Intel says the agent era is pushing data center CPU-to-GPU ratios toward 1:1

At its Intel Connection technology and industry ecosystem conference in Suzhou, Intel laid out a new view of AI computing: in the agent era, the balance of CPUs to GPUs in the data center is shifting from roughly 1:4 toward nearly 1:1, and the value of the CPU needs to be re-evaluated.

Intel CEO Lip-Bu Tan said in a video address that x86 remains the foundation of modern computing and continues to drive computing toward intelligent applications. Chen Baoli, vice president of Intel's data center group and general manager for China, and Gao Song, vice president and China general manager of the edge computing and physical AI group, spoke on stage, while Justin Yifu Lin, dean of Peking University's Institute of New Structural Economics, discussed AI's economic impact.

Chen used a simple "summarize the news" task to explain why CPUs are becoming scarce. A user gives the agent one sentence of instruction — collect yesterday's AI news every morning at 8 — but the agent must fetch web pages, PDF reports and even video, extract and process those formats, and only then hand the material to a large model. Much of that preprocessing work, he said, suits the CPU.

In his framing, a data center has three cooperating parts: CPU servers or clusters that execute and orchestrate agents, GPU and other accelerators that run model computation, and storage that holds task data, conversations and newly generated documents. CPU work runs through all of them, from data preprocessing for accelerators to data scheduling for storage.

The question becomes concrete at procurement time. Chen said conversations over recent months with Chinese software partners and internet companies showed customers care both about how many agents a system can run and about the capability and reliability of each agent, with some workloads needing many sandboxes started in a very short window. Launching a large number of agents and keeping that number working reliably are two different things, which pushes CPU evaluation toward concurrency, response time and support for real tasks.

Intel's answer is to optimize CPUs and supporting software together inside real workloads. Chen described using the built-in acceleration in Xeon processors to speed up data preprocessing, and KV cache optimizations that combine lossless compression with faster data transfer as agents repeatedly process context. The company is also working with Chinese software partners on high-performance storage systems to keep CPUs and GPUs from sitting idle, and it repeatedly pointed to open-source projects including Linux and PyTorch.

The same split shows up at the edge. Gao described CPU, GPU and NPU working together: the CPU handles planning and decisions that need quick responses, the GPU handles matrix-heavy AI computation, and the NPU covers long-running audio, video and computer-vision workloads at lower power. He named home AI, in-car AI, devices running on private data, and robots that must sense their environment, plan tasks and control motion in real time.

Justin Yifu Lin argued that AI's economic effect will show up through industries using it to raise efficiency. For Intel, that means whether partners build products with Xeon, Core and its software stack. QbitAI also noted that Intel's shares jumped 12% the previous day, read as a market response to the CPU's role in the agent era. The actual CPU-to-GPU ratio still depends on the task and system configuration, and how much of this business Intel wins will depend on the solutions it delivers.