摩根士丹利于7月27日发布题为《Playing the AI Infrastructure Dip》的120页深度报告，分析了6月下旬以来全球AI板块的显著回调[1]。报告认为此次回调主要由技术面因素驱动，包括仓位拥挤度消化、去杠杆和动量反转，而非基本面恶化[1]。
报告通过实际案例论证AI需求远超供给的结构性判断。企业级AI应用平均单次调用可节省约55美元劳动力成本，通过Agent协作完成一项任务平均成本仅为2-5美元，投资回报率超过10倍[1]。根据芯片迭代规划，基于Blackwell的数据中心Token销售净利率约58%，Rubin和Feynman代际GPU部署后利润率分别升至约80%和90%[1]。NVIDIA 2025-2028年AI芯片销售复合年增长率约140%[1]，而Google高管表示公司可能需要每6个月将算力翻倍，即5年内实现1000倍增长[1]。
电力供应制约成为市场定价最大偏差所在[1]。2026-2028年美国数据中心电力需求约68GW，扣除在建设施（15GW）和已签约电网容量（15GW），潜在缺口达38GW[1]。电网并网排队周期在部分地区已延长至5-7年[1]。报告指出，当前Powered Shell Provider的EV/Watt仅处于2-4美元区间，基于成熟数据中心运营商20-25倍EV/Watt参照，折价目标估值为15倍EV/Watt[1]，暗示相关公司具有估值修复空间。
Morgan Stanley released a comprehensive 120-page research report titled "Playing the AI Infrastructure Dip" on July 27, analyzing the significant retreat in global AI equities since mid-June [1]. The report attributes this correction primarily to technical factors rather than fundamental deterioration, identifying crowded positioning, deleveraging, and momentum reversal as key drivers [1].
The bank directly counters three major market concerns—the narrative reversal around tokenmaxxing, improvements in Chinese model efficiency, and supply constraints—to reinforce its structural thesis that AI demand vastly exceeds available supply [1]. According to the report, enterprise-level AI applications can save approximately $55 in labor costs per single invocation, while Agent-based task completion costs only $2-5 on average, yielding returns on investment exceeding 10 times [1]. For next-generation hardware, Blackwell-based data center deployments generate net margins around 58%, while Rubin and Feynman era GPUs are projected to achieve margins of approximately 80% and 90% respectively [1].
Supply-side constraints emerge as the binding factor. Google executives have indicated the company may need to double computing capacity every six months, translating to a 1,000-fold increase over five years [1]. NVIDIA's AI chip sales are forecast to grow at a compound annual rate of approximately 140% from 2025 through 2028 [1]. On the power infrastructure front, U.S. data center electricity demand is projected to reach approximately 68GW through 2028, but when accounting for facilities under construction (15GW) and committed grid capacity (15GW), a potential 38GW shortfall emerges [1]. In certain regions, grid interconnection queues have extended to five to seven years [1].
The report identifies time to power access as the area where market pricing divergence is greatest [1]. Powered Shell Providers currently trade at EV/Watt valuations between $2-4, compared to the 20-25x EV/Watt multiples typical of mature data center operators, suggesting a potential revaluation target of 15x EV/Watt [1].
