OKLAHOMA CITY—Meeting the power demand from AI will likely require the United States to increase natural gas production by 10%-15% by the early 2030s, according to a study published on Nov. 13 by the Hamm Institute for American Energy. This will coincide with expanding LNG exports, which will drive similar supply growth, the study finds.

Data center construction is accelerating rapidly, the study observes. “Until recently, the U.S. data center industry had been adding a total of about 500 MW of new capacity annually over the prior decade. That rate tripled in 2022 and has doubled again since,” it relates. “Thus far, construction underway is on track to double again in 2025.

“The capacity added in 2025 alone, likely over 10 GW, is comparable to the peak daily electric demand of New York City,” it remarks.

The study, “The Rise of AI: A Reality Check on Energy and Economic Impacts,” is authored by Mark P. Mills, the executive director of the National Center for Energy Analytics. Mills is also a distinguished senior fellow at the Texas Public Policy Foundation, a distinguished fellow of the Hamm Institute for American Energy, and a contributing editor for the Manhattan Institute’s City Journal. He has published several books that explore the relationship between technology and energy, including 2021’s The Cloud Revolution: How the Convergence of New Technologies Will Unleash the Next Economic Boom and a Roaring 2020s, as well as 2020’s Digital Cathedrals: The Information Infrastructure Era.

“Forecasters, pundits, and policymakers have all been surprised, even shocked, by the magnitude and speed of electricity demand emerging to power AI and the hundreds of billions of dollars of data centers under construction and planned,” Mills commented after the study’s release. “The economic and strategic importance of the AI race sidelines the idea of an energy transition and highlights the old 'all-of-the-above' approach, especially the immediate needs for more natural gas.”

Data centers’ total impact on electricity demand extends beyond the buildings themselves. “The wired (fiber) and wireless (cellular) networks that connect data centers and users consume a quantity of electricity rivaling data centers,” the study illustrates.

Manufacturing the centers’ components also requires energy. “Fabricating global semiconductors alone uses about one-third as much electricity as global data centers,” the study details. “Energy is also used to produce all the related digital hardware.”

But the largest demand increases will come from the productivity gains AI enables. “While some analysts believe AI can boost the U.S. economy to a 4% (annual) productivity growth rate, the highest level (since WWII)—and there is good evidence to support that optimism—consider instead the implications of merely restoring the rate to the postwar long-run average of 2.2%. That would, arithmetically, induce a cumulative $10 trillion of greater economic growth over the coming decade than is currently forecast,” the study calculates.

That extra wealth will enable people to engage in activities that consume energy, the study suggests. It estimates that the $10 trillion GDP uplift would “lead to increased overall energy use equivalent to about five billion barrels more energy over the next decade. Such a wealth-induced increase in energy will be far greater than the quantity of energy needed to power the wealth-creating AI.

“If the promise of AI-driven biological discovery does lead to longer, healthier lifespans, that too will induce more energy use,” the study adds.

Complications

The study acknowledges that predicting when and how much AI will affect demand can be tricky. However, the topic is important enough to policymakers, utility planners, investors, and vendors that many market observers have weighed in.

“Myriad forecasts and analyses are in circulation, including from organizations such as Goldman Sachs, Brookfield Management, DNV, SemiAnalysis, Lawrence Berkeley National Laboratory, Rand, Rystad Energy, S&P Global, Bain & Company, Thunder Said Energy, and of course the International Energy Agency,” Mills observes.

These analysts’ projections vary widely, with some putting data centers’ annual electricity demand in 2030 around 200 terawatt hours while others see it reaching 600 terawatt hours.

To provide a minimum, the study looks at the data centers that are under construction or have construction permits. As of mid-year 2025, that list suggests that more than 1,000 data centers will come online in the next five years with a capacity of 75 gigawatts. “At expected utilization levels, 75 gigawatts is about 550 terrawatt hours of new demand,” Mills says.

While some of those projects will face delays and others may be cancelled, Mills says these losses will likely be compensated for by other data centers that have not yet been announced. He predicts that AI infrastructure will require 75-100 gigawatts of new electricity generation by 2030.

“The two biggest variables in forecasting future digital energy demands are guessing how many unanticipated new uses will emerge for AI tools and, on the other side of the equation, estimating how much more efficient the underlying AI hardware will become,” Mills writes. “There is no doubt that far greater efficiency gains are coming rapidly to all relevant compute technologies: GPUs, CPUs, memory chips, and information transport, as well as power management and cooling.”

In addition to improving hardware, Mills says the AI sector will find ways to train and operate models more efficiently. “Innovators have already deployed better techniques for learning (different logic, new math) and means for curating data to avoid brute-force data crunching,” the study says. “Since the primary fuel for AI software is a massive quantity of data, techniques to select, pre-qualify, meter-as-needed, and move data save enormous amounts of energy. Samsung researchers, for example, recently published a new technique to run a large language model that increases learning speed twelvefold and cuts energy use.”

But “invoking efficiency as a reason to expect a moderation of overall AI energy use gets it backward,” the author warns.

“History shows that efficiency gains have led, overall, to increased—not decreased—energy use,” he explains. “Aircraft, for example, are three times more energy efficient than the first commercial passenger jets. That efficiency didn’t save fuel. Instead, it propelled a fourfold rise in overall aviation energy use since then.”

Who Benefits?

Data centers tend to go to areas where power is already available or should become available. “Based on a survey of existing, permitted plans, over three-fourths of the expected 75 GW of new data center demand that has already been permitted is taking place in about a dozen states, though nearly all have some activity,” Mills writes. “The top six states, in order, are Virginia, Texas, Oregon, Arizona, Georgia, and Ohio. California is number seven—despite being home to the world’s first data center, built by Exodus Communications in Santa Clara in 1998.

“However, as the early players exhaust local grid capacities, the trends show new plans and applications increasingly chase regions with underlying energy resources that can accommodate the scale and velocity of demand,” he says. “Thus, plans are increasingly found not only in Texas but also in other energy-rich states such as Louisiana, Oklahoma, Pennsylvania, and North Dakota.”

In the short term, the study says rising power demand will increase coal plant utilization and potentially extend the life of plants that would otherwise be retired. “In the first half of 2025, U.S. coal-fired generation was 15% higher than in the same period of 2024,” it shares. “The odds are that coal generation will not shrink in the near term and may even continue to slightly expand, not just because of administration policies and executive orders but also due to frontline realities for utilities.

“However, based on decisions visible thus far at dozens of projects under construction—e.g., Meta’s 1.5 GW project in Louisiana using natural gas turbines and Cat reciprocating engines—and visible in vendor order books, natural gas generation is the primary source of new electricity planned for near future data centers,” it finds.

Industry Readiness

It’s possible for the natural gas industry to meet that demand, the study assures. “For illustration, assuming the low end of additional data center capacity at 55 GW and that it is entirely supplied with natural gas, roughly 10 billion cubic feet per day of additional gas would be required by 2030,” it says. “For context, the U.S. industry increased gas production and delivery by about 15 Bcf/d since 2017 for LNG export.”

Given the right price, Mills says the industry can supply the necessary gas. He says recent history also shows that as long as the permitting environment is reasonable, the industry is more than capable of expanding pipeline networks to get that gas where it needs to go.

However, labor could be a concern. “The construction of most infrastructures involves similar activities and skills, regardless of whether it entails a chip manufacturing plant, a shipyard, highways, houses, pipelines, power plants, or data centers. All such construction involves the need for similar people with comparable skills, from heavy equipment and vehicle operators to electricians, plumbers, welders, and carpenters,” he says.

Assuming data center construction continues at $50 billion a year, that work alone will require about 250,000 people in the skilled trades, Mills estimates. “This will come at a time when the energy supply industries will need additional skilled labor,” he says.

“The new demand for labor will likely accelerate the use of emerging options for such things as semi-automated and fully automated construction machines, including robotrucks (the latter have long been used at mine sites),” the study predicts. “The effect will be to free up people in that limited labor pool to do work on other difficult-to-automate tasks.”

In addition to deploying new technologies, Mills says companies will expand the labor pool by encouraging skilled workers to delay retirement and bringing more young people into the industry. “Trends suggest this has begun. Shop classes are returning to many high schools and apprenticeships are growing,” Mills reports.

The full study provides more detail on the arguments summarized above, discusses whether AI represents a bubble or an enduring technological shift, explains why natural gas is the preferred option for meeting AI’s near-term power needs, and outlines some of the steps utilities and other stakeholders are taking to mitigate AI’s potential impacts on grid reliability. Almost every argument the study makes is supported by graphs. To read it, see The Rise of AI: A Reality Check on Energy and Economic Impacts.