AI agents need more than GPUs

An AI agent checking an invoice does more than ask a model for an answer.

The agent retrieves the purchase order, checks company policy, and sends exceptions to a person.

Each step adds time and cost.

On November 10, AMD announced general availability of its sixth generation EPYC SP7 server processors. AMD says partners are already shipping systems.

One example of how AMD pairs CPUs and GPUs is Helion Air Inference, which combines agent execution and local GPU inference in the same rack.

Here’s how the work breaks down:

• Training: CPUs prepare and feed data. GPUs handle the heavy math used to train the model.
• Inference: GPUs run the model. CPUs process requests, retrieve data, and return results.
• Agentic AI: CPUs also coordinate the steps between model calls. They run tools, check permissions, and manage isolated environments where agents execute tasks. More CPU capacity helps when those steps limit how many agents your system handles at once.

AI is only part of the job.

General purpose computing remains the core role for these server CPUs.

Cloud services, databases, enterprise applications, and many scientific workloads still rely on CPUs, including in data centers without GPUs.

My takeawaySize your infrastructure for the work you need to complete.

• For training, measure how often GPUs wait for data and what each run costs.
• For inference and agents, track end-to-end response time and cost per correctly completed task.
• For traditional workloads, measure performance and cost against the service levels your business needs.

At your next infrastructure review, ask your team: Where is our biggest bottleneck, and what measurements justify our next investment?