How much electricity does AI use

Every AI answer, image, or model update is the result of billions of mathematical operations running on specialized chips. Those chips draw power while they work, and the buildings that hold them draw still more power for networking, storage, and cooling. The result is that artificial intelligence is, at the physical level, an electricity story before it is anything else.

When people ask how much electricity AI uses, the honest answer is that it is hard to measure AI on its own, because AI shares data centers with ordinary cloud computing, storage, and streaming. The clearer picture comes from looking at total data center electricity, which has become one of the fastest growing parts of the power system and which researchers track carefully.

That framing matters because it keeps the discussion grounded. Instead of guessing at the cost of a single chat reply, it is more useful to look at measured totals from credible sources and watch how quickly they are changing.

Two trends sit behind the rising totals. The first is scale. Modern AI models are larger and more capable than earlier ones, and running them takes more compute per task. The second is adoption. As more people and companies use AI tools every day, the number of requests served keeps climbing, and each request consumes power.

The International Energy Agency notes that electricity use by accelerated AI servers has been growing around 30 percent per year. In the United States, Lawrence Berkeley National Laboratory projects that data center electricity could reach 325 to 580 TWh by 2028, depending on how fast the buildout continues. These are not distant forecasts. They describe a curve that is bending sharply right now.

There is also a feedback loop. As AI becomes more useful, it gets embedded in more products, which draws in more users, which raises the number of inference calls, which lifts power demand again. Each turn of that loop pushes electricity higher.

At about 1.5 percent of global electricity in 2024, data centres are still a modest slice of total demand. The reason the topic gets so much attention is the speed and the concentration. Demand is rising quickly, and it tends to cluster in a few regions where land, power, and connectivity come together.

That concentration is why electricity, not just chips, now shapes where and how fast AI can expand. A region can have hardware available and still wait years for the grid capacity to run it. The IEA notes that the United States alone accounted for roughly 45 percent of global data centre electricity in 2024, a sign of how uneven the map has become.

So the right way to read the numbers is less about the global percentage and more about local strain. In the places where AI capacity is being built, the demand is intense, fast-growing, and very real.

Adding large new electricity loads is not as simple as plugging in. A data center that needs steady, heavy power requires new transmission, substation upgrades, and generation that can run around the clock. Those projects take years to plan and approve, which is why a site can sit ready with hardware while it waits for the grid to catch up.

This is also why location has become strategic. Operators look for places where power is available, affordable, and reliable, and where the local grid can absorb a large new customer without strain. The IEA notes that the United States accounted for roughly 45 percent of global data centre electricity in 2024, a sign of how the map is concentrating around regions that can deliver power at scale.

None of this means the growth stalls. It means the pace is set by the slowest physical input, and for now that input is increasingly power rather than chips. Understanding that shift is the key to reading where AI capacity will actually appear next.