According to a recent research report from Morgan Stanley, as the NVL72 rack-scale architecture becomes the mainstream approach for AI data center construction, gains in computing efficiency have not eased energy pressure — instead, they have created a massive power supply gap. The report warns that three constraints — electricity, manpower, and policy approvals — are becoming the core bottlenecks limiting the expansion of the U.S. AI industry, with "Time-to-Power" replacing raw chip performance as the new high-value position in the supply chain.

The iteration toward rack-level architecture is the direct trigger for the widening gap. The industry is moving away from the traditional 8U server model toward rack-scale solutions that integrate 72 GPUs in a single rack. Power consumption for next-generation racks such as Vera Rubin and Rubin Ultra has risen sharply, and rack power draw is not simply the sum of GPU power — supporting systems including memory, high-speed interconnect, and liquid cooling further push up electricity load. Although hardware delivers a leap in computing efficiency, with tokens produced per watt of electricity growing nearly sixfold from 2025 to 2028, the Jevons paradox comes into play: falling cost per unit of compute stimulates explosive growth in high-load tasks such as Agent-based AI, and the efficiency dividend is swallowed by expanding total compute demand, keeping overall power demand on a rising trend.

Data shows that cumulative total U.S. data center power demand will reach 97 gigawatts from 2026 to 2028. After deducting projects under construction and capacity available from the existing grid, the power gap — before accounting for remedial solutions — is as high as 57 gigawatts. Even after including various on-site power supply methods such as gas turbines, fuel cells, nuclear power co-location, and converted mining sites, the net power gap under a base-case scenario still stands at 33 gigawatts, or 34% of total power demand — roughly equivalent to the baseline electricity load of six New York Cities. Extending the timeline to 2029, with the arrival of the next-generation Feynman architecture, the U.S. net power gap will expand to 72 gigawatts. Grid interconnection approval cycles often stretch for years, and the pace of public grid expansion is far behind the speed of AI compute buildout, making "having power first" a real problem for cloud providers before they can begin production.

The supply gap is giving rise to an entirely new commercial track, with Power Shell Providers (PSP) and Behind-the-Meter (BTM) self-generation seeing a value reassessment. Many PSP companies converted from crypto mining firms hold ready land and grid interconnection resources, and can provide data center infrastructure capable of being energized quickly. Companies sign long-term leases of 15 to 25 years, and projects can achieve unlevered free cash flow yields of 15% to 19%. Energizing one year earlier can create incremental revenue of about US$4.5 per watt, and the commercial premium for speed of energization is being repriced by the market. Institutions judge that within the next six months, names such as CIFR, RIOT, and HUT are expected to land multiple heavyweight leasing deals, and a large wave of behind-the-meter on-site generation projects will cluster in West Texas.

Power constraints are also reshaping the geographic landscape of global compute. With domestic U.S. power supply under strain, hyperscale cloud providers are being forced to expand outward, and data center construction in the Nordics, ASEAN, India, and the Iberian Peninsula has room to rise beyond expectations. At the same time, the industry's cost structure is undergoing a qualitative shift, with the share of memory, network interconnect, and liquid cooling facilities rising rapidly while GPUs' share of total rack cost continues to decline, and profit distribution across the compute supply chain is gradually shifting toward power equipment and infrastructure segments.

It should be noted that while the power gap brings industrial opportunities, it also harbors multiple risks. Limited production capacity for core hot-end components of gas turbines, project approvals, equipment delivery cycles, and capital expenditure returns falling short of expectations will all disrupt the pace of industry realization. For market participants, in the second half of AI competition, beyond chips and algorithms, securing stable and timely power supply has become the key variable determining the upper limit of corporate development.