Why the CPU plays a central role

NvidiaThe graphics processing unit of has been the best-selling chip for many years, but the sudden emergence of agent-based artificial intelligence has led to the resurgence of a more modest host chip: the central processing unit.

Now, Nvidia is poised to announce new details about its agent-optimized CPUs at its annual GTC conference, which begins Monday, and CPU-only racks could be hitting showroom floors.

Dion Harris, Nvidia’s head of AI infrastructure, told CNBC this week that “the CPU is becoming the bottleneck in terms of scaling this AI and agent workflow,” calling it a “great opportunity.”

The chip giant announced its first data center CPU, Grace, in 2021, and the next generation Vera is now in production. CPUs are typically deployed in full rack-scale systems alongside Nvidia’s famous Hopper, Blackwell, or Rubin GPUs.

Explosive demand for GPUs has made Nvidia a household name and the world’s most valuable publicly traded company with a market capitalization of $4.4 trillion. The company’s broader chip strategy took a major turn in February when the company signed a multi-year deal with Nvidia. meta This includes the first large-scale deployment of its proprietary Grace CPUs, with plans to deploy Vera in 2027.

Nvidia’s thousands of standalone CPUs also help power supercomputers at the Texas Advanced Computing Center and Los Alamos National Laboratory, Nvidia told CNBC.

bank of america We predict that the CPU market could more than double from $27 billion in 2025 to $60 billion by 2030. In the most recent quarter alone, Nvidia’s data center revenue exceeded $62 billion, an increase of 75% year-over-year.

The resurgence of CPUs is being driven by a fundamental shift in computing needs as AI adoption moves from call-and-answer chatbots to task-oriented agent apps.

GPUs are ideal for training and running AI models because they have thousands of small cores that specialize in performing many operations simultaneously, whereas CPUs have fewer powerful cores that perform sequential general-purpose tasks.

Agentic AI requires large amounts of general computing power to move large amounts of data and coordinate across multiple agents for AI workflows.

“These agent systems come from different agents working as a team,” CEO Jensen Huang said on Nvidia’s earnings call last month. “The number of tokens being generated is really increasing exponentially, so we need to infer faster.”

Huang mentioned agent AI more than a dozen times on the conference call, saying that as hardware needs change, “it’s literally all about the best performance per watt.”

The company said in a press release that its standalone CPUs will significantly improve performance per watt in Meta’s data centers.

“This is new infrastructure. It’s a greenfield extension of the CPU rack whose only role is to run agent AI,” said Ben Bajarin, a chip analyst at Creative Strategies. “The software is in a different place and the accelerator just runs the token, but something has to sit in the middle to coordinate it.”

“Silent supply crisis”

The once-sluggish central processor market is now facing what Futurum Group calls a “silent supply crisis,” predicting that CPU market growth could outpace GPU growth by 2028.

Major CPU providers AMD and intel The company has warned customers in China about the supply shortage, Reuters reported. According to reports, delivery times for CPUs have taken up to six months and prices have increased by more than 10%.

“The increase in demand has been unprecedented over the last six to nine months,” Forrest Norrod, head of data centers at AMD, said in an interview with CNBC.

Norrod said that while there is no “foreseeing that this situation will slow down or stop anytime soon,” AMD is anticipating increased demand and is “working diligently” to meet it.

An Intel spokesperson told CNBC that the company expects inventories to reach “lowest levels” this quarter, “but we are working aggressively and expect supply to improve in the second quarter through 2026.”

“Wafers don’t grow on trees,” Bajarin says. “You can’t just harvest 10% more silicon wafers. Demand is tight across the industry. So, unfortunately, CPU wafers are constrained.”

As for whether Nvidia has experienced any delays in CPU shipments, Harris told CNBC, “So far, things are going well.”

He said NVIDIA’s “robust supply chain” has been able to manage demand largely because many of the company’s CPUs are sold alongside GPUs in rack-scale systems.

Optimized for “feeding to GPU”

Harris said NVIDIA takes a fundamentally different approach in designing its CPUs to make them “ideal” for data processing and agent AI workflows compared to the more general-purpose CPUs made by industry leaders Intel and AMD.

The major difference is the number of cores on each CPU.

AMD’s EPYC series and Intel’s Xeon high-performance server CPUs typically have 128 cores, while Nvidia’s Grace CPUs have 72 cores.

“If you’re a hyperscaler, you want to maximize the number of cores per CPU, which essentially lowers your cost, the amount per core. So that’s one business model,” Harris explained.

Instead, Nvidia specifically designed the CPU to help its star GPUs run AI workloads.

“Single-threaded performance becomes much more important than the per-core amount because you’re trying to make sure you don’t have very expensive resources like GPUs waiting there,” Harris says.

Nvidia is also based on CPU arm Intel and AMD base their CPUs on the traditional x86 architecture, which is typically used for chips in low-power devices such as smartphones. Introduced by Intel nearly 50 years ago, x86 is the primary instruction set that has dominated PC and server processor design since its inception.

AMD’s Norod said Nvidia “I think they’ve optimized the chip very well for powering the GPU. It’s not optimized well enough for general-purpose applications.”

In fact, Nvidia relies on more versatile CPUs in some of its products. For example, Nvidia combines its GPUs with host CPUs from Intel or AMD in its HGX Rubin NVL8 platform, which customers use as building blocks for their own AI racks.

“Platform independent”

Nvidia’s foray into standalone CPUs comes as more customers are building their own Arm-based processors for data centers.

Amazon was the first major hyperscaler to launch its own CPU with the release of Graviton in 2018. googleAccording to Futurum Group, its Axion processors, released in 2024, currently handle about 30% of its internal applications. microsoft released its second generation Cobalt processor in November. Arm plans to release its own CPUs this year with Meta as an early customer.

Mercury Research estimates that Intel had 60% of the server CPU market share in the final quarter of 2025, AMD had 24.3%, and Nvidia had 6.2%, with the remaining share split between in-house Arm-based CPUs from hyperscalers like Amazon, Microsoft, and Google.

In the face of an insatiable need for computing, Nvidia typically takes a welcoming attitude toward competition. In keeping with that tradition, Nvidia opened its NVLink networking technology to third-party licenses in May.

The remainder of 2025 saw a flurry of NVLink deals with Intel. Qualcomm, FujitsuArm eases the path for third-party CPUs to integrate with Nvidia GPUs in AI servers.

These deals include CPUs made with Arm or x86 architecture, but Nvidia now also supports the open instruction set architecture RISC-V. RISC-V, which has gained momentum in recent years, allows companies to design custom processors without paying licensing fees to companies like Arm.

Nvidia signed a deal in January that will allow US chip company SiFive to use NVLink to connect its RISC-V chip designs to Nvidia GPUs.

Harris said Nvidia’s strategy remains “platform agnostic” no matter how CPU demand is met.

“While we do develop Arm-based CPUs, we have significant investments in the x86 community and ecosystem as a whole, so we will be in a strong position either way.”

Bajarin describes NVIDIA’s changing strategy as “soup to nuts.”

“To compete, NVIDIA’s answer is not: buy GPUs from us or buy something else,” Bajarin said. Whether it’s GPUs, CPUs or specialized hardware, “we need to scale our products to accommodate a variety of workloads,” he said.

WATCH: CNBC’s exclusive first look at Nvidia’s Vera Rubin AI system