The electrical power problem that will decide the AI race

Every leap in artificial intelligence is, underneath the silicon, a bet on electrons. The models grab the headlines and the GPUs absorb the capital, but the binding constraint on the next decade of computing is neither chips nor talent. It is power; firm, abundant, affordable, and ideally clean. The companies racing to build frontier AI have quietly become some of the largest prospective buyers of electricity on the planet, and the scramble to secure megawatts has become the defining infrastructure story of the era. The scale is hard to overstate. US data centers drew roughly 25GW in 2024; BloombergNEF projects that figure could reach 106GW by 2035, more than quadrupling the sector's appetite in a decade. McKinsey pegs US data-center consumption at 606TWh by 2030, up from about 147 in 2023, or close to 12 percent of the nation's electricity. The Department of Energy and Lawrence Berkeley National Laboratory estimate data centers account for roughly 50GW of the new peak capacity the grid must add by 2030, and could consume up to 12 percent of US electricity by 2028. Other analysts project US data-center electricity demand could grow by as much as 130 percent by 2030. Globally, the International Energy Agency expects data-center electricity use to more than double to roughly 945TWh by 2030, with the United States driving the single largest increase. Why power, not silicon, is the bottleneck The economics of an AI cluster are brutal in their simplicity: a GPU that is not computing is a depreciating asset burning money. Large training and inference clusters must run as close to continuously as possible to justify their staggering capital cost, which means the electricity behind them has to be available every hour of every day. Intermittency or curtailment directly erodes the return on billions of dollars of compute. A single 100MW IT campus running flat out consumes on the order of 876GWh a year, and can produce tens to hundreds of millions of AI tokens per second. That appetite has collided with a grid that cannot move fast enough. Interconnection queues stretch for years, transmission is congested, and new high-voltage lines take most of a decade to permit and build. The result is a pivot that would have seemed exotic two years ago: hyperscalers are increasingly building generation directly alongside their facilities, also known as "behind the meter," to sidestep the grid and guarantee supply on their own timeline. The power menu: Eight ways to feed a data center A developer's choices differ enormously along the two axes that matter most: reliability, can the source deliver around the clock, and sustainability, what that power costs the climate. Today's US mix is led by natural gas near 43 percent, then nuclear around 19 percent, coal near 16 percent, wind around 10 percent, solar near seven percent, hydropower around five percent, and geothermal at a sliver under one-half of one percent. Natural gas is the incumbent and the default; dispatchable, fast to build, politically favored, but carbon-intensive, with a wave of new gas plants now being announced specifically to serve data centers behind the meter. Oil and diesel survive only as expensive, emissions-heavy backup. Coal remains a legacy baseload source but is both the dirtiest option and a retiring fleet. Nuclear is the gold standard for firm, carbon-free power, running above 90 percent of the time, but new plants take five to ten years or more, hence the rush toward reactor restarts and small modular designs. Hydropower is clean and partly firm but geographically capped and increasingly drought-exposed. Wind and solar are the cheapest new energy and nearly carbon-free, but capacity factors of 30–40 percent and 20–30 percent betray their core problem for always-on compute: the wind drops and the sun sets. Geothermal is the sleeper, under one percent of supply today, yet the one source that combines a capacity factor above 90 percent, near-zero emissions, a small water and land footprint, and independence from the transmission grid. The table below sets the contenders side by side.