Study

Power Shortages From A.I. Data Centers: The Gas Power Comeback Has Already Begun

co-produced by AI28. September 2026, 09:24

A gas-fired power plant illuminated at night — A gas-fired power plant at night. © American Public Power Association on Unsplash

Anyone building an A.I. data center has so far expected the power grid to deliver, at any time and at full capacity. An international research team led by TU Wien is now challenging that principle. In a study just published in the journal Joule, the researchers conclude that data centers must become flexible electricity customers if their expansion is to be compatible with the energy transition. Otherwise, they warn, artificial intelligence threatens to bring about a renaissance of fossil fuel power plants.

Gas Turbines on Trucks

How far the industry is currently willing to go is illustrated by a case from the United States that TU Wien cites as its starting point: the company xAI ran gas turbines mounted on trucks in Southaven, Mississippi, to power a data center just across the state line in Memphis, Tennessee. A power plant would have required environmental permits; the mobile turbines ran without the necessary air quality approvals. The case is still being disputed, and we have already reported on the environmental concerns around the Colossus supercomputer.

The numbers behind it: 39 percent of the gas power plant capacity that was in planning in the United States at the end of 2025 is intended for data centers, according to the study. Even in Norway, where electricity comes almost entirely from renewable sources, the grid is not strong enough everywhere to meet the demand of large data centers.

“We see that data centers are not necessarily built where electricity is cheapest. For many companies, the decisive factor is: Where is the grid well developed? Where can you get a suitable grid connection as quickly as possible?” says Reda El Makroum of the Institute of Energy Systems and Electrical Drives at TU Wien.

Throttling for at Most 35 Hours a Year

The team led by Mr. El Makroum worked with research groups from other universities to examine how energy systems and data centers can interact in the future. The key recommendations:

Flexible consumers instead of fixed load: In grid planning, data centers have so far been treated as ordinary consumers that must receive the desired power at all times. That principle, the team says, should be reconsidered.
Shifting computing tasks: Training large A.I. models can be moved in time, for instance to periods when plenty of renewable energy is available. Data centers could thus even help stabilize the grid.
New contract models: One option is contracts with a guaranteed base capacity and an additional capacity that may be temporarily reduced when the grid is overloaded. A model calculation for U.S. sites shows that power draw would have had to be throttled for at most 35 hours per year at the sites studied to stay within existing grid capacity.
Better data: “To plan grid infrastructure well, we would need to know much more about how electricity consumption changes,” Mr. El Makroum says. Anonymized operating data and information on the hardware used could help.

Huang: ‘We’re Going to Use a Lot More Fossil Fuel’

That the industry knows the problem but is solving it differently is evident in the remarks of Nvidia chief Jensen Huang. “There’s no question that in four or five years’ time, we’re going to use a lot more fossil fuel,” he said recently in an interview with Ezra Klein of The New York Times, as we reported. The United States, he argued, underinvested in energy production out of climate concerns and now lacks the sustainable capacity for the A.I. boom. Gas, in his view, is a bridge, and A.I. will ultimately accelerate the energy transition because data centers steer capital toward solar, batteries, nuclear and hydropower. It is precisely that sequence, fossil first, clean later, that the TU Wien study calls into question with its flexibility approach.

Who Pays for Grid Expansion?

The study also addresses the question of what a region gets out of it. Data centers bring large investments, but they create jobs mainly during the construction phase. At the same time, the necessary grid expansion can cause additional costs. The researchers therefore propose binding agreements to ensure that the local population does not end up footing that bill. Using the waste heat from data centers could also benefit municipalities.

The team’s conclusion is nonetheless not a negative one: data centers can become part of a sustainable energy system, provided that planning starts early on how they fit the power grid, ecological goals and the needs of the local population. The paper, “Governing AI infrastructure on decarbonizing grids,” is freely accessible.