Damo points out the “hard restraint” of the AI computing power arms race: the US electricity gap is highlighted, and “fast power supply” is on the cusp of the industry

The Zhitong Finance App learned that Morgan Stanley recently released a research report. As the NVL72 cabinet architecture becomes the mainstream construction plan for AI data centers, the increase in computing power efficiency has not relieved energy pressure; on the contrary, it has spawned a huge power supply gap. Research shows that the triple restrictions on electricity, manpower, and policy approval are becoming the core bottleneck limiting the expansion of the US AI industry. “Time-to-power (time-to-power)” has replaced simple chip performance and has become a new value highland in the industrial chain.

Rack-level architecture iterations are a direct driver of a widening gap. The industry bid farewell to the traditional 8U server model and switched to a complete cabinet solution with 72 GPUs integrated in a single cabinet. The power consumption of Vera Rubin and Rubin Ultra's next-generation cabinets has increased dramatically. Cabinet power consumption is not just a simple superposition of GPU power consumption; supporting systems such as memory, high-speed interconnection, and liquid cooling are further driving up the electricity consumption load.

Although hardware achieved a jump in computing power efficiency, and the number of tokens produced per watt of electricity increased nearly 6 times from 2025 to 2028, the decline in the cost of computing power per unit stimulated explosive growth in high-load tasks such as agent agents. The efficiency dividend was swallowed up by the expansion of total demand for computing power, and overall electricity demand continued to rise.

According to the data, the total cumulative electricity demand for US data centers reached 97 gigawatts in 2026-2028. After deducting the available capacity of projects under construction and existing grids, the electricity gap not included in the remediation plan is as high as 57 GW. Even if various local power supply methods such as gas turbines, fuel cells, nuclear power facilities, and mine rehabilitation are included, the net electricity gap is still 33 gigawatts under a neutral scenario, accounting for 34% of total electricity demand, and the scale is about equal to the basic electricity consumption load of the six New York cities.

The period is extended until 2029. Combined with the implementation of Feynman's next-generation architecture, the net electricity gap in the US will expand to 72 gigawatts. The approval cycle for connecting power grids can easily take several years. The pace of public grid expansion is far from keeping up with the speed of AI computing power construction. “Only when there is electricity can be put into production” has become a real problem facing cloud vendors.

The supply gap has spawned a new commercial circuit, and power supply case service providers (PSP) and self-built electricity generation (BTM) have ushered in a revaluation. A large number of PSP companies transformed from crypto mining companies have ready-made land and grid access resources, and can provide data center infrastructure that can be quickly powered up. The company signs a 15-25 year long-term lease, and the project's unleveraged free cash flow yield can reach 15-19%.

Powering up one year in advance can generate incremental revenue of about $4.5/watt, and the commercial premium on electricity delivery time is being repriced by the market. According to the agency's judgment, CIFR, RIOT, and HUT are expected to land multiple heavyweight leasing transactions within the next six months, and West Texas will concentrate on the emergence of a large number of off-site power generation projects.

Electricity constraints are also reshaping the geographical pattern of global computing power. Power supply in the US is under pressure, forcing large-scale cloud vendors to move outward. Data center construction in regions such as Northern Europe, ASEAN, India, and the Iberian Peninsula has room to rise beyond expectations. At the same time, the cost structure of the industry has changed qualitatively. The share of memory, network interconnection, and liquid cooling facilities is rapidly rising, the share of GPUs in the overall cost of cabinets continues to decline, and the profit distribution of the computing power industry chain is gradually shifting to power equipment and infrastructure.

It is important to note that while the electricity gap brings industrial opportunities, it also harbors multiple risks. Limited production capacity for gas turbine core hot-end components, project approval, equipment delivery cycles, and returns on capital expenses fall short of expectations, all disrupt the pace of industry implementation. For market participants, in the second half of the AI competition, in addition to competing for chips and algorithms, competing for stable and timely electricity supply has become a key variable in determining the upper limit of enterprise development.