As AI agents spread, the AI infrastructure landscape that had been built around GPUs is shifting toward CPUs.

Amazon News shared why CPUs matter in the age of AI agents, describing how the CPU role is growing in agent environments.

Amazon News said the discussion of AI infrastructure has previously centered on chips used to train large models. AI accelerators such as Trainium and GPUs are optimized to process large-scale data in parallel, making them suited to model training.

Agent AI is different. If an LLM is like a calculator that takes prompts and produces outputs through parallel computation, an AI agent is like a manager that autonomously completes multi-step tasks. For example, if asked to "research this company and write a report," an agent breaks the goal into steps, opens a web browser to explore links, analyzes files and runs code to produce results.

Logical processing other than inference, file management, network calls and code execution are all tasks handled by CPUs.

AWS Graviton is a CPU designed for always-on, low-latency workloads like these.

Agent AI quickly repeats data lookups, tool calls, action execution and evaluation of the next step, and Graviton is designed to minimize the time it takes for each part inside the processor to exchange data, Amazon News reported.