Sinolink: Agent industry has entered a new round of prosperity cycle with rigid demand for CPU

Sinolink releases research report stating that as large models evolve from Chatbot to Agent, the center of computational load is shifting. Agents not only require GPUs for model inference, but also rely on high-performance CPUs for complex logic orchestration, tool invocation, and memory management. With the exponential growth in the number of Agents, task complexity, and Token consumption, the CPU industry has entered a new cycle of prosperity. Intel and AMD server CPU inventory is tightening, lead times are extending, and prices have been steadily increasing since 2026. Key points from Sinolink are as follows: Agents have a rigid demand for CPUs, putting CPUs back into a central position As large models evolve from Chatbot to Agent, the center of computational load is shifting. Agents not only require GPUs for model inference, but also rely on high-performance CPUs for complex logic orchestration, tool invocation, and memory management. Sinolink believes that the rigid demand for CPUs by Agents mainly comes from three aspects: 1) the OS scheduling pressure brought by Multi-Agent architecture, as well as the continuous consumption of CPU power for sandbox environment creation, scheduling, and destruction; 2) the higher requirements for CPU memory and bandwidth posed by KV Cache unloading in long context scenarios; 3) the significant CPU power consumption brought by high-concurrency tool invocation. According to Intel's paper, in most Agent workloads, CPU time can account for 40%-90% of the end-to-end latency. Sinolink believes that with the exponential growth in the number of Agents, task complexity, and Token consumption, the CPU industry has entered a new cycle of prosperity. Intel and AMD server CPU inventory is tightening, lead times are extending, and prices have been steadily increasing since 2026. Expansion of CPU TAM, raising CPU/GPU deployment ratio 1) In terms of TAM, AMD and Arm have significantly revised the market space for server CPUs, with the global server CPU TAM expected to exceed 100 billion dollars by 2030. According to AMD, CPU demand can be divided into general computing CPUs, AI head node CPUs, and Agentic AI CPUs, with AI-related demands being the largest source of incremental growth. 2) In terms of deployment ratios, the CPU/GPU deployment ratio in AI data centers is transitioning from the traditional HGX era of 1:4, 1:8 to 1:2, 1:1, or even higher. For example, NVIDIA's NVL72 with 72 GPUs paired with 36 Grace CPUs has achieved a 1:2 ratio, and Vera Rubin further raises the overall CPU ratio by attaching independent Vera CPU cabinets. All CPU architectures benefit, with ARM showing more significant changes in the medium term 1) The low power consumption and high core density characteristics of ARM architecture are more suitable for Agent workloads. Compared to x86, ARM has better energy efficiency and scalability in high-concurrency, low-power scenarios, especially suitable for light computing and high-concurrency tasks such as massive API calls and KV Cache scheduling. 2) The open licensing ecosystem of ARM architecture is highly suitable for cloud vendors to build their own AI infrastructure. Current solutions such as AWS Graviton, NVIDIA Grace, and Microsoft Cobalt are accelerating deployment. According to ARM's FY26Q4 results meeting, ARM architecture is expected to have the largest market share by CPU type in 2030. Agentic AI drives CPU restructuring, as global manufacturers embark on a new round of architecture upgrades 1) Overseas, Intel, AMD, Arm, NVIDIA, and others are conducting a new round of product iterations around high core density, heterogeneous collaboration, and energy efficiency optimization. CPU competition is shifting from pure performance competition to system-level computational efficiency competition. 2) Domestically, manufacturers such as Hygon, Feiteng, Loongson, Huawei HiSilicon, and Yizhi Electronics are continuously breaking through in the directions of x86, ARM, and proprietary instruction sets, rapidly increasing core counts, thread counts, memory bandwidth, and ecosystem capabilities. With the surge in CPU demand brought by Agentic AI and the deepening trend of self-control, domestic CPUs are expected to undergo a scale replacement and a reevaluation of their industrial status. Risk Warning Risks of intensified industry competition; risks of technological research and development progress falling behind schedule; risks of cyclical fluctuations in downstream capital expenditures in specific industries.