For much of the recent AI infrastructure buildout, procurement has revolved around securing accelerators. GPU availability, memory capacity and interconnect performance have shaped purchasing decisions.
The growth of AI agents is expanding that discussion. Running a model is one part of an agent’s workload. Executing its instructions—running code, accessing business applications and processing results—creates a separate requirement for computing capacity.
For data center buyers, this raises a practical question: does the next expansion require more inference capacity, more execution capacity, or both?
The market signal: rising demand for server CPUs
In research published on October 22, 2026, TrendForce forecasts server CPU shipments to increase 31% in 2026 and more than 40% in 2027, linking that growth to inference demand and the adoption of AI agents.
These figures are analyst forecasts, rather than completed annual results. They point to increasing CPU demand across AI infrastructure, but do not establish availability or lead times for a specific Intel or AMD processor. Source: TrendForce, October 22, 2026.
The implication for procurement is broader than securing additional processors. Buyers need to understand where CPU demand will appear: inside GPU servers, in separate execution clusters, or across the application infrastructure supporting the service.
Why agent execution changes the hardware requirement
An agent completing a business task may alternate between model inference and actions in other systems. A coding workflow, for example, can involve generating code, executing it, running tests, inspecting failures and requesting another model response.
Several of those steps depend on the result of the previous step. Increasing inference throughput cannot remove time spent waiting for code execution or other dependent operations.
NVIDIA’s October 21 technical discussion of its OpenShell architecture highlights this distinction. It describes agent execution as a combination of sequential steps and concurrent workflows, requiring both strong individual-core performance and sufficient capacity to serve many tasks at once. Source: NVIDIA, October 21, 2026.
For operators, the purchasing implication is that two services using the same model can require different supporting infrastructure. A service generating short responses and an agent repeatedly compiling and testing software should not automatically receive the same CPU-to-GPU allocation.
AMD’s Helion announcement illustrates two deployment approaches
On October 12, AMD introduced Helion, an open blueprint for agentic AI infrastructure.
The announcement describes CPU-only racks for agent execution connected to separate model inference infrastructure, alongside integrated configurations combining CPU and GPU nodes. AMD said partner configurations were under development and planned for availability in November 2026. That planned timing should be distinguished from confirmed systems available for delivery today. Source: AMD, October 12, 2026.
Our interpretation is that execution capacity is becoming a purchasing decision in its own right.
An operator may expand the CPU infrastructure supporting an existing GPU fleet. An enterprise using externally hosted models may still need local servers to execute tools and process application data. Other deployments may benefit from acquiring CPU and GPU capacity together.
These are different requirements, even when all three projects are described as “AI infrastructure.”
What should change in the next RFQ
The most useful starting point is the workload at its intended concurrency.
A lightly loaded demonstration can hide resource contention that appears when many agents run simultaneously. Before choosing the next configuration, buyers should establish whether delays come from CPU execution, memory pressure, storage access, network communication or the inference service itself.
That distinction changes the specification:
- CPU execution limits: compare task completion time and throughput under realistic concurrent load. Core count alone is insufficient.
- Memory limits: specify the capacity required to keep the intended number of execution environments active.
- I/O limits: identify the actual data path and required connections between execution nodes, storage and inference infrastructure.
The acceptance criterion should connect these resources to the service being delivered: how many tasks can the deployment complete successfully within its latency target?
A faster accelerator remains valuable when inference is the limiting stage. Additional CPU capacity matters when execution is limiting progress. Neither purchase necessarily solves delays caused by an external service.
What this means for buyers and sellers of existing equipment
For buyers, this development creates a reason to evaluate available CPU servers alongside accelerator offers.
A compatible server configuration may provide execution capacity without requiring an expansion of the GPU fleet. Its value still depends on workload performance, power consumption, memory configuration, support and integration costs. A new use case does not make every older platform suitable.
For sellers, complete configurations deserve consideration before systems are separated into components. A buyer expanding an execution cluster may value a consistent group of servers with documented memory, storage and networking. An integrator with existing platforms may prefer individual CPUs or component lots.
The commercial opportunity lies in matching equipment to a defined requirement. Broad CPU demand forecasts cannot determine the resale value of an individual lot.
At REVO.tech, we connect businesses sourcing and selling surplus servers and components. Share your required platform, memory configuration, quantity and deployment window—or the specifications of equipment you are looking to sell.
As AI agents move into production, the procurement brief needs to cover the infrastructure that executes their work as carefully as the hardware that runs their models.
For sourcing and surplus enquiries: [email protected]