AI runs in data centres filled with specialised servers, memory, networking equipment and cooling systems. The biggest question is no longer whether AI needs more electricity, but how quickly grids can supply it.
The International Energy Agency estimated that data centres used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global consumption. Its base case projects consumption of around 945 TWh by 2030 — more than double the 2024 level and just under 3% of global electricity demand.
AI is the main reason growth is accelerating. Advanced servers use GPUs and other accelerators at much higher power densities than traditional computing. The IEA projects electricity use by accelerated servers to grow far faster than conventional servers.
Where will the power come from? No single source. The IEA expects renewables to meet close to half of the additional data-centre demand through 2035. Natural gas also grows, particularly where grids need dependable generation. Nuclear power is expected to contribute more later in the decade, while existing coal-heavy grids will still supply some regions.
The problem is local concentration. A data centre can be built in two or three years, while transmission lines and power plants can take much longer. A cluster of large facilities can therefore overwhelm a local grid even if data centres are only a few percent of global demand.
Cooling and water also matter, but electricity remains the main bottleneck. New transformers, substations, transmission and storage are often needed before servers can be switched on.
That is why technology companies are signing long-term power contracts and exploring nuclear, geothermal, gas and renewable projects. The AI race is becoming an energy-infrastructure race as much as a software race.