摩根士丹利：Crypto-to-DC Conversion Analysis

Investment Rating - The report expresses a bullish outlook on the non-linear rate of AI capability improvement, particularly highlighting the exponential growth in AI performance metrics over the past six years [3]. Core Insights - The total cumulative spend on AI infrastructure is projected to exceed $3 trillion through 2028, with approximately $2.6 trillion allocated for data centers, including chips and servers [5][11]. - Generative AI (GenAI) is expected to create a revenue opportunity of around $1 trillion by 2028, with software spending projected to rise from $16 billion in 2024 to $401 billion by 2028, representing about 22% of total software spending [12][14]. - Consumer spending on GenAI is anticipated to grow from $29 billion in 2024 to $683 billion by 2028, driven primarily by eCommerce, search, and autonomous technologies [14]. Summary by Sections AI Infrastructure and Power Demand - The report indicates that over 110 gigawatts (GW) of power will be needed through 2028, with associated costs for power plants estimated between $210 billion and $330 billion [11]. - A survey by Schneider Electric highlights that grid constraints are the primary barrier to new data center projects, with nearly half of respondents reporting average new data centers of 100+ MW [20]. Data Center Development - Cushman & Wakefield is tracking 47 GW of US data centers in development, with a projected demand of 62 GW through 2028, indicating a significant focus on training-focused data centers [24]. - The report discusses various "de-bottlenecking" solutions for data centers, including building power plants on-site and redirecting power from Bitcoin sites, although these options face execution risks [25][26]. Economic Metrics and Valuation - The report outlines the potential for high returns in building and leasing "powered shells" to hyperscalers, with indicative enterprise value/EBITDA multiples ranging from 10.0x to 15.0x [30]. - Bitcoin stocks are noted to trade at low enterprise value/watt levels, suggesting potential for conversion transactions to high-performance computing (HPC) data centers [27]. AI Adoption and Innovation - The report emphasizes that the level of AI adoption is under-appreciated, with significant investments expected in training AI models due to the high value of improved cognitive capabilities [31]. - The cost per unit of computational power is projected to drop by approximately 90% over a six-year period, indicating rapid innovation and depreciation risk in the GPU replacement cycle [32].