汇丰：AI 需求深度分析，仍大幅低估

Investment Rating - The report maintains a constructive outlook on AI demand trends, indicating a bullish sentiment towards the sector [4][29]. Core Insights - The easing of AI hardware supply chain constraints is expected to shift investor focus towards the sustainability of demand for AI compute and software applications, with a conclusion that demand will justify the current infrastructure build-out [2][4]. - Three primary sources of AI demand are identified: training for large language models (LLMs), consumer-facing AI products, and enterprise AI applications [3][15]. - The report emphasizes that while model training and consumer demand are robust, enterprise AI spending is lagging but is anticipated to ramp up significantly in the coming years [4][29]. Summary by Sections Executive Summary - AI model providers and hyperscalers have announced significant capital expenditure plans, with estimates exceeding $400 billion in 2025 for AI infrastructure [9][11]. - The report addresses concerns about whether the massive investments will yield sufficient returns, focusing on the balance between supply and demand [11][12]. Demand Pillars - Demand for AI infrastructure is driven by three distinct buyers: LLM training, consumer AI products, and enterprise applications [15][62]. - The first pillar, model training, shows no signs of slowing down, with significant GPU demand from providers like OpenAI and Google [17][39]. - The second pillar, consumer AI products, is thriving, with OpenAI's ChatGPT reaching over 600 million weekly active users, indicating strong consumer engagement [18][49]. - The third pillar, enterprise AI adoption, is currently slow but is expected to become a major demand driver by 2026-2027 as organizations mature their AI readiness [22][63]. Modeling Demand - A quantitative approach is taken to model GPU demand based on compute intensity for various AI workloads, projecting significant growth in the AI accelerator market from approximately $125 billion in 2024 to over $300 billion by 2027 [24][25]. - The report highlights that hyperscaler AI infrastructure spending is expected to remain high due to increasing compute requirements and the emergence of sovereign AI initiatives [25]. Bottom-Up View - Insights from 130 UBS analysts indicate that AI is being actively utilized across various industries, with a notable focus on sectors like insurance and IT services [26][28]. - The report tracks AI mentions in conference calls as a proxy for management focus on AI product development, revealing a growing trend in AI adoption across the economy [26][28]. UBS Bottom Line - The report concludes that model training and consumer-led growth in inference workloads will sustain GPU demand for years, driving further investment in cloud infrastructure and related technologies [29][30]. - Key stock picks based on this analysis include Nvidia, Broadcom, Oracle, Snowflake, and Meta, reflecting confidence in their ability to capitalize on AI demand [30].