For much of the AI boom, the spotlight stayed on GPUs. Demand for those processors surged first, and memory became the next major pressure point. CPUs, meanwhile, sat outside the main AI story because their relative lack of parallelization made them a poor fit for AI model inference.
That picture is changing. Amazon Web Services and its engineers have seen an explosion in wait times for CPU server capacity as AI workloads strain the company’s cloud infrastructure, while 2026 has brought a spike in CPU demand tied in large part to agentic AI.
Agentic AI Changes the Hardware Equation
Agentic AI systems do more than produce a single response. They can spawn agents that take on separate tasks, and those tasks can multiply across an enterprise. Matt Kimball, vice president and principal datacenter analyst at Moor Insights & Strategy, described the scale this way: “It’s one thing to have this agentic workload, and let’s say, it spawns 100 agents. If I’m going to roll this out across my enterprise, those 100 become tens of thousands, hundreds of thousands, or millions of agents.”
That expansion changes the demand calculation. One agentic workload may begin with 100 agents, but an enterprise deployment can grow to tens of thousands, hundreds of thousands, or millions of agents. Each agent needs to use a computer, and computers need CPUs.
The important point is that agentic AI relies on many steps beyond the AI model itself. A system must parse the model’s output, decide which tool to invoke, make an API call or run code, collect the result, and feed that result back into the process. Those actions create CPU work even when the broader task also includes AI model inference.
Souvik Kundu, senior staff research scientist at Intel, put the shift plainly: “Many components of an agentic AI task are inherently CPU based jobs.” The growing use of agents therefore connects AI workloads to a part of computing infrastructure that received less attention during the first phase of the boom.
Why CPU Capacity Matters Again
The CPU’s role in an agentic pipeline is practical and repeated. It handles the instructions and coordination that let an agent interact with a computer, use a tool, run code, and process what comes back. The CPU does the job of parsing output, figuring out which tool to invoke, making the API call or running the code, collecting the result, and feeding it back.
Madhu Rangarajan, vice president of compute and enterprise AI at AMD, described the balance found in testing: “in our testing, seven of the eight stages in realistic agentic AI pipelines run entirely on the CPU.” That figure helps explain why CPU capacity has become a pressing issue even as GPUs remain central to AI model inference.
The hardware story is no longer just about choosing the processor best suited to run an AI model. It also involves handling everything around that model, especially when one workload creates many agents and each agent must perform a chain of computer-based actions. Seven of the eight stages in realistic agentic AI pipelines running entirely on the CPU puts that supporting work at the center of the discussion.
AWS’s wait times for CPU server capacity show how this demand is reaching cloud infrastructure. The strain is not limited to the processors used for model inference; it also affects the computing capacity needed to coordinate agents and complete their tasks.
The result is a broader AI hardware picture. GPUs helped drive the first surge, memory followed, and CPUs are now facing their own demand spike as agentic systems move from individual workloads toward enterprise-scale use. The more those systems expand, the more computer-based steps they create, and the more CPU capacity those steps require.
Agentic AI has changed the question from “Which processor runs the model?” to “Which processors handle the full chain of work?” For 2026, that distinction is putting CPUs back at the center of AI infrastructure.
