Agentic AI Could Quadruple Demand for Data Center CPU Cores

Agentic AI could quadruple CPU core demand, rising from approximately 30 million cores per gigawatt of data center capacity to 120 million, according to Arm estimates cited by TrendForce.

TrendForce says GPUs remain essential for model training.

CPUs coordinate the complex action sequences that make agents operate autonomously, according to TrendForce.

That matters because operators may need many more CPUs.

They may also need a different balance between CPUs and GPUs in server infrastructure, according to TrendForce.

The shift could affect system performance, energy consumption and infrastructure costs, according to TrendForce.

Why do AI agents need more CPUs?

TrendForce says agentic AI creates a 'more CPU-intensive workload' than conventional large language model, or LLM, inference.

TrendForce says GPUs still execute model training.

TrendForce says CPUs handle the coordination work around those calls.

According to TrendForce, agents rely on CPUs and conventional computing resources for these tasks:

• Planning tasks.
• Scheduling subtasks.
• Calling external tools.
• Managing interactions between multiple agents.
• Evaluating results.

TrendForce says agents also use conventional computing resources for routine execution.

They run Python code, search databases, crawl websites and process information, according to TrendForce.

GPUs cannot perform those coordination steps alone, according to TrendForce.

TrendForce therefore presents CPUs as a significant part of AI infrastructure, not merely support for GPUs.

That marks a change in infrastructure design, according to TrendForce's analysis.

TrendForce's analysis links the change to a simple shift: as agents move from generating responses to completing multistep tasks, the processor handling planning, tools and evaluation becomes central to each workflow.

How much could CPU demand change?

Traditional AI data centers typically deploy one CPU for every four to eight GPUs, according to TrendForce.

That ratio reflects TrendForce's cited industry emphasis on accelerated, compute-heavy AI workloads.

TrendForce expects agentic infrastructure to move toward one CPU for every one or two GPUs.

Arm estimates cited by TrendForce put baseline core needs at approximately 30 million per gigawatt.

They project 120 million cores per gigawatt as agentic workloads become more prevalent.

That is a potential fourfold increase in CPU core demand, according to TrendForce.

Under TrendForce's forecast, the new ratio would place processors much closer to accelerator density.

It would also exceed CPU deployment in many existing AI installations, according to TrendForce.

The forecast does not require every operator to reach the same ratio immediately, because TrendForce presents it as an outcome tied to the wider adoption of agentic workloads.

For data center operators, GPU count alone becomes an incomplete measure under TrendForce's forecast.

TrendForce's analysis makes CPU core density and processor-to-accelerator balance part of the same infrastructure decision.

TrendForce's figures summarize the expected change:

Metric	Existing AI data center pattern	TrendForce's agentic outlook
CPU-to-GPU ratio	Approximately 1 CPU per 4 to 8 GPUs	Toward 1 CPU per 1 to 2 GPUs
CPU cores per gigawatt	Approximately 30 million	120 million
Core-demand change	Baseline	Potential fourfold increase

Where can CPUs become bottlenecks?

Research cited by TrendForce finds CPU-based tool processing can reach 90.6% of total latency in some agentic workloads.

That makes CPU speed central to end-to-end agent performance, according to TrendForce.

CPU dynamic energy consumption can also form a substantial share of total system energy use under certain conditions, according to TrendForce.

TrendForce says more GPU capacity may not improve performance when coordination processors become bottlenecks.

Adding accelerators alone therefore does not guarantee a faster overall system, according to TrendForce.

Balanced resources could ease latency, energy and cost pressures, according to TrendForce's analysis.

This shifts attention from the speed of individual AI components to the speed of the complete agent workflow, because planning, tool execution, coordination and evaluation all occur around model computation.

That system-level view matters as operators compare energy use and infrastructure costs, according to TrendForce.

What is Nvidia doing with CPUs?

Network World points to Nvidia as a clear sign of growing CPU demand.

Network World reports that the chipmaker spent years downplaying and even disparaging CPUs.

Nvidia had positioned the GPU as the solution for nearly everything, according to Network World.

Network World says Nvidia changed course a few years ago with the Grace processor.

Grace now appears in Nvidia's Vera generation, according to Network World.

Nvidia says Vera features '88 CPU cores and 176 threads.'

The platform includes a high-bandwidth NVLink-C2C interconnect, according to Nvidia.

It supports communication and memory sharing with Nvidia accelerators, according to Nvidia.

Vera is designed to work closely with Nvidia's GPU platforms, according to Nvidia.

Nvidia is also developing dedicated Vera CPU racks, according to the company.

Customers can deploy those CPU systems independently of GPU-heavy or GPU-only systems, according to Nvidia.

Under Nvidia's rack plan, customers gain a CPU route that does not require a dense GPU configuration.

The standalone rack strategy fits TrendForce's forecast that agents will require more CPU capacity, because it treats the processor as useful infrastructure in its own right rather than only as an accessory to an accelerator.

Network World's report draws the same conclusion: CPUs handle work that GPUs cannot.

What changes for AI infrastructure?

TrendForce frames a broader change in AI computing.

TrendForce says systems are moving from generating individual responses toward executing multistep tasks autonomously.

Conventional LLM inference concentrates on producing an answer, according to TrendForce.

Agentic systems add planning, subtask scheduling, tool calls, inter-agent management and result evaluation, according to TrendForce.

That extra orchestration shifts CPU work from the background into the main control path, according to TrendForce.

TrendForce ties the market stakes to CPU deployment, system performance, energy use and infrastructure costs.

Data center operators may need more cores, different server ratios and better control of coordination bottlenecks, according to TrendForce.

Arm's estimate shows the potential scale of that redesign.

Nvidia's Vera products show how one supplier is preparing for a larger CPU role.

Neither Arm's estimate nor Nvidia's roadmap settles the pace of adoption.

TrendForce's forecast depends on agentic workloads becoming more prevalent in AI infrastructure.

The next operational signal will be CPU deployment inside AI data centers, according to TrendForce's outlook.

Operators moving toward one CPU for every one or two GPUs would confirm TrendForce's central premise.

If they do, Nvidia's Vera racks could become a template for deploying CPU capacity independently of GPU-heavy systems, according to Nvidia's roadmap.

For the AI market, TrendForce's forecast shifts the central issue beyond how many accelerators a data center can buy and toward how much processor power agents need to turn model outputs into completed work.

TrendForce's 120-million-core estimate sets the scale; Nvidia's Vera generation supplies the product response.

TrendForce's forecast faces its next test as agentic workloads become more prevalent and operators decide whether CPUs deserve a larger share of each data center.