The rise of personal AI agents is creating an unexpected beneficiary in the artificial intelligence boom: the humble central processing unit.
For much of the generative AI era, Nvidia’s graphics processing units have captured the attention, investment, and spending of the technology industry because of their ability to process the enormous volumes of computation required to train and run AI models. Now, the emergence of AI agents capable of working autonomously for hours at a time is shifting part of that workload back toward CPUs, giving AMD and Intel a new opportunity in a market they spent years watching Nvidia dominate.
The change is already showing up in investor enthusiasm. AMD and Intel shares have risen 32% and 21%, respectively, over the past month, outperforming the technology sector’s megacap companies. AMD’s expanding position in both CPUs and GPUs has also pushed the company into the trillion-dollar market capitalization club.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
According to a CNBC report, the catalyst is the rapid development of agentic AI systems. OpenAI recently released its AI agent Dots, following Meta’s launch of Muse in early September. Muse gained traction quickly enough to reach the top of Apple’s App Store in less than two weeks.
Unlike conventional chatbots, these systems are designed to execute multistep tasks autonomously. They can plan actions, interact with software, and continue working in the background, sometimes for hours or days. That has created a different infrastructure requirement from simply responding to a prompt.
“As more agents are created and developed, and more people start to use them for more tasks, it’s going to start to shift the workload away from GPUs and onto CPUs,” said Ryan Shrout, president of Signal65, which advises companies on AI hardware and costs.
Analysts expect that shift could give AMD and Intel a fresh growth market precisely as the AI infrastructure industry begins to move beyond the initial GPU-centric phase.
Why Agents Need CPUs
The distinction between CPUs and GPUs is becoming increasingly important as AI agents evolve.
GPUs remain crucial for inference, the process of running AI models to generate responses and make decisions. But an autonomous agent needs more than the model itself. It also needs a computer capable of executing the instructions generated by that model, running applications and maintaining tasks in the background.
“CPUs are actually performing the workflows while GPUs are doing the thinking,” said Daniel Newman, CEO of Futurum Group.
That has created a complementary relationship rather than an immediate replacement for Nvidia’s GPUs. The more successful AI agents become, the more computing infrastructure may be required around them, including CPUs capable of handling the operational workload generated by each agent.
Users can get a glimpse of that infrastructure simply by asking the agents what powers them.
Muse tells users that it runs on an AMD-powered computer, while Dots says it runs on a virtual computer powered by AMD’s EPYC CPU. Dots has been observed using an AMD EPYC 9V74 processor, while the EPYC 9D25 has reportedly been used by Meta for Muse.
Meta, however, says it is not dependent on AMD.
“We take a diverse approach to our hardware and are largely CPU-agnostic by design, which gives us the most flexibility in acquiring capacity,” a Meta spokesperson said.
OpenAI likewise uses multiple CPU suppliers, meaning the growth of agents does not automatically translate into an AMD monopoly or even a guaranteed AMD win.
Still, the economics are attractive.
AMD’s EPYC processors can contain as many as 192 CPU cores. Benchmarks and queries suggest Muse’s virtual machine uses two cores while Dots uses nine. Because cloud providers can divide physical servers into virtual machines, a single server can support numerous agent users simultaneously.
That is considered a game changer because CPUs are significantly cheaper than high-end GPUs.
The EPYC 9V74 used by Dots is available from resellers for less than $3,000, while the EPYC 9D25 reportedly used by Meta for Muse costs even less on the secondary market. Nvidia GPUs can cost more than 10 times as much for a single chip and are typically deployed in clusters containing hundreds or thousands of processors.
But the economic implication is significant. This is because if an AI agent spends hours working on behalf of a user, the cost of maintaining the computing environment in which that work occurs can become a major component of the service’s economics.
Meta already faces that calculation with Muse. Sensor Tower estimates that downloads have surpassed 5 million since its launch, while Morgan Stanley estimates the cost of serving each user at between $3 and $130 per month, with an average of $37 depending on inference usage.
Morgan Stanley estimates Muse could account for 20% of AMD’s 2026 chip sales.
AMD Has An Early Advantage, But Nvidia is Not Out
AMD’s position is especially interesting because it can supply both CPUs and GPUs, giving it exposure to different parts of the AI computing stack. Its data-center business has already become a major growth engine. Revenue in the division more than doubled to $6.7 billion in the quarter ended June, representing almost 60% of AMD’s total sales.
“The conversation has changed quite a bit over the last eight to 10 months in terms of agentic usage,” said Dan McNamara, AMD’s senior vice president and general manager of compute and enterprise AI. He said CPU sales are going to “really ramp.”
AMD also has an important advantage in the hyperscaler market, where many of the virtual machines supporting AI agents will be deployed. Amazon, Google, Meta, and Microsoft all use CPUs from AMD or Intel across their server infrastructure.
AMD’s x86 architecture gives it another advantage because software written for existing x86 systems can generally continue running without major compatibility changes. That matters for enterprises that want to deploy agents without rebuilding large portions of their software environments.
AMD currently controls about 46% of the x86 CPU unit market, according to Mercury Research.
“It’s just a new product and they’re going to be competing against a lot of different players,” said Jordan Klein, an analyst at Mizuho Securities, who believes AMD currently has the “best product.”
Klein said AMD is “probably getting the most of this because they have the largest market share and the cloud hyperscaler world for CPUs,” although he added that he “wouldn’t rule Nvidia out.”
Nvidia is already positioning itself for the same opportunity. Earlier this year, it introduced Vera, a redesigned central processor, alongside an entire rack built around CPUs. The company says Vera was designed specifically for agentic workloads and expects CPUs to become a $200 billion market by 2030. That puts Nvidia in a new competitive position. The company has spent the past several years building an ecosystem around GPUs, but its enormous financial resources and existing relationships with cloud providers mean it can also attack markets that emerge around its dominant franchise.
Arm is another potential challenger. The company announced its own CPU for AI agents in March, with Meta as its first customer. Arm’s stock has more than doubled since then, suggesting investors are already pricing in expectations that agentic computing could expand the market for alternative processor architectures.
A New AI Infrastructure Cycle
The size of the opportunity is beginning to change forecasts for the CPU industry.
Futurum estimates total CPU sales will reach $118 billion in 2027, almost twice its previous forecast from May. AMD went further in July, predicting that the entire CPU market could reach $220 billion in annual sales by 2030, compared with its previous 2025 forecast of $60 billion. AMD expects to capture more than half of that market.
Those forecasts depend heavily on whether agentic AI becomes a mainstream computing model rather than remaining a collection of consumer experiments.
If millions of users begin delegating tasks to autonomous agents, the infrastructure requirements could be substantially different from those of today’s chatbot market. Each user may effectively require access to a persistent virtual computer capable of running applications, storing temporary information and executing tasks while the underlying AI model handles reasoning.
Analysts say that could result in a much broader market for CPUs without eliminating demand for GPUs.
“If you are a company that can supply both types of chips, you’re in the best position,” Klein said.
That is increasingly the strategic advantage AMD is developing. Nvidia still controls the most important layer of AI computing, particularly high-performance GPUs used for model training and inference. But agentic AI is adding another layer of computation around those models, and that layer is far more CPU-intensive.
For AMD and Intel, the significance is not simply that CPUs are selling again. It is that the next phase of AI could require a much larger and more diverse computing infrastructure than the first.
The generative AI boom made GPUs the defining hardware of the industry. Now, industry experts believe the agentic AI boom may make the relationship between CPUs and GPUs just as important. And if autonomous software becomes a persistent part of everyday computing, AMD and Intel could benefit from a market in which the processor that executes the work becomes nearly as important as the processor that does the thinking.