电子行业：Deepseek拉动推理需求爆发，服务器架构创新催生PTFE PCB需求

Investment Rating - The report recommends a "Buy" rating for stocks related to the AI infrastructure and PCB sectors, indicating a potential upside greater than 15% compared to relevant market indices [43]. Core Insights - Deepseek is expected to drive a surge in application-side AI demand, leading to a significant increase in computational power requirements. The model has achieved performance comparable to top global models at a very low training cost, which raises concerns about future computational needs [3][11]. - The demand for customized ASICs is strong, confirming the sustainability of AI's growth. Broadcom anticipates that the demand for ASICs and networks from three clients will reach between $60 billion and$90 billion by 2027, with a projected CAGR of 45% from 2023 to 2028 [3][25]. - The design of AI servers is undergoing iterative upgrades, which is expected to enhance the value of PCBs. The complexity of deploying new architectures like the GB200 has led to challenges in assembly yield and thermal management, necessitating design optimizations in future models [3][28]. Summary by Sections Section 1: Deepseek and AI Demand - Deepseek has significantly reduced pre-training costs, but the demand for inference power is expected to explode as applications become more widespread. The total training cost for Deepseek-V3 is estimated at $5.576 million, with operational costs potentially exceeding$944 million [11][12][13]. - The inference cost per million output tokens for Deepseek is $0.28, which is substantially lower than competitors, but the overall cost for serving billions of users remains high [15][16][17]. - Major cloud service providers (CSPs) are increasing their capital expenditures significantly, with Google projecting$75 billion for 2025, a 43% increase year-over-year [19][22]. Section 2: ASIC Demand and Market Growth - ASICs are expected to see explosive growth, with Broadcom predicting a demand of $60 billion to$90 billion by 2027 from three clients. Marvell also anticipates strong growth in AI inference accelerators [25][26]. - The CAGR for the ASIC market from 2023 to 2028 is projected to be 45%, indicating robust growth potential in this segment [27]. Section 3: AI Server Design and PCB Value - The deployment of the GB200 architecture has introduced complexities that affect thermal management and assembly yield. Overheating issues have been reported, leading to delivery delays [28][32][33]. - Future designs may adopt orthogonal architectures to improve data transmission speeds and reduce interconnect distances, with PTFE materials being considered for high-performance applications [36][37][38].