Natural gas supplies more than 40% of US data center electricity today and is expected to remain the largest source of new AI power through 2030
The Fusion Report has written a number of pieces on artificial intelligence (AI) and AI data centers over the past year, and one topic which is particularly timely is the question of, “what is powering these data centers?” While the use of clean (‘green’) energy for AI data centers originally had gotten significant headlines, the reality is that most of the power for AI data centers isn’t renewable energy sources. The purpose of this article is to look into what are the likely sources of power, now and in the future, for data centers in the United States, and whether that power will be from the grid or behind the meter.
Today: Powering AI Data Centers Is A Mix of Sources
Electrical power in the United States today is a mix of sources. Depending on what data you look at, there is variation between the time scale (2023 is the latest data from the US-EIA), and what is being measured (usage vs. generating capacity). The largest source of electrical power in the US is based on natural gas, contributing roughly 41% of all electricity generated. The second largest source is renewables, which are roughly one-quarter of all electricity, followed by nuclear power at 18% and coal at 17%. Petroleum and other gases are less than 1%. Within renewables, wind power today provides 10%, followed by solar at 8.6%, hydropower at 5%, and biomass and other renewable sources at 1%.
Since US AI data centers in 2026 are powered primarily by the regional electric grids to which they connect, their actual fuel mix varies by location rather than by a single dedicated “AI” supply. At the national level, natural gas is the largest source: the Energy Information Administration projects it will provide about 40% of U.S. electricity generation in 2026. Nuclear supplies roughly 18%, renewables about 25%, and coal about 16%; renewables combine wind, solar, hydropower, biomass, and geothermal generation. Recent estimates specific to US data centers likewise put natural gas above 40% of their electricity supply, followed by renewables at about 24%, nuclear at about 20%, and coal at about 15%. The reality is that even if a hyperscaler has signed a renewable-power contract, its servers generally still consume the instantaneous grid mix unless matched by dedicated or time-aligned clean generation.
US AI Data Center Electricity by Fuel Mix, 2030
By 2030, U.S. AI data centers are likely to be supplied by a more diverse electricity mix than today, but natural gas is still expected to be the largest incremental fuel source. AI facilities need very high load factors and near-continuous power, while grid interconnection and transmission additions often take years; as a result, gas-fired combined-cycle plants, gas turbines, and—in some cases—on-site gas generation are expected to provide firm capacity for much of the new load. The International Energy Agency projects that roughly 15–27 GW of on-site natural-gas capacity could be serving data centers globally by 2030, mostly in the United States. In one U.S. base-case projection, gas generation associated with data centers rises from just under 100 TWh in 2024 to well above 200 TWh by 2030. Gas’s role does not mean every AI campus will burn gas on site: most will still draw from regional grids, whose marginal supply is often gas-fired, particularly during periods when renewable output is low or demand is high.
The biggest factor making forecasting of the fuel types to power AI data centers in the next few years difficult is the speculative nature of US plans for building new power plants, especially “behind the meter” ones. According to the Global Energy Monitor, the United States has 378GW of new gas generation plants under proposal, more than the next nine countries combined. These projects range in planning levels from the early planning stages to construction, but many of these projects don’t yet have vendors named for gas turbines or other critical components; moreover, only about 30% of these early projects historically actually reach conclusion. The one confounding issue for US data centers is that, despite the large supply of natural gas in the United States, the data centers will be forced to compete with an export market in Asia and Europe, which is very hungry for liquefied natural gas.
Additionally, by 2030 behind-the-meter power plants for US AI data centers will have become popular as a way to bypass interconnection agreement delays. These systems will take the form of campus-scale microgrids, built next to or within large data-center campuses to bypass long interconnection queues and provide highly reliable power; examples of these include Elon Musk’s xAI facility in Memphis, Tennessee, the OpenAI/Stargate facility in Abilene, Texas, Meta’s Prometheus data center in Ohio, and others. The core of these microgrids will most often be modular natural-gas generation—either gas turbines for large blocks of capacity or reciprocating engines for faster, phased deployment—supplemented by medium-duration batteries, switchgear, controls, backup generation, and a grid connection where available. Some sites will also add solar, fuel cells, and potentially geothermal or other local resources, but these will generally complement rather than replace firm generation because AI loads operate around the clock. Forecasts suggest behind-the-meter solutions could supply roughly 25–33% of incremental data-center demand through 2030, while industry estimates indicate that 38% of data centers may use on-site generation at least in part; however, the exact share will depend on regional permitting, gas-pipeline availability, emissions rules, and the speed of transmission and utility interconnection upgrades. Nuclear solutions may appear in a limited number of projects, but widespread small-modular-reactor (SMR) deployment is more likely to be a post-2030 development.
Conclusion: Powering AI Data Centers in 2035 and Beyond
Beyond 2035, AI data centers will most likely continue to be powered by a portfolio of electricity sources, rather than a single fuel. Wind, solar, hydropower, and—in some regions—geothermal are poised to provide most of the incremental energy because they are generally the fastest-growing and often lowest-cost new generation options. For AI operators, however, buying annual renewable energy is not sufficient: training clusters and inference services need electricity every hour, and their concentrated demand can be measured in hundreds of megawatts or even gigawatts at individual campuses. That makes transmission, grid-scale batteries, long-duration storage, demand flexibility, and geographically diversified renewable procurement as important as the generating fuel itself. The International Energy Agency projects renewables to meet most additional global data-center electricity demand through 2035 and expects them to remain the dominant source of incremental supply afterward; its base case still has total data-center demand above 1,300 TWh globally in 2035.
The balancing and reliability layer will determine how clean this mix becomes. Existing and new nuclear power—including life extensions, large reactors where available, and potentially small modular reactors later on—offers firm, low-carbon output that matches data centers’ around-the-clock load, so it is likely to gain strategic importance after 2035. Natural gas will nevertheless remain a substantial bridge and backup fuel, especially in U.S. regions with constrained transmission or slow clean-power permitting; gas plants may increasingly operate fewer hours but provide capacity during low-wind, low-solar periods. Additionally, fusion power is also a potential dark horse in the AI Data Center electricity race, though even an early rollout of fusion is not likely before 2040 because of the new technology’s inherent technical and schedule risks. The most plausible post-2035 outcome is therefore renewable-heavy electricity backed by nuclear, hydro, storage, and flexible gas—rather than an entirely renewable or entirely nuclear AI grid. Even in the IEA’s scenarios, fossil fuels still account for roughly 40% of data-center electricity generation in 2035, underscoring that policies, grid build-out, and the commercial success of firm clean technologies will shape the longer-term result.
