After the Q1 2026 earnings report, AMD CEO Lisa Su revealed during a conference call that the demand for CPUs in data centers is rapidly increasing with the arrival of the Agentic AI era. Su pointed out that the number of CPUs and GPUs in a single computing node is shifting from a traditional one-to-many ratio to approaching a one-to-one ratio, and in the future, there may even be more CPUs than GPUs in a single node.

She explained that in the past, AI training and inference typically used a configuration of "one CPU paired with four to eight GPUs," with the CPU mainly responsible for scheduling and initiating GPU computing tasks. However, in an Agentic AI model, a large number of autonomous agents rely on the CPU for state updates and task coordination, a change that is profoundly affecting the form of computing nodes. As the number of agents grows rapidly, the ratio of CPUs to GPUs is moving closer to 1:1.

Su further suggested that if a large number of agents run in a cluster in the future, it is entirely possible to achieve a configuration where the number of CPUs in a single node exceeds that of GPUs. This means that the previous trend of GPU-dominated computing expansion is being replaced by a CPU demand trend driven by "agent workloads." Agentic AI refers to running multiple autonomous agents on top of large language models to automatically complete complex tasks. For example, in software development, agents can review code, make modifications, compile, wait for feedback, and continue fixing bugs when they find them, with almost no human intervention required.

To effectively coordinate and schedule these parallel-running agents, the system must rely on the CPU to provide continuous control and management functions. Under this new type of workload, the CPU is no longer just a "starter" for GPU training or inference, but has become the core power of the entire Agentic AI system. Although GPU-accelerated computing is still expanding rapidly, CPU utilization is also continuously increasing.

AMD stated that almost all CPUs currently available have already been sold to AI labs and hyperscale cloud service providers to meet the surge in demand from agent tasks. This trend indicates that in future AI infrastructure design, the relationship between CPUs and GPUs may shift from "master-slave" to a more equal one, and even the CPU usage density may be higher.

Key Points:

🖥️ The ratio of CPUs to GPUs is gradually approaching 1:1, and in the future, there may be more CPUs than GPUs.

🤖 In the Agentic AI era, agents rely on CPUs for state updates and task coordination, changing the form of computing nodes.

📈 AMD is currently selling almost all of its CPUs to AI labs and cloud service providers to meet the surging demand for agent tasks.