The global semiconductor market is undergoing a structural transformation, with the CPU sector, traditionally viewed as a mature category, becoming a focal point for capital markets. As of January 21st, Intel's stock price hit a nearly four-year high, with a year-to-date surge exceeding 44%; AMD continued its upward trajectory; in the A-share market, Loongson Technology and Hygon Information Technology recorded a 20% limit-up and a single-day gain of over 13%, respectively. This market movement reflects a repricing based on the transmission logic of the "computing tax." Following the significant price increases for GPUs driven by AI training demand, CPUs are now emerging as the second wave bearer of rising computing costs.

A consensus among institutions is rapidly forming. Guolian Minsheng Securities and Western Securities pointed out concurrently in recent reports that the current shifts in CPU market supply and demand are not cyclical fluctuations but a structural change driven by the large-scale deployment of AI agent applications. Unlike AI training centered on GPUs, in agent workloads, CPUs handle a vast amount of non-AI-native computations, including tool invocation, task orchestration, and real-time decision-making. These processes account for a staggering 80%-90% of total task latency. This makes the CPU a potential system-level performance bottleneck even before the GPU. The demand outlook is already supported by data. IDC predicts that the number of active AI agents globally will surge from approximately 28.6 million in 2025 to 2.216 billion by 2030, representing a compound annual growth rate of 139%. Under a neutral scenario, the corresponding long-term CPU demand could exceed 11.73 million units, creating a substantial incremental market. The supply side for CPUs is also under extreme pressure. J.P. Morgan data indicates that Intel's advanced process capacity utilization has reached an overloaded state of 120%-130%, while the bottleneck in TSMC's advanced packaging capacity has extended CPU delivery cycles from the normal 8-10 weeks to over 24 weeks. Amid this trend, domestic CPU manufacturers are encountering dual opportunities from both industry dynamics and policy support. The CPU, long perceived as a "traditional" computing component, is re-establishing its system-level value in the wave of AI agents. AI Agents Catalyze Reshaping of "Peripheral CPU" Demand Traditional AI computing placed the entire computational burden solely on the GPU, primarily for model training and inference acceleration. However, as AI evolves towards agents capable of autonomous planning and execution, the structure of computational loads is undergoing a fundamental restructuring. For an agent to complete a practical task, such as "analyzing a batch of resume data," its workflow far exceeds a simple API call. It needs to autonomously execute steps like: creating an independent sandbox environment, accessing a specified cloud drive to download files, decompressing archives, running data analysis scripts, generating visual reports, and finally cleaning up and releasing environment resources. In this complete task chain, only the task decomposition and result generation phases rely on the GPU for inference. The intermediate steps, constituting 80%-90% of the entire process duration—including file operations, code execution, data processing, and system scheduling—are all handled by the CPU. A white paper published by Intel, titled "A CPU-Centric View of Agent AI," explicitly states that the latency in agent workloads primarily originates from tool-processing tasks on the CPU side. Agent Architecture Paradigm Unifies: Mainstream Platforms Shift En Masse to "Sandbox Execution" Model As AI agents transition from concept to large-scale application, the underlying industrial technical architecture is undergoing a fundamental restructuring. According to industry research by Guotai Junan Securities' Haitong Electronics, since the second half of 2025, major AI platforms, including Doubao and Zhipu, have fully transitioned to a "sandbox execution" architectural model. The core of this model involves creating an independent, isolated virtual execution environment for each agent task to safely perform external calls like file operations, code execution, and network access. This architectural shift directly gives rise to new computing demand characteristics: CPU resource consumption is strongly correlated with user scale and task concurrency, but only weakly linked to the number of GPUs. Breakthroughs in engineering practices provide crucial technical support for this architectural evolution. The DeepSeek research team demonstrated a milestone "compute-storage separation" solution in a paper: successfully storing a 100-billion-parameter embedding table entirely in the CPU's host memory, instead of the traditional GPU VRAM. Through a sophisticated PCIe asynchronous data transfer mechanism, this solution introduces less than 3% of additional inference latency, achieving a critical breakthrough in engineering feasibility. This technological breakthrough reveals two major industry trends: technically, the dependency of model parameter scale on GPU VRAM capacity is effectively broken, making more cost-effective host memory a viable option for large-scale parameter storage; architecturally, the role of the CPU undergoes a fundamental change, transforming from an auxiliary computing unit into the core hub for data scheduling and system management, taking on key functions like real-time retrieval, intelligent filtering, and efficient forwarding of massive parameters. Supply-Demand Imbalance Accelerates Price Hike Expectations The dramatic shift in demand structure coincides with a dual squeeze from supply-side capacity bottlenecks. According to TrendForce's January 2026 supply chain monitoring report, the advanced process capacity (e.g., N2, N3) at TSMC through 2027 has already been allocated by giants like Apple, NVIDIA, and Broadcom. Given that high-end GPUs and custom ASICs offer significantly higher "value per wafer" compared to traditional CPUs, foundries exhibit a clear bias in capacity allocation. Simultaneously, bottlenecks in advanced packaging technologies like CoWoS have further exacerbated the supply chain—IDC analysis indicates that its capacity utilization exceeded 100% as early as Q4 2025, causing CPU shipment cycles to extend from the normal 8-10 weeks to over 24 weeks. Intel's internal ecosystem is also facing extreme pressure. As its 18A process ramps up to peak production, the company must not only ensure supply for its own Core and Xeon series but also fulfill commitments to external foundry customers like Microsoft and Amazon. A J.P. Morgan research report notes that capacity utilization at Intel's core nodes has climbed to an overloaded state of 120%-130%, forcing the migration of some non-core components to secondary foundries like UMC. The latest industry commentary from Western Securities indicates that to address the supply-demand imbalance and ensure stable supply, Intel and AMD have planned to increase server CPU prices by 10%-15%. Furthermore, the server CPU capacity for 2026 from both manufacturers is reportedly "largely pre-sold." In summary, as AI progresses from "content generation" to "task execution," the core of computing demand is undergoing a structural migration—shifting from GPU-centric parallel computing to CPU-hubbed system scheduling and resource coordination. Under the dual pressures of supply-side capacity hitting physical limits and demand-side growth being driven exponentially by agent applications, CPUs not only face sustained upward price pressure but are also experiencing a systemic re-evaluation of their strategic value within the entire computing architecture.