Gf Securities released a research report stating that in AI's "Memory" moment, AI memory is becoming the underlying capability supporting contextual continuity, personalization, and the reuse of historical information. It continuously expands the boundaries of model capabilities and is expected to accelerate the adoption of applications like AI Agents. The value of AI memory is shifting from a "cost item" to an "asset item," and the value and importance of related upstream infrastructure will continue to increase. The report recommends focusing on core beneficiaries within the industrial chain.
The increased demand for CPUs from AI Agents primarily stems from three aspects: (1) Amplified application invocations. In the Agent era, a single user can simultaneously invoke multiple types of Agents to drive various applications and toolchains on the server side. The overall invocation frequency and coverage are significantly higher than human usage, leading to more system requests, data movement, and control flow overhead, thereby increasing CPU load. (2) Orchestration and scheduling becoming a bottleneck. In the Agent era, LLMs are coupled with decision orchestrators and various external tools (search, Python, databases, APIs, etc.), evolving the overall computational process into a closed-loop structure of "perception-planning-tool invocation-re-reasoning." Since key steps like tool invocation, task scheduling, and information retrieval all rely on CPUs, as Agent penetration and tool invocation density increase, the utilization of the CPU as the central scheduling hub expands linearly. (3) Sandbox isolation raising fixed overhead. To mitigate the impact of misoperations on tenants/networks/servers, Agent runtimes commonly use sandboxing/virtualization isolation, which introduces additional process and kernel overhead, increases IO bandwidth and local SSD cache requirements, further pushing up CPU/memory/storage configurations and suppressing cluster utilization.
The CPU-to-GPU ratio is continuously increasing. According to a SemiAnalysis report, the current CPU ratio per GPU megawatt (MW) is below 10%. It is projected that by Q2 2026, the CPU ratio per GPU MW will rise to 15%. Consequently, the report assumes: (1) Under a 300 MW GPU building power, if CPU ratios are 16% and 25% respectively, the number of additional general-purpose server CPUs per GPU would be 0.29 and 0.45; the total number of CPUs per GPU would be 0.79 and 0.95. (2) Assuming global shipments of B200 and ASICs are 8 million and 7 million units respectively, the total number of X86 CPUs in AI clusters would be 6.1 million and 8.55 million units, with incremental AI CPU demand at 25% and 36% respectively.
The potential for memory modules and interface chips is vast. (1) Volume: General-purpose servers typically use half-population (1 DPC) memory configurations to balance cost and frequency; AI servers, to meet higher capacity and bandwidth demands, tend towards full population (2 DPC). Thus, DIMM demand can be expressed as: DIMM stick growth = CPU volume growth × 2. (2) Price: The memory form factor for X86-CPU configurations is expected to gradually evolve from traditional RDIMMs towards MRDIMM solutions, requiring the introduction of more complex interface chips and driving up the ASP of supporting chips. The combination of increasing population rates and MRDIMM penetration jointly opens up significant potential for memory modules and interface chips.
Risks include the AI industry's development and demand falling short of expectations; AI server shipments falling short of expectations; and slower-than-expected technological and product progress by domestic manufacturers.
Comments