TechPro
Harju maakond, Tallinn, Kesklinna linnaosa, Tartu mnt 25-46, 10117 smarttek.ou@gmail.com

For the past few years, the narrative in tech has been simple: GPUs are king, and CPUs are gradually taking a backseat in the AI revolution. However, as the industry shifts from simple chatbots to autonomous AI agents, that narrative is quickly unraveling.

According to new insights from semiconductor research firm TrendForce, the rapid adoption of agentic AI is breathing new life into the traditional CPU—and could soon trigger a fourfold surge in data center CPU core requirements.

Why AI Agents Need CPUs More Than Chatbots Do

Traditional Large Language Model (LLM) inference relies heavily on GPUs to crunch massive amounts of data and generate responses. Autonomous AI agents, however, operate very differently. Instead of just answering a single prompt, agents execute complex, multi-step workflows on their own—and that requires a completely different kind of computing power.

While GPUs handle the heavy lifting of model training and core inference, CPUs act as the orchestrators. They are responsible for the logic-heavy, sequential tasks that GPUs simply aren’t built to handle efficiently, including:

  • Planning complex tasks and breaking them down into subtasks

  • Calling external APIs and software tools

  • Managing communication between multiple AI agents

  • Running Python scripts, querying databases, and crawling the web

In short, the CPU is evolving from a basic supporting component into a mission-critical pillar of modern AI infrastructure.

A Dramatic Shift in Data Center Hardware

Historically, AI-focused data centers have been heavily skewed toward graphics processors, typically deploying one CPU for every four to eight GPUs.

Agentic AI is set to upend that ratio. Citing estimates from Arm, TrendForce notes that as agentic workloads scale, data center CPU requirements could skyrocket from roughly 30 million cores per gigawatt of capacity to 120 million cores per gigawatt. Consequently, the standard server architecture is expected to shift closer to a 1:1 or 1:2 ratio of CPUs to GPUs.

Avoiding the Next Big Bottleneck

Ignoring CPU capacity in an agent-driven world doesn’t just limit scale—it kills performance. Research highlighted by TrendForce shows that CPU-based tool processing can account for up to 90.6% of total latency in certain agentic workflows. Furthermore, dynamic energy consumption from CPUs is becoming a much larger piece of the overall power puzzle.

Simply throwing more GPUs at the problem won’t make an AI system faster if the CPUs coordinating the tasks become a bottleneck.

Even Nvidia is Betting Big on CPUs

Perhaps the clearest proof of this shift comes from Nvidia itself. For years, the chipmaker positioned the GPU as the ultimate solution for virtually every computing workload, often downplaying the role of traditional processors.

That strategy has noticeably changed. Following its Grace processor, Nvidia is now pushing its Vera generation—a chip featuring 88 CPU cores and 176 threads, linked to accelerators via high-bandwidth NVLink-C2C interconnects. Even more telling, Nvidia is developing dedicated Vera CPU racks, allowing data centers to deploy standalone CPU clusters independent of GPU-heavy systems.

The Takeaway

We are witnessing a fundamental evolution in AI computing. As artificial intelligence moves beyond generating isolated text or images and starts autonomously executing real-world workflows, the hardware behind it has to adapt. GPUs may be the muscle of the AI era, but CPUs are rapidly proving to be its indispensable nervous system.

Share:

Leave a Reply

Your email address will not be published. Required fields are marked *