Morgan Stanley: Power shortages are becoming the core bottleneck for AI.

Morgan Stanley identifies electricity as the key bottleneck in AI infrastructure construction. Morgan Stanley’s Chief Fixed Income Strategist Vishwanath Tirupattur wrote in a recent research note: “These constraints are not peripheral frictions, but the core of AI infrastructure construction.” One of the most serious constraints lies upstream of data centers—electricity. This means that simply looking at data center investment plans is no longer sufficient. Even if a project secures land, servers, and funding, if there is not enough local power generation capacity, transmission network, or critical electrical equipment, construction progress will still be delayed. 128 weeks for transformer delivery—the bottleneck is data center commissioning time The expansion of the electricity system is much slower than that of data centers. Transmission networks and key equipment supply chains all have much longer construction cycles. Data centers can rapidly ramp up capital expenditure plans, but the grid and equipment delivery cannot match that speed. Data shows: The average delivery cycle for power transformers has reached 128 weeks, and generator step-up transformers are at 144 weeks. Before the pandemic, the normal delivery time for such equipment was only 12–16 weeks. This is not a minor delay, but a shift from a few months to more than two years. For AI data centers, transformers are not optional components. Without these devices, generated power cannot be accessed and used; computing rooms cannot release their computational capacity as planned. Queuing for grid connection is harder than building the project—electricity is only usable when it reaches the grid Another bottleneck lies in grid connection. By early 2025, the backlog for grid interconnection in the U.S. has exceeded twice the total installed power capacity. This backlog refers to the wait for new energy projects from construction to actual grid connection. This is crucial for AI infrastructure. Data centers need not “electricity” in the abstract, but specific, connectable power at a certain location, time, and stability. If a power generation project isn’t grid-connected, the electricity cannot be supplied to the data center. This changes the site selection logic: developers are prioritizing locations where power is easier to obtain, and exploring ways to bring power generation and computation closer together. In other words, the logic is shifting from “where is suitable for a data center” to “where is fast, stable, and cheap power available.” The boundary between AI and energy financing is becoming blurred Once electricity becomes a roadblock, AI infrastructure and energy asset financing demands start to merge. “Off-grid” power solutions are entering the AI construction framework, including fuel cells, gas turbines, and energy storage. Some trades in investment grade and high yield bond markets already reflect this convergence. AI companies are no longer just waiting for utility companies to invest in power assets; they are starting to directly acquire, sign contracts, or finance related assets. This will change how capital markets view AI infrastructure. In the past, data centers, cloud computing, electricity, and utilities might have been priced in different capital pools. Now, computational capacity expansion requires these pools to interlock: whoever gets power fastest is closest to deployable computational resources. Besides electricity, there are electricians, water, and local approvals Electricity is not the only constraint. In terms of labor, the U.S. is expected to be short about 300,000 electricians over the next decade. Over one-fifth of the current electrician workforce is already 55 or older, nearing retirement. Expanding power infrastructure requires people to install, maintain, and upgrade—labor shortages will directly impact project delivery. Water resources are also becoming a limit. S&P analysis shows 43% of global data centers are located in high water stress areas. Data center cooling relies on water or alternative cooling solutions; if new construction continues to concentrate in these areas, sustainable expansion will face more scrutiny. Policy resistance is also rising. New York State has draft proposals to pause new data center projects. Texas Governor Abbott recently asked regulators to ensure new data center demand won’t raise costs for other users. More projects are being denied at the local level. More broadly, legislatures in 14 U.S. states are considering some form of data center moratorium. For data centers, power costs, residential electricity prices, water pressure, and local approvals are becoming interlinked. Computational supply may not keep up with demand—scarcity will shift pricing power These constraints all point to one result: computational supply may not grow in line with the current demand curve. If data center construction is slower than expected, the market won’t see simply a capital expenditure expansion, but a supply-demand mismatch. Electricity, grid connection, equipment, labor, water, and approval—any one can delay a project. In this environment, those with scalable, reliable computational capacity will gain stronger pricing power. The framework uses a phrase: these players are becoming “computing merchants.” The core is not who announces more capital expenditure, but who truly owns deliverable, stable computing power. On the demand side, sensitivity remains low, especially for enterprise applications. Higher computing prices may not slow enterprises’ adoption of AI much. On the contrary, usage may concentrate on higher-value applications, as these use cases still make economic sense even at higher computational prices. Risk warning and disclaimer The market has risks; investments are to be made cautiously. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment goals, financial situation, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article suit their particular circumstances. Investing accordingly is at their own risk.