Morgan Stanley's latest research report indicates that as rack-scale AI architectures rapidly proliferate, the power deficit facing U.S. data centers is far more severe than market expectations. The bank projects that from 2026 to 2028, cumulative power demand from U.S. data centers will reach 97 gigawatts (GW). After deducting grid capacity under construction and available, the power deficit before implementing any solutions is as high as 57 GW, an increase of approximately 50% from the previously estimated 38 GW. Even when "fast-power" solutions such as natural gas turbines, fuel cells, and nuclear power plant co-location are included in calculations, the median net deficit still reaches 33 GW, accounting for 34% of U.S. data center power demand during the same period.

The report points out that the migration of AI chip deployment from 8-GPU server architectures to 72-GPU rack-scale architectures (NVL72) is the direct driver of the widening deficit, leading Morgan Stanley to significantly raise its power consumption assumptions. Analysts believe that obtaining power one year earlier creates value equivalent to approximately $4.5/watt, which is 5.9 times the annual electricity cost. Therefore, "the urgency for AI players to secure power is far greater than market perception." Former Bitcoin mining companies that have transitioned into "power supply shell" providers (PSPs) are benefiting from this, with unlevered free cash flow yields reaching 15% to 19%.