Power supply has become one of the core bottlenecks in the construction of AI data centers in the United States, according to a research report released by CITIC SEC (ASX: 06030).
Citing Aterio, the report noted that the capacity of data centers currently under construction in the US stands at 71.7GW, while new grid capacity expected between 2027 and 2028 is estimated at only 25.2GW. The brokerage judged that the current boom in the US power industry is underpinned by the contradiction between the relatively high return on investment of data centers and the scarcity of power supply resources: investment in the power system is relatively limited, yet connecting to power earlier can accelerate capital recovery; according to its estimates, if the first-year capacity utilization rate rises from 50% to 75%, the project ROIC can climb from 16.5% to 19.8%, providing an economic basis for customers to procure high-efficiency power supply solutions. Given the long cycle for grid expansion and the difficulty of fully off-grid operation, the brokerage expects grid upgrades and behind-the-meter generation to advance in parallel, with hybrid power supply gradually becoming mainstream, and the grid equipment, power transmission and transformation EPC, and behind-the-meter generation industrial chains are expected to benefit together.
Core viewpoints of CITIC SEC are as follows
Power supply has become the core bottleneck in US AI data center construction, and capital expenditure on computing infrastructure is spreading to the power supply industrial chain. According to Aterio's official website, as of early September 2026, the US already had 52.4GW of data center capacity, with 71.7GW under construction; while according to SemiAnalysis, new US grid capacity in 2027-2028 is expected to be only 25.2GW, just 35% of the capacity currently under construction, corresponding to a gap of about 46.5GW. As the scale of individual data centers rises toward the GW level, project requirements for power supply capacity, reliability, and delivery speed have increased simultaneously, but the cycles for grid planning, equipment production, and engineering construction are all significantly longer than the data center construction cycle. The brokerage believes that the constraint on data center construction has shifted from server and GPU supply to power access, and North America will simultaneously advance transmission and distribution grid upgrades, cross-regional power transmission, and campus-side behind-the-meter generation, with power equipment and engineering services expected to become an important incremental direction for AI infrastructure investment.
Demand side: time is money, efficiency is life
Based on the requirements of the Uptime Institute's official website for Tier IV data centers, and drawing on Tom's Hardware, alpha-matica's official website, and announcements and official websites of companies including Nvidia, ETN, WDC, and GEV, the brokerage estimates that the total capital expenditure for a 1GW AI data center is about US$49.9 billion, of which the main power supply, backup power supply, and grid system together account for about 17%; annual operating costs are about US$8.4 billion, with electricity accounting for about 8%. Investment in the power system is relatively limited, but the timing of power supply determines when computing capacity begins operation. Under the leasing price and utilization assumptions set in the report, the project's annual revenue can reach US$18.9 billion after full production, with a full-lifecycle ROIC of about 16.5%. If the power supply solution enables the project to come online three months earlier, the corresponding additional revenue would be about US$4.7 billion, close to the scale of capital expenditure on the main power supply system; if the first-year capacity utilization rate rises from 50% to 75%, the project ROIC can rise to 19.8%, an increase of 3.3 percentage points. The brokerage believes that data center customers have a high willingness to pay for shortening grid connection and power-on cycles, and transformers, switchgear, gas turbines, fuel cells, and engineering services that can accelerate delivery are expected to benefit.
Supply side: grid expansion speed struggles to match data center construction demand, and the industry is entering a hybrid power supply stage
According to SemiAnalysis, new US data center capacity demand in 2027 is about 39GW, while available grid capacity is only 24.1GW, with about 14.8GW needing to be supplemented by behind-the-meter generation; in 2028, demand rises to 54GW, while new grid supply is only about 11.1GW, expanding the behind-the-meter power supply gap further to about 42.6GW. On the grid side, Wood Mackenzie expects shortages of power transformers and substation equipment of about 15% and 8%, respectively, in 2026; ETN disclosed that transformer lead times once approached two years, and transmission and transformation construction is also constrained by a shortage of skilled workers. On the campus side, fuel cells, aeroderivative gas turbines, and reciprocating gas engines can help projects connect to power earlier, and GEV's aeroderivative gas turbine shipments in the first half of 2026 have already approached full-year 2025 levels. However, fully off-grid operation still requires solving issues related to power supply redundancy, load fluctuation, and long-term operations and maintenance. The brokerage believes that hybrid power supply combining front-of-meter and behind-the-meter is more feasible, and the grid equipment, power transmission and transformation EPC, and behind-the-meter generation industrial chains are expected to benefit together.
Grid expansion: grid expansion speed struggles to match data center construction demand, and equipment and construction bottlenecks drive parallel development of behind-the-meter generation
On the equipment side, ETN's transformer delivery cycle once approached two years and was extended again in 2026Q2. On the construction side, transmission and transformation projects are constrained by a shortage of electricians and skilled workers, with the training cycle for experienced technicians at about four years. Supply constraints have gradually been reflected in company orders: in 2026Q2, ETN's North American electrical business had backlog orders of US$15.2 billion (up 33% year on year), GEV's electrification business received more than US$5 billion in data center orders in 2026Q2, and PWR's power infrastructure backlog reached US$43.79 billion in 2026Q2. The brokerage believes that order accumulation provides strong support for subsequent revenue growth in the grid equipment and grid EPC industries.
Behind-the-meter generation: fuel cells, aeroderivative gas turbines, and reciprocating gas engines are accelerating deployment
Fuel cells offer advantages including modularity, high efficiency, and relatively fast deployment; aeroderivative gas turbines have a single-unit capacity of tens of megawatts and can complete installation and commissioning in as little as about one month; natural gas reciprocating engines have rapid response and multi-unit parallel capabilities, which can match AI load fluctuations and phased campus expansion. On the demand side, the agreement between Bloom Energy and Oracle supports a procurement scale of up to 2.8GW; GEV received 70 new orders for aeroderivative gas turbines in 2026H1 (up 94% year on year); the joint venture between FTAI and Jereh obtained an initial order worth US$1.465 billion; and CAT also secured an initial 2GW unit order for the Monarch campus. On the supply side, Bloom Energy plans to increase the capacity of its existing plant to 2GW by the end of 2026, and CAT will raise its large reciprocating engine capacity target to nearly three times the 2024 level, providing support for subsequent deliveries. Considering that fully off-grid operation still faces difficulties such as redundant configuration, load management, and long-term operations and maintenance, the brokerage expects that hybrid power supply combining the grid and campus power sources will become mainstream.
Risk factors
Risks include AI data center construction progress falling short of expectations; gas turbine capacity expansion falling short of expectations; fuel cell capacity expansion falling short of expectations; grid construction progress falling short of expectations; large fluctuations in fuel prices such as natural gas; intensifying industry competition; unexpected changes in nuclear power policy; and a global macroeconomic slowdown leading to an unexpected decline in corporate electricity spending.
Investment recommendations
Current AI data centers offer relatively high investment returns, but project ROIC is highly sensitive to the timing of production and first-year capacity utilization. The brokerage estimates that the current full-lifecycle ROIC of AI data centers is about 15% to 20%; if the first-year capacity utilization rate rises from 50% to 75%, project ROIC can rise from 16.5% to 19.8%; for a 1GW AI data center, if the project is put into operation three months earlier, the additional cash recovery can basically cover the capital expenditure of the main power supply system. Because US grid expansion and grid connection cycles struggle to match data center construction demand, the brokerage believes that behind-the-meter generation is not only a temporary solution to the power gap, but also an important means of improving data center project ROIC. At the same time, because fully off-grid operation still requires solving issues related to power supply redundancy, load fluctuation, and long-term operations and maintenance, the brokerage believes that hybrid power supply combining front-of-meter and behind-the-meter is more feasible, and the grid equipment, power transmission and transformation EPC, and behind-the-meter generation industrial chains are expected to benefit together. In the behind-the-meter generation direction, it recommends focusing on companies related to fuel cells, aeroderivative gas turbines, and reciprocating gas engines; in the grid expansion direction, it recommends focusing on transformer and grid EPC-related companies.