CITIC Securities: Agentic AI reshapes the value of general computing.

CICC released a research report stating that the paradigm evolution of artificial intelligence from pre-training to reinforcement learning and Agentic AI has significantly increased the demand for CPUs. When the concurrent volume of Agents exceeds the traditional CPU carrying limit, problems such as request queue stacking, tool call timeout, and delay in multi-round reasoning loops will occur. In different types of intelligent agent tasks, CPUs have a higher proportion in terms of running time and energy consumption compared to GPUs, becoming the main bottleneck for Agent operation. In the past, CPUs have been continuously marginalized in the value distribution of the computing power industry chain. With the sharp increase in CPU demand brought by Agentic AI, coupled with upstream price increases and limited capacity, the global CPU market has seen continuous price increases and shortages this year. Arm CEO Haas points out that traditional AI workloads require around 30 million CPU cores per GW of data center capacity, while in the Agentic AI era, the demand for CPU cores will surge to 120 million per GW. At the same time, Intel stated in its Q1 2026 earnings call that the future CPU-to-GPU ratio will increase from the current 8:1 to 1:1.