Science & Tech
US AI data centres face a 32-gigawatt power shortfall by 2028
Electricity access now limits which AI companies can scale.
Electricity access now limits which AI companies can scale.
AI companies may have enough chips but nowhere to run them. Morgan Stanley estimates that US could lack 32 of power by 2028. That equals about 34% of expected demand, even after counting short-term fixes such as and . US power use is expected to reach record levels in 2026 and 2027, partly because of AI . Power lines and take longer to build than data-centre campuses.
Companies are turning to gas plants, batteries, nuclear-site upgrades and old industrial sites. These choices can increase emissions, water use and local power prices. Large technology firms may secure limited power first and help pay for projects. Smaller AI companies could then face higher costs or long delays. Utilities and governments must decide who pays for new power plants, and grid upgrades. Communities must weigh jobs and investment against risks to reliable power and the environment.
EIA analysis warned that demand growing faster than supply could produce higher or periods of . Communities near new data centres also face questions about emissions, water use and who pays for generation, transmission and local grid upgrades. The infrastructure race now depends on whether utilities, regulators, developers and customers can build enough without shifting unacceptable costs to other electricity users.
The analysis concluded that expanding chip supply alone may not sustain AI growth if companies cannot secure electricity, suitable land and operating data centres. Generation, transmission, and finance determine how quickly chips can become . The latest estimate therefore frames reliable power access as a potential in the AI race.
Morgan Stanley estimated that US data-centre developers could face a net 32- power shortfall through 2028, equal to about 34% of projected requirements. The estimate reportedly includes additional generation solutions, so it is not the same as total AI electricity demand. It captures the remaining gap after expected grid supply and temporary options such as onsite generation.
Morgan Stanley said limited electricity, data-centre capacity and could give financially stronger technology companies an advantage over smaller , potentially making infrastructure harder for smaller rivals to secure.
Transmission limits and delayed grid connections pushed more data-centre developers toward onsite . These installations can provide faster than new transmission, but they may increase emissions and local infrastructure demands. They also represent temporary or supplemental solutions rather than proof that the broader power gap has been solved.
The projected US would reach another record in 2026 and rise again in 2027. Data-centre development, including AI computing, was identified as a major source of commercial and industrial growth. The forecast describes total expected demand, not a net shortfall after new generation and temporary substitutes.
Large technology companies and data-centre operators pursued campuses, utility connections and onsite or dedicated generation options as grid interconnection delays and transmission constraints made dependable power harder to secure.
Texas grid planning confronted a concentration of proposed large loads in a region already experiencing rapid electricity growth. Revised connection procedures and limits highlighted that projects could not all receive immediate . Developers increasingly had to consider staged, provisional or while permanent upgrades were completed.
US power consumption reached a record level in 2025, with data-centre expansion among the major drivers of increasing demand. Large computing facilities joined and manufacturing as forces reshaping . Rising total demand did not by itself prove a shortage, but it increased pressure on generation and .
and AI developers began planning much larger to train and serve new models. These projects required substantial electricity, land, and financing before chips could be deployed. The scale of the announcements made power supply a strategic constraint alongside .
The rapid expansion of increased demand for and . The change shifted infrastructure planning from conventional cloud growth toward . Chips became useful only when companies could place them in operating facilities with sufficient electricity.
Reliable electricity access, including , utility commitments, and generation options, increasingly affected whether data-centre projects could be developed, financed and operated. Sites without a credible path to power could remain commercially unusable.
Battery storage, and emerged as ways to reduce peaks and improve reliability. These tools can help data centres respond to constrained system conditions without eliminating their total electricity needs. They separate from the larger requirement for sustained energy over the year.
AI developers and infrastructure operators explored locations near nuclear plants, and converted crypto-mining sites. Existing generation, and could shorten the path to operation. The search made physical access to power infrastructure a competitive asset independent of land availability.
Developers examined and other technologies as temporary or supplemental sources for AI campuses. These systems can be installed near the load, but their economics and environmental performance depend on fuel, scale and operating conditions. Onsite power therefore became a plug-and-play supplement, not equivalent to new .
Data-centre companies increasingly considered and other local sources to obtain more quickly. The approach can reduce dependence on delayed interconnections and support continuous operations. It also creates trade-offs involving , emissions, and cost.
Grid constraints led utilities, operators and planners to consider flexible operations, storage, onsite generation and arrangements that could reduce or during system stress. These approaches can provide near-term flexibility but do not substitute for firm, long-term grid capacity.
began incorporating much faster load growth from data centres and other large computing facilities. Higher forecasts required utilities and grid operators to reconsider generation, transmission, and . The revisions showed that AI demand was becoming a rather than only a technology-sector issue.
encountered lengthy studies and construction requirements before receiving permanent grid connections. Data-centre buildings and computing equipment could be delivered faster than substations, transmission lines and generation. This made an increasingly valuable to developers and investors.
Utilities and grid planners faced a surge in proposed as AI projects expanded. Many requests exceeded the size and speed of ordinary commercial connections. The resulting planning pressure exposed a mismatch between AI developers’ construction schedules and the electric system’s ability to add and wires.