Vera is now in production. These CPUs are typically deployed in full rack-level systems alongside Nvidia’s renowned Hopper, Blackwell, or Rubin GPUs.

The surge in GPU demand has made Nvidia a household name and the world’s most valuable publicly traded company, with a market cap of $4.4 trillion. In February this year, Nvidia’s overall chip strategy underwent a major shift. At that time, Nvidia reached a multi-year agreement with Meta, including the first large-scale deployment of the Grace CPU independently, with plans to deploy Vera in 2027.

Nvidia told CNBC that thousands of independent Nvidia CPUs are also powering supercomputers at the Texas Advanced Computing Center and Los Alamos National Laboratory.

AI Agents Drive CPU Revival

Bank of America predicts that the CPU market size could more than double, growing from $27 billion in 2025 to $60 billion in 2030. Just in the last quarter, Nvidia generated over $62 billion in data center revenue, up 75% year-over-year.

The revival of CPUs stems from a fundamental shift in computing demands: as AI becomes widespread, application scenarios are shifting from question-answer chatbots to task-oriented AI agents.

While GPUs are well-suited for training and running AI models due to their thousands of small cores focused on parallel computations, CPUs have fewer high-performance cores capable of running continuous general-purpose tasks.

AI agents require substantial general computing power because they need to transmit and process large amounts of data within AI workflows and coordinate and schedule among multiple agents.

Jensen Huang

Nvidia CEO Jensen Huang (Jensen Huang) said during last month’s earnings call, “These agent systems are evolving into different types of agents working together like a team. The number of tokens generated has grown exponentially, so we need to perform inference at higher speeds.”

Huang repeatedly mentioned AI agents during the call and stated, “In the face of hardware demand shifts, the most important thing is performance per watt.”

Nvidia said in a press release that its standalone CPUs significantly improved performance per watt in Meta’s data centers. Creative Strategies chip analyst Ben Bajarin (Ben Bajarin) commented, “This is a brand-new infrastructure: a new scale-out of pure CPU racks dedicated solely to running AI agents. Your software will be elsewhere, your accelerators only handle tokens, but there must be an intermediate layer to coordinate and schedule.”

CPU Supply Crisis

Today, the once-silent CPU market is facing what consulting firm The Futurum Group calls a “quiet supply crisis.” The firm predicts that by 2028, the growth rate of the CPU market may surpass that of GPUs.

According to Reuters, leading CPU suppliers AMD and Intel have issued supply shortage warnings to Chinese customers. Reports indicate that CPU delivery cycles are up to six months, with prices rising over 10%.

“Demand has surged unprecedentedly over the past six to nine months,” said Forrest Norrod, head of data center at AMD, in an interview with CNBC.

Norrod believes that CPU demand will not slow down or stop growing in the short term, but AMD has anticipated the increase and is “ramping up efforts” to meet demand.

An Intel spokesperson told CNBC that the company’s inventory is expected to hit a “bottom” this quarter, “but we are actively responding, and supply is expected to gradually improve from Q2 through the end of 2026.”

Creative Strategies chip analyst Bajarin said, “Wafers don’t grow on trees. We can’t harvest 10% more silicon wafers out of thin air. The entire industry is facing supply tightness. Unfortunately, CPU wafer supply is constrained.”

When asked whether Nvidia might face delays in CPU shipments, Nvidia AI infrastructure head Harris told CNBC, “Everything is normal at the moment.”

He said Nvidia has a “robust supply chain” capable of handling demand pressures, largely because the company sells many CPUs alongside GPUs in rack-level systems.

Optimized for GPUs

Harris stated that compared to more general-purpose CPUs produced by Intel and AMD, Nvidia has taken a fundamentally different approach in design, making its CPUs “most suitable” for data processing and AI agent workflows.

One key difference is the number of cores per CPU. AMD’s EPYC series and Intel’s Xeon high-performance server CPUs typically have 128 cores, while Nvidia’s Grace CPU has 72 cores.

AMD EPYC Server CPUs

Harris explained, “If you’re a hyperscale cloud provider, you’d want to maximize cores per CPU, which is essentially about reducing cost per core. That’s a business model.”

However, Nvidia designed its CPUs specifically to support its flagship GPUs running AI workloads. “In this case, single-thread performance is more important than cost per core because you want to ensure that the extremely expensive GPU resources don’t sit idle,” Harris said.

Nvidia’s CPUs are based on the ARM architecture, which is more commonly used in smartphones and low-power devices, whereas Intel and AMD’s CPUs are based on the traditional x86 architecture. x86 was introduced by Intel nearly 50 years ago and has dominated PC and server processor design since.

AMD data center head Norrod said, “I think Nvidia has optimized their chips very well for providing compute power to GPUs, but they haven’t optimized for general-purpose applications.”

In fact, Nvidia does rely on more general-purpose CPUs for some products. For example, in the HGX Rubin NVL8 platform, Nvidia pairs its own GPUs with Intel or AMD host CPUs as foundational components for customers building their own AI racks.

As Nvidia advances into the independent CPU market, more of its customers are developing ARM-based processors for their data centers.

Amazon was the first major hyperscale cloud provider to launch a self-designed CPU, releasing the Graviton processor in 2018. The Futurum Group states that Google’s Axion processor, launched in 2024, now handles about 30% of internal applications. Microsoft released its second-generation Cobalt processor last November. ARM is expected to launch its own self-developed CPU this year, with Meta as an early customer.

Research firm Mercury Research estimates that in Q4 2025, the server CPU market share will be led by Intel at 60%, with AMD at 24.3%, Nvidia at 6.2%, and the remaining share held by hyperscale cloud companies like Amazon, Microsoft, and Google developing ARM-based self-designed CPUs. (Author/Shao Yu)