Data center buildout surge projects $10 trillion capex and threatens grid capacity, raising power supply bottlenecks as binding constraint on AI infrastructure expansion.
Artificial intelligence has spent the past several years driving semiconductors into one of the market's hottest sectors. Now the physical world is beginning to push back. AI models may live in the cloud, but the servers running them require enormous amounts of electricity, cooling, transformers, substations, and grid connections. These physical components cannot be manufactured or permitted at software speed.
This mismatch is becoming one of the defining investment themes of the AI boom. Chip supply is no longer the binding constraint on the industry's growth. The next constraint on AI growth may not be Nvidia chips at all—it may simply be finding somewhere to plug them in.
Morgan Stanley estimates that U.S. data center IT power demand will rise from 9.19 gigawatts in 2025 to 78.57 GW by 2029, a 755% increase in four years, with acceleration beginning now. Demand is expected to nearly double to 17.96 GW in 2026, nearly double again to 35.46 GW in 2027, then reach 52.31 GW in 2028. Increasingly dense AI hardware compounds the challenge.
The investment bank's September research estimates that an Nvidia Vera Rubin rack could require roughly 234 kilowatts of power, while a future Rubin Ultra rack could consume around 600 kW. The firm projects new U.S. data center power needs during 2026 through 2028 at 97 GW.
Improving chip efficiency does not necessarily solve the problem. Faster chips lower the electricity cost of producing each AI token, making it economical to generate far more tokens. Morgan Stanley estimates tokens generated per watt could rise nearly sixfold between 2025 and 2028, yet total electricity consumption continues climbing.
The grid mathematics are stark. Morgan Stanley projects that data centers will require 97 GW of new capacity from 2026 through 2028, while only 21 GW of data centers are under construction and just 19 GW of grid capacity is available—leaving an initial shortfall of 57 GW. Alternative supplies such as onsite natural gas generation, fuel cells, and nuclear-related projects could reduce that deficit, but would still leave about 33 GW uncovered through 2028.
This does not mean the U.S. is destined for rolling blackouts. Rather, hyperscalers increasingly cannot assume the grid will provide power whenever a new data center is ready. Developers are already turning toward smaller onsite gas turbines because they can be deployed faster than traditional grid-connected generation. Enverus Intelligence Research expects 29.6 GW of behind-the-meter gas generation to be added by 2030, with data centers accounting for 88% of it.
Nvidia remains central to AI computing, but scarcity value is migrating downstream toward companies that can provide electricity and the equipment needed to deliver it. Power equipment companies such as GE Vernova, Eaton, and Vertiv Holdings, along with alternative power providers such as Bloom Energy, stand squarely in the path of this spending wave. Morgan Stanley argues that infrastructure and energy constraints should keep AI-enabling companies relevant even as AI adoption spreads into other industries.
AI's biggest problem is evolving from acquiring enough GPUs to acquiring enough megawatts. A projected 57-GW power gap makes electricity infrastructure one of the clearest second-order AI investment themes. Smart investors should still watch the chipmakers, but the next leg of the AI boom could increasingly reward the companies building the power plants, transformers, cooling systems, and electrical infrastructure that keep those chips running.