Morgan Stanley: Power shortages are becoming a core bottleneck for AI infrastructure, and computing power expansion has entered the "power-constrained era."

Mars Finance News, June 15—In its latest research, Morgan Stanley pointed out that power shortages have risen from ancillary/supporting issues to become the core limiting factor for AI infrastructure development. The delivery lead time for power transformers has been extended dramatically from the 12–16 weeks before the pandemic to 128–144 weeks; at the same time, the backlog of new energy projects awaiting grid connection in the United States has exceeded more than 2 times the current installed capacity. Meanwhile, a shortage of 300,000 electricians and the fact that 43% of data centers are located in high water-resource pressure zones are jointly suppressing the pace of expansion of computing power supply.

The expansion rate of the power system is far slower than the construction schedule of data centers, and the cycles of transmission network expansion and key equipment supply chains are significantly longer. Currently, the average delivery cycle for power transformers has reached 128 weeks, and generator step-up transformers are about 144 weeks, compared with only 12–16 weeks before the pandemic. This means that even if AI data centers complete financing, site selection, and equipment procurement, they may still be unable to begin operations on time due to delays in power connection.

In the grid connection stage, the backlog of new energy projects queued for grid connection in the United States has already exceeded twice the nation’s existing installed power capacity, creating a structural problem in which “generation completed does not equal available power.” Power must complete grid connection before it can be converted into supply that is usable for data centers, causing the site-selection logic to shift from “suitable for building data rooms” to “areas where power can be rapidly and stably connected.”

At the same time, the financing boundaries between AI infrastructure and the energy system are becoming blurred. Some projects have begun adopting off-grid or semi-off-grid solutions, including direct power-supply paths such as gas turbines, energy storage, and fuel cells. AI companies are also gradually shifting away from relying on utility-led expansion, and toward directly investing in power assets and locking in power supply capacity—driving integrated pricing of AI and energy assets in capital markets.

In addition to power, labor and resources are also constraints. Over the next decade, the United States is expected to face a gap of about 300,000 electricians, with more than 20% of workers aged 55 and above. At the same time, about 43% of data centers are located in high water-resource pressure areas, making cooling water and alternative solutions important limiting factors for new construction. Moreover, multiple states have begun discussing or pushing tighter restrictions and more stringent approvals for data center development, further increasing project uncertainty.

Overall, power, grid connection, equipment, manpower, water resources, and policy approval are forming multiple overlapping constraints, which may cause the expansion speed of computing power to lag behind demand growth. The report believes that this supply-demand mismatch will strengthen “the scarcity of computing power,” giving participants with stable and deliverable computing power stronger pricing power. The market is gradually shifting from “competition in computing power expansion” to “competition for control over available computing power.”