In the migration from generative AI to agent-based AI, a new bottleneck has emerged, this time the CPU. Previously, discussions in the market about the shortage of computing power were mainly focused on GPU, HBM, optical modules, power, and other aspects, with relatively little attention paid to CPU. In fact, the shortage of CPUs has been circulating for some time. The core driving force behind the recent trends of Intel and AMD is the shortage of CPUs. Even Lenovo Group, which was not well received in the past, has been strong in the past two weeks. 1. Why will the proportion of CPU increase in the era of agentic AI? Traditional AI (mainly large model training/inference) heavily relies on GPUs because the core of Transformers is parallel matrix operations, and GPUs excel at high-throughput parallel computing. At this point, the CPU is mainly responsible for "auxiliary" tasks such as data routing, memory compression, GPU scheduling, etc., resulting in a very low CPU to GPU ratio in the data center (typically 1:4~1:8, even one CPU managing eight GPUs). Low CPU utilization, basically supporting role. Agentic AI is completely different. It is not a single "Q&A", but an autonomous multi-step loop (Planning → Tool Use → Act → Observe → Reflect → Iterate), involving: 1) Arrangement: Scheduling subtasks, multi-agent collaboration, branching logic, retry mechanism. 2) Tool calls: web search, API calls, code execution, database queries, vector retrieval (RAG), file processing, etc. 3) Other CPU intensive tasks: context management, KV Cache processing, reinforcement learning (RL) simulation evaluation, data preprocessing/post-processing. These tasks are highly serial, I/O intensive, and have multiple logical branches, which GPUs are not good at (and may even idle). Research shows that the tool processing phase can account for 50% to 90.6% of the total latency on the CPU (GPU waiting for CPU). The dynamic energy consumption of CPU in Agentic workflow can reach 44%, which is 3-4 times higher than traditional AI. Simply put, Agentic AI hands over "thinking" to the GPU, but "doing/coordinating" to the CPU. The CPU has changed from a "steward" to a "commander-in-chief" and must be significantly increased in order for the entire system to operate efficiently. This is the core driver for the expansion of CPU proportion (consensus among Intel, AMD, Arm, TrendForce, etc.). 2. CPU becomes the real evidence of the new tight link This year's Q1 Intel/AMD server CPU lead time has been extended to 6-12 weeks, with some models sold out and prices increased by more than 10%. Manufacturers themselves say 'demand far exceeded expectations'. It's not that the production capacity is insufficient, but rather that Agentic AI has transformed the CPU from being "optional" to a "must-have" commander-in-chief. The data center project is currently experiencing the most severe CPU bottleneck, in addition to electricity. Traditional x86 (Intel/AMD) has high power consumption and tight production capacity, leading to a direct disruption in the supply chain. 3. How big will the CPU gap be? The industry consensus is that the CPU: GPU ratio will significantly decrease, and CPU demand will increase significantly: from the traditional 1:4~1:8 (CPU: GPU) to 1:1~1:2 (in some scenarios, even 1.4:1, which means there are more CPUs than GPUs). According to Arm's previous estimate, the CPU cores required per GW of computing power have surged from 30 million to 120 million (a fourfold increase) CPU computing power share: In the Agentic workflow, the computing power borne by the CPU may shift from being "GPU dominated" to being more balanced than future racks/clusters, and even dedicated CPU racks may appear to support Agentic orchestration; The AMD/NVIDIA next-generation platform has begun to be designed in a 1:2~1:4 ratio This brings about a real turning point in CPU demand, which is a real hardware refactoring. 4. Especially, would ARM server CPUs benefit more? What Agentic AI needs most is "high core count+low power consumption+stable serial processing". ARM is naturally multi-core and scalable, leading in perf/watt: Arm AGI CPU (136 cores, TDP only 300W) consumes over 40% less power compared to x86 of the same specifications, doubling performance per rack. Air cooled racks can accommodate 8000+cores, while liquid cooled racks can accommodate up to 40000+cores, perfectly solving the "power wall" of data centers. Even more ruthless is the ecological shift: AWS Graviton, Google Axion, and Microsoft Cobalt have long developed their own ARM, and cloud giants collectively "de x86". Arm directly launched its self-developed AGI CPU (the first mass-produced chip) in March, with Meta, OpenAI, and Cerebras as initial partners, and OEMs including Lenovo Supermicro。 Counterpoint predicts that ARM's share in AI ASIC server CPUs will increase from 25% in 2025 to 90% in 2029. Arm himself said that this wave can increase the royalties of data center CPU TAM from 3 billion to over 100 billion, and in the next few years, server CPU revenue is likely to surpass mobile phones and become the largest growth pole. See the changes in actual CPU shipments and the actual price changes of CPUs during Intel and AMD's financial conference calls next week and early May. This can demonstrate how scarce it really is. 5. Which companies will benefit from CPU shortage? After sorting out which companies will benefit, we will pay attention to them in the future: The core of the US stock market: Intel (INTC) continues to be the dominant player in the server CPU market. The shortage wave will increase the profit margin of its previous models, and its combination of Gaudi and Xeon has strong demand on the agent inference end. AMD: Reason: In the Agentic AI server market, AMD's EPYC processors continue to increase their market share among cloud vendors due to their multi-core advantages and high cost-effectiveness, making them the preferred choice for GPU+CPU balanced configurations. Arm Holdings (ARM): More and more cloud vendors (Amazon, Microsoft, Google) are starting to develop their own CPUs based on ARM architecture. No matter who wins, as long as the demand for Agent drives up the number of CPU cores, Arm's licensing fees will skyrocket. Hong Kong stocks (key points of manufacturing and distribution) SMIC (0981): Although it is limited in its most advanced processes, the overflow of demand for a large number of non core logic control chips (auxiliary chips supporting CPU operation) and mid-range CPUs will significantly improve its capacity utilization. Lenovo Group (0992): The world's largest server and PC manufacturer. In the early stages of the shortage wave, large factories with strong supply chain management capabilities and inventory can seize more market share in the government and enterprise markets by raising prices and ensuring supply. A-shares (domestic substitution and supporting industry chain) Haiguang Information (688041): The leader in domestic x86 server CPUs. In the era of Agentic AI, due to its architecture having the best compatibility with the global ecosystem, Haiguang is the first priority substitute for domestic computing centers in filling the CPU shortage. Loongson Zhongke (688047): A representative of autonomous architecture CPUs. With the increasing demand for domestic independent controllability, it benefits from the application of agents in the party, government, and key infrastructure. Shennan Circuit (002916)/Shanghai Electric Power Co., Ltd. (002463): Reason: Benefit from supporting facilities. The increase in the number of CPU cores and the adjustment of GPU+CPU ratio require more complex PCBs (printed circuit boards) and packaging substrates. These companies are the main suppliers of high-end server PCBs worldwide. Lanqi Technology (688008): Leading memory interface chip manufacturer. As long as there are more CPUs, there will be more memory modules. In the era of agents, there is a high demand for memory bandwidth, and its MRIMM and memory interface chips are essential for the explosive performance of CPUs. The core of investment logic actually consists of two points: 1) Both quantity and price increase: CPU manufacturers (AMD, Intel, ARM, Haiguang) are the most direct. 2) The person selling shovels: Due to the high bandwidth required by agents, the demand for memory matching (Lanqi) and advanced packaging/substrate (Shennan) is even more stable than the CPU itself.