花旗：全球半导体-个人人工智能服务器时代开启

Investment Rating - The report does not explicitly state an investment rating for the semiconductor industry but highlights significant growth potential driven by advancements in AI technology and architecture changes in portable devices. Core Insights - The emergence of portable AI servers is expected to revolutionize the semiconductor market, with a projected increase in demand for on-device AI capabilities across various form factors such as smartphones, laptops, and robots [22][35][36]. - AI distillation technology, particularly innovations from DeepSeek, is anticipated to lower the cost and complexity of AI model training, thereby accelerating the adoption of on-device AI solutions [24][68]. - The report identifies three major architectural shifts in on-device AI products, emphasizing the integration of AI modules into existing architectures and the development of new memory technologies to enhance performance [28][33][69]. Summary by Sections AI Hardware Shift - The report discusses the transition from centralized AI computing to personal on-device AI servers, predicting that consumers will eventually carry portable AI servers as easily as smartphones [22][36]. - It highlights the historical context of computing evolution, suggesting a return to distributed computing models as personalized AI devices become mainstream [23][37]. AI Model Efficiency - DeepSeek's distilled AI models are noted for their lower training costs and comparable performance to traditional models, which could significantly enhance on-device AI capabilities [24][50]. - The report outlines the algorithmic innovations of DeepSeek, including Mixture-of-Experts architecture and Knowledge Distillation, which improve model efficiency and reduce hardware dependency [62][68]. On-Device AI Technology - The report identifies three architectural directions for on-device AI: adding AI kits to conventional architectures, utilizing near memory technologies, and placing low-power DRAM adjacent to processing units [28][33][69]. - It anticipates broader adoption of heterogeneous integration and mobile DRAM technologies, with LPDDR6 and LPW DRAM expected to become mainstream by 2028 [4][34]. Market Growth Projections - The report forecasts substantial growth in AI smartphone shipments at a CAGR of 78% from 2023 to 2028, with AI PC shipments projected to grow at 28% CAGR from 2024 to 2029 [5][107]. - Overall AI DRAM demand is expected to grow at a CAGR of 75%, reaching 331 billion pieces by 2028, driven by the increasing adoption of next-gen memory solutions [5][68]. Key Beneficiaries - Potential beneficiaries of the shift towards portable AI servers include major semiconductor and technology companies such as Hynix, Samsung, TSMC, Micron, Nvidia, and Mediatek, among others [6][68].