What energy sources will supply the AI boom?

The AI boom will be powered by an all-of-the-above portfolio whose mix changes over time. Through 2030, most added supply is likely to come from solar, wind, and natural gas, backed by batteries, transmission, the existing grid, and harder-running gas and coal plants where connections lag. Existing nuclear can help through extensions, uprates, and restarts; new nuclear and enhanced geothermal are more plausible at scale in the 2030s. The exact mix will vary by region and depends heavily on uncertain demand, grid buildout, clean-energy deployment, efficiency, and flexible computing. Corporate clean-energy contracts may not match electricity consumed locally hour by hour.

What energy sources will supply the AI boom?

Bottom line

An “all-of-the-above” portfolio will supply it, but the mix changes over time. Through 2030, the most likely marginal sources are renewables—especially solar and wind—plus natural gas, supported by batteries and the existing grid. Existing coal and gas plants will also run harder where new projects cannot connect fast enough. Existing nuclear can contribute through life extensions, uprates, and restarts; new nuclear and next-generation geothermal are more plausible material contributors in the 2030s than immediate solutions.

The global mix: cleaner, but still fossil-heavy

The IEA’s base case projects electricity physically supplying data centers to rise from 460 TWh in 2024 to more than 1,000 TWh in 2030 and 1,300 TWh in 2035. Renewables meet nearly half of growth to 2030, while gas and coal together meet more than 40%. By 2035, low-emissions sources could supply roughly 60% and fossil fuels about 40%—a reversal of today’s approximate ratio, not a fossil-free outcome.

Regional grids matter. In the United States, gas is currently the largest data-center power source and the IEA expects it to add more than 130 TWh through 2030, versus 110 TWh from renewables. China’s near-term supply remains more coal-heavy; Europe’s additional demand is expected to be served mainly by renewables and nuclear. Corporate clean-energy contracts therefore should not be confused with the local physical mix consumed hour by hour.

Why gas and existing fossil plants matter near term

Data centers want continuous power and are being built faster than generation, transmission, and interconnections. Rhodium finds that, under a faster-growth U.S. scenario, these bottlenecks leave little time for extra capacity by 2030, causing existing gas and coal plants to serve 55–85% of demand above its baseline. EPRI likewise finds that least-cost procurement under current U.S. policy favors substantial new gas, whereas 24/7 carbon-free procurement shifts investment toward wind, solar, nuclear, and storage.

Onsite gas may bypass some connection delays, but it is not a frictionless answer: the IEA estimates 15–27 GW could serve data centers by 2030, while noting turbine shortages and a need to overbuild onsite capacity by 30–70% for reliable service. Heavy gas dependence also exposes operators and consumers to fuel-price volatility and emissions.

Renewables will supply the largest growth share—but need integration

Solar and onshore wind are modular, comparatively fast to deploy, and already dominate much new capacity. RMI says projects combining them with batteries can typically be built in under two years, and co-location at existing grid connection points could serve more than 50 GW of U.S. data-center load. Their variability means the practical product is not solar or wind alone, but a portfolio of renewables, storage, transmission, flexible load, and some firm generation.

Efficiency and flexibility can materially shrink the supply problem. RMI estimates that allowing new U.S. data centers to curtail just 0.5% of annual load could accommodate nearly 100 GW without expanding generation. Batteries are also becoming essential for second-to-second AI load swings; the IEA sees 20–25 GW potentially installed at data centers globally by 2030.

Nuclear and geothermal: important, but mostly later

Nuclear’s best near-term contribution is preserving, uprating, or restarting existing reactors. New small modular reactors are more likely after 2030. Carnegie calculates that all announced U.S. hyperscaler-backed nuclear projects—about 13 GW—would still cover less than 20% of a mid-range increase in data-center demand through 2035, before accounting for construction, licensing, financing, fuel, workforce, and waste risks.

Enhanced geothermal offers firm, low-carbon power with a high capacity factor and broader geographic potential than conventional geothermal. Rhodium estimates it could economically serve 55–64% of modeled hyperscale growth in the early 2030s under historic clustering, or all modeled growth if sites follow the best resources. That is an opportunity estimate, not a forecast: commercialization, drilling costs, siting, permitting, and developers’ willingness to relocate remain decisive.

What could change the answer

Demand itself is highly uncertain: EPRI projects data centers at 9–17% of U.S. electricity use by 2030, while the IEA emphasizes rapidly improving efficiency alongside much more energy-intensive AI uses. Faster grid reform and clean-energy deployment would raise renewables’ share; cheap gas or delayed clean projects would raise fossil generation; successful SMRs or geothermal scale-up would increase firm clean supply after 2030. The most defensible conclusion is therefore near-term solar, wind, gas, batteries, and existing generators; longer-term more nuclear and geothermal—enabled throughout by transmission, storage, efficiency, and flexible computing.