Power-hungry AI data centers weigh viable energy sources
AI data centers are encountering political resistance in some communities, but they'll be built elsewhere. Survival depends on the power sources chosen and their availability.
U.S. data centers used about 176 terawatt-hours (TWh) of electricity in 2023, or 4.4% of the national total, according to Lawrence Berkeley National Laboratory's 2024 report to Congress. By 2028, that power demand could nearly double or even triple, reaching 6.7% to 12% of the electricity the U.S. will have generated.
There currently isn't enough grid capacity to absorb this much demand, especially for the larger data center campuses, said Parag Nathaney, an engineer at a U.S. public electric utility company. Spare capacity is shrinking as AI-driven demand outpaces new supply, and adding substantial new capacity now takes five to seven years. In addition, local residents are increasingly pushing back, fearing their community would bear the perceived increased energy costs, pollution and other environmental issues associated with data center buildouts.
As a result, data center planning teams in the process of selecting an energy system to run behind the electricity meter must balance several factors, including time to market, power grid demands, equipment availability, ratepayer pushback and environmental footprint. And the cheapest energy source available isn't necessarily able to supply the amount of power needed to handle the load. Estimates prepared for the U.S. Energy Information Administration (EIA) in 2024 estimated onshore wind and utility solar at about $1,500 per kilowatt (kW) versus roughly $850 for a large gas combined-cycle plant, which then incurs ongoing fuel charges. Battery storage pushes that utility solar plant past $2,000 per kW, and offshore wind runs nearly $3,700, according to EIA 2024 estimates.
Andrei Romanescu, CMO at LumaDock, a provider of GPU servers for data centers, said the local resistance his company encounters "cluster around land, water, noise and what a new load does to household electricity prices." Carbon emissions are further down the list of concerns.
Timing is everything
"The market is selecting for speed, not for the cleanest electron," said Alex Marshall, group business development and marketing director at Clarke Energy and vice president of the Cogen World Coalition, a global community of companies and institutions focused on cogeneration. Data center planners confront a range of buildout issues, including energy costs, energy waste, property values and emissions, but the concern gaining greater attention is the time it takes data centers to power up.
Time to power remains the primary objective for the data center projects we're participating in.
Michael StadlerCo-founder and CTO, Xendee
"Time-to-power remains the primary objective for the data center projects we're participating in," said Michael Stadler, co-founder and CTO of microgrid decision support platform provider Xendee. Some data center operators pay well above the going rate for power even though it might not arrive for quite some time. The cheaper power options typically lose out to the more expensive ones that can be connected faster because data centers must also consider the lifespan of their AI chips and the cost to replace them.
Lawrence Berkeley National Laboratory pegs the median wait time for a data center power connection at more than five years. Coordinating utility interconnection, which can range from three to seven years, with data center buildout cycles that might take just one to two years, exacerbates the time-to-power problem.
As a result, operators are starting to include minigrids in their data center infrastructure. These small-scale localized electricity systems operate and generate power independently without straining and adding costs to community power grids.
Still, capital costs and fuel prices are moving targets. Data centers therefore need to conduct multiyear analyses of their operations "and strategically adapt their on-site power strategy to avoid stranded assets or unexpected costs," Stadler advised.
Choosing the right power source is a critical part of that strategy. Some of the more common data center energy alternatives are listed here, along with more detailed information on 14 power source options in the accompanying table.
Natural gas
Gas is winning the battle of power sources, particularly in the U.S., thanks to its abundance, availability and about three million miles of pipeline. New combined-cycle gas turbines are the most efficient gas option, but capital costs have roughly doubled since 2022, ranging from $2,000 to $2,500 per kW, Nathaney said. Major gas turbine manufacturers like GE Vernova, Siemens Energy and Mitsubishi Power are taking orders with delivery times approaching 2030 or later.
It's a speed and reliability play, not a decarbonization one.
Alex MarshallGroup business development and marketing director, Clarke Energy
Reciprocating internal combustion engines as an on-site bridge or primary power source can come online much faster. They commission in 12 to 18 months, start in about two minutes and maintain efficiency with partial loads better than turbines, which matters when GPU clusters fluctuate, Marshall said.
Another option is Jenbacher power and cogeneration systems, which run on natural gas, hydrogen-rich blends or renewable gases. "It's a speed and reliability play, not a decarbonization one," Marshall noted. Natural gas emits carbon dioxide (CO2) and nitrogen oxide (NOx) emissions, which can create permitting issues in some communities.
Very few new coal plants are being commissioned for AI in most parts of the world, but existing coal-fired power plants are expected to fill some of the electrical power gaps for data centers. Natural gas and, in some geographical areas, coal should continue to cover peak and baseload demand as AI workloads ramp up, said Damir Špoljarič, founder and managing partner of Gi21 Capital, which is developing a large European AI data center platform.
Fuel cells
Fuel cells have become a credible middle-of-the-road option. They run on gas, are quiet and eliminate local air pollution, although they still emit CO2. Bloom Energy announced last year it planned to deliver onsite power to Oracle AI data centers within 90 days, and Oracle expanded its agreement with the energy company in April.
In a related development, Oracle and BorderPlex Digital Assets announced in April that Project Jupiter -- a massive 1,400-acre AI data center campus under construction in Doña Ana County, N.M. -- could use as much as 2.5 gigawatts of gas-powered fuel cells from Bloom to replace previously planned gas turbines and diesel generators, and consolidate the facility into one single microgrid campus. However, there could be a significant pipeline delay in delivering natural gas to the facility.
Fuel cells can theoretically run on hydrogen or ammonia, but hydrogen is less dense and can damage existing pipes. In addition, Amazon canceled plans to use Bloom gas-powered fuel cells in three of its Oregon data centers in 2024 due to the state's concerns the fuel cells would increase the data centers' carbon footprint. Fuel cells also carry higher recurring costs than turbines, because the Cisco Catalyst stacks of network switches need periodic replacement.
Solar and wind
While solar and wind don't emit pollutants into the atmosphere, solar requires a large land footprint that invites community pushback similar to the resistance surrounding the amount of real estate data centers require. Solar panels typically last 25 to 30 years, but they carry a cumulative retired photovoltaic waste stream at the end of life that can create economic, environmental and regulatory hurdles. Solar power can only operate efficiently during sunny daylight hours.
National Renewable Energy Laboratory cost modeling put the cost of recycling a solar panel at roughly 20 times the cost of landfilling it, and the International Renewable Energy Agency projects as much as 78 million tons of global panel waste by 2050.
Wind turbines can produce the same amount of energy as solar using a much smaller footprint, but they also face community resistance due to their size and appearance. However, wind is an all-purpose energy source for powering IT equipment and cooling systems. Although it can operate any time of day, there has to be sufficient wind velocity.
Geothermal
Geothermal energy promises a small footprint, low waste volume, no direct emissions and steady output. The EIA estimates put a 50 megawatt binary-cycle plant at nearly $4,000 per kW with no NOx, sulfur dioxide or CO2. Its fixed operating costs are among the highest compared to other power sources. Other drawbacks are longer time to power and difficulty finding acceptable geographical locations. In addition, poorly managed geothermal operating fluids can be a water pollution risk, which could be eased by newer drilling approaches and heat engines.
Small and large nuclear power plants
Until recently, nuclear was generally considered too risky and expensive. But the timeline for nuclear has moved up, owing to new regulatory approaches and substantial investment.
"SMRs [small modular reactors] may be commercially available earlier than we were thinking at this time last year," Stadler noted. His group initially estimated nuclear power availability would be about 2035, but it now sees 2030 as more realistic. Small nuclear plants aren't necessarily comparatively cheaper to build and operate than large plants.
Fission reactors produce highly toxic radioactive waste. The volume of waste is much smaller than other kinds of energy waste, but after several decades, the nuclear waste disposal problem hasn't been entirely solved. However, Finland's Onkalo repository for spent nuclear fuel received a favorable safety assessment from the Radiation and Nuclear Safety Authority of Finland in August and could be operational later this year to help address the waste disposal issue.
While some experts consider SMRs to be the future of nuclear energy because of its potential cost and safety advantages over other reactors, SMR waste per megawatt-hour can be 2 to 30 times higher than a more common conventional reactor, according to an April 2026 paper by Progress in Nuclear Energy.
Powering a 2030 AI data center
Exciting technologies typically get the most press, but the likely power source trajectory is more mundane. "Based on our modeling," Stadler said, "we still anticipate natural gas being the primary fuel source, with SMRs starting to enter the picture after 2030, supplemented by solar photovoltaic and battery storage where it makes economic and operational sense."
The biggest wildcard is not a single technology but the speed of policy and market design.
Damir ŠpoljaričFounder and managing partner, Gi21 Capital
Hydrogen has been proffered as a green energy carrier but hasn't emerged as a near-term priority in Xendee's projects, Stadler added. Most gas engines installed today can also run on biogas, so operators can shift to that power source if it makes sense.
Enhanced geothermal could also play a role, thanks to new techniques that can access usable heat at more sites. The U.S. Department of Energy estimates it could deliver 90 gigawatts of firm (weather-independent, low-emissions) power to the U.S. grid system by 2050, compared to just 4 gigawatts today. Nuclear fusion has been bandied about as cheap, clean and affordable energy option since the 1950s, but the goalposts keep getting pushed back for AI data centers.
"The biggest wildcard is not a single technology but the speed of policy and market design," Špoljarič said. Nuclear energy, long-duration storage and demand response could gain wider usage, if flexibility and low-carbon firm capacity become priorities. Until then, natural gas will continue to be the most widely used energy source for AI data centers.
George Lawton is a journalist based in London. Over the last 30 years, he has written more than 3,000 stories about computers, communications, knowledge management, business, health and other areas that interest him.
