America Is Running Out of Power for AI Data Centers

The race to build artificial intelligence infrastructure in the United States is running into a hard physical limit: electricity. A series of research notes published by Morgan Stanley between September and October 2026 has revived fears that American data center developers will not be able to energize a large share of the computing capacity they are planning, with the shortfall expected to widen sharply in 2027 and again in 2028.

The central concern revolves around what the industry calls a powered shell, a data center building that already has a grid connection or on-site generation and is ready to receive servers. Chips can be manufactured and delivered within months, but securing a large grid connection, building a substation or obtaining heavy electrical equipment can take several years. As a result, the scarce asset in the AI boom is no longer the processor itself but a building with enough power to run it.

In its report released on September 21, 2026, Morgan Stanley raised its estimate of cumulative US data center power demand for the 2026 to 2028 period to 97 GW, roughly 43% above its previous forecast of 68 GW. After deducting about 21 GW of facilities already under construction and around 19 GW of grid capacity that is available or can still be contracted, the bank calculated a gross gap of 57 GW, compared with 38 GW in its earlier model.

Even after counting faster ways to bring power online, the hole remains large. The bank expects new gas turbines and engines installed directly on data center sites to add about 19 GW by 2028, fuel cells from Bloom Energy around 6 GW, and data centers located next to operating nuclear plants about 3 GW. Once these options are weighted by their probability of success, they provide roughly 24 GW, leaving a net shortfall of about 33 GW, or 34% of the power needed to run the chips expected to ship over the period.

The gap grows quickly over time. According to the model, the cumulative midpoint shortfall stands at about 5 GW in 2026, rises to around 12 GW in 2027 and reaches 33 GW in 2028. Extending the forecast to 2029 pushes total demand to 165 GW and the residual gap to roughly 72 GW. Morgan Stanley stressed that these figures are model estimates based on projected chip shipments, not capacity that utilities have formally approved.

The main driver of the upward revision is the shift toward rack-scale AI systems, where dozens of processors, memory, networking and liquid cooling are packed into a single cabinet. The bank now assumes that a rack built around the Nvidia Vera Rubin platform will draw about 234 kW, up from 149 kW in its previous model, while the following Rubin Ultra generation is expected to reach around 600 kW per rack. Each new generation is far more efficient per unit of computing, but the absolute amount of electricity consumed keeps climbing.

The analysts also corrected a methodology issue. Earlier versions of the model counted some powered shell sites twice, once as a separate fast-to-power solution and again inside the grid and construction figures. Removing this double counting added about 15 GW to the residual gap and showed that the pool of ready-to-use buildings is smaller than many investors had assumed.

The grid situation on the ground confirms the pressure. PJM Interconnection, the largest grid operator in the country, which covers 13 states and Washington DC, failed for the first time in its December 2025 capacity auction to secure enough resources to meet its reliability target, ending about 6.6 GW short for the delivery year starting in June 2027. Northern Virginia, and Ashburn in particular, hosts the densest cluster of data centers in the world and sits inside the PJM footprint, which makes it one of the most exposed areas when new projects wait for a connection.

This scarcity is reshaping the market. Morgan Stanley calculated that bringing a site online one year earlier creates about 4.50 dollars of value per watt, nearly six times the annual cost of the electricity itself. Lease rates for powered shells have climbed from about 1.80 dollars per watt of IT capacity in the summer of 2025 to between 2.30 and 2.40 dollars. Former bitcoin mining companies that already hold large grid connections have become key players. Hut 8 signed leases covering 704 MW at its Beacon Point campus, and TeraWulf agreed to a 20-year lease with Anthropic for about 401 MW in Kentucky. The bank expects Cipher Digital, Riot Platforms, Galaxy Digital, Hut 8 and TeraWulf to sign further large contracts in the coming months, while many new projects are moving to West Texas, where on-site gas generation is easier to develop.

On October 5, 2026, Morgan Stanley published a follow-up note on the consequences for the semiconductor supply chain. The bank said Nvidia and Broadcom remain relatively protected because they know where their chips will be installed and can redirect deliveries to other regions. Suppliers of memory, optical modules, power management and analog chips are more exposed. If a data center is delayed by several months for lack of electricity, its operator can postpone or cancel component orders, which could trigger inventory corrections even while memory remains in short supply.

Other constraints add to the power problem. Large transformers and gas turbines still face multi-year delivery times, skilled electricians are in short supply, and local opposition to new data centers has grown, with protests held in dozens of states during the summer of 2026 and several local councils considering moratoriums. Each of these factors slows the conversion of planned capacity into operating capacity.

As of mid-October 2026, the shortage is still developing rather than easing. The industry is responding with on-site gas plants, fuel cells, nuclear partnerships and the conversion of Bitcoin mining farms, but current estimates suggest these measures will cover only part of the need. Unless grid connections and equipment deliveries accelerate, a significant share of the AI chips scheduled for deployment in 2027 and 2028 is likely to wait in warehouses or be installed outside the United States.