According to a research report from Guotai Haitong Securities Co., Ltd., the CPU is reclaiming its position as the core hub of AI infrastructure, with heterogeneous collaboration driving a revaluation of industry value. As AI advances towards inference and agentic applications, computational bottlenecks are shifting from GPU processing to CPU scheduling and execution. The CPU's role in task planning, data processing, KVCache management, tool invocation, and multi-agent collaboration is becoming increasingly prominent, leading to a reassessment of its valuation and creating supply-demand imbalances. This may position the CPU as the next scarce and premium component following memory. Over the next two to three years, tight CPU supply and pricing pressures are expected to persist. The main viewpoints of Guotai Haitong Securities Co., Ltd. are as follows:

AI inference and agent workloads are driving a structural inflection point in CPU demand. With the growth in inference requests, multi-user concurrency, long contexts, and multi-agent collaboration, CPU load has significantly increased. The CPU-to-GPU ratio has evolved from 1:8 during the training phase to 1:3–1:4 in the inference phase, approaching 1:2 in the agent phase. Rising CPU prices and extended delivery times have preliminarily validated the resurgence in demand. In the future, AI data centers will shift from simply stacking GPUs to holistic optimization of "CPU + GPU + memory + network + software."

Heterogeneous systems and new CPU architectures present long-term investment opportunities. The increasing complexity of AI workloads is driving CPU architecture towards system-level collaborative optimization, including more cores, high-bandwidth caching, CXL interconnects, and new shared memory architectures. CPU-GPU collaboration is becoming the core of AI data centers, enhancing the value of the industrial chain. The acceleration of domestic production in upstream semiconductor materials and equipment, midstream chip design and packaging/testing, and high demand in downstream servers, AI acceleration, and edge scenarios create long-term investment opportunities across the entire CPU industry chain.

Domestic CPUs are entering a strategic layout window in the industry chain. Internationally, Intel, AMD, and ARM dominate the market by optimizing architectures, core counts, and heterogeneous collaboration capabilities, with AI inference and agent workloads continuing to drive demand for their server CPUs. Domestic CPU manufacturers, such as Phytium and Hygon Information Technology Co., Ltd., are accelerating their strategic layout. Through proprietary instruction sets, the RISC-V ecosystem, and domestic EDA/IP support, they are achieving breakthroughs in the server and AI computing markets.

Risk warnings: Risks include slower-than-expected adoption of AI agents, slower-than-expected commercialization of AI, intensified industry competition and market share loss, and fluctuations in capital expenditure cycles.