How Much Electricity Do AI Data Centres Use?
Data centres used about 1.5% of the world's electricity in 2024. Here is what the IEA and the US Department of Energy report, what drives the growth and how uncertain the forecasts are.
In this article
Data centres, the buildings full of servers behind websites, apps and AI chatbots, used about 415 terawatt-hours (TWh) of electricity in 2024, according to the International Energy Agency (IEA). That is roughly 1.5% of all the electricity used in the world. AI is one reason the number is rising, but it is not the only one, and the forecasts come with wide error bars. This article sets out the main figures from two public reports and explains what sits behind them.
What a terawatt-hour is
A kilowatt-hour (kWh) is the unit on a home electricity bill: a 1,000-watt heater running for one hour. A terawatt-hour is one billion kilowatt-hours. So 415 TWh is 415 billion kWh. The size only makes sense next to something else, which is why reports quote it as a share of total electricity use.
The global picture from the IEA
The IEA's Energy and AI report says data centre electricity use has grown by about 12% a year over the last five years. In its Base Case, consumption reaches about 945 TWh by 2030, a little under 3% of global electricity. That growth of roughly 15% a year from 2024 is, according to the IEA, more than four times faster than the growth of electricity use in all other sectors combined.
The growth is not spread evenly. The IEA says the United States and China together account for nearly 80% of the global increase to 2030. It projects US data centre use rising by about 240 TWh (up 130%), China's by about 175 TWh (up 170%), Europe's by more than 45 TWh (up 70%) and Japan's by about 15 TWh (up 80%).
The US view from the Department of Energy
On 20 December 2024 the US Department of Energy published a report written by Lawrence Berkeley National Laboratory, the 2024 Report on U.S. Data Center Energy Use. It found that data centres used about 4.4% of US electricity in 2023. It projected a share of roughly 6.7% to 12% by 2028. The wide range is itself a message: the authors could not say how fast demand will grow.
Where the electricity goes inside a data centre
Not all of a data centre's power runs AI. The IEA gives an approximate split of the net increase in demand to 2030:
- Accelerated servers, the specialised chips used mainly for AI, account for almost half of the increase
- Conventional servers account for about 20%
- Other IT equipment accounts for about 10%
- Cooling and other infrastructure account for about 20%
Accelerated servers are growing the fastest, at about 30% a year in the IEA's Base Case. Cooling matters because the chips turn nearly all the electricity they use into heat, and that heat has to be carried away or the equipment fails.
Why forecasts disagree
The IEA stresses that there is substantial uncertainty about both today's and tomorrow's data centre use, so it publishes scenarios instead of one forecast. For 2035 it describes three:
- Lift-Off: above 1,700 TWh, about 4.4% of global electricity
- High Efficiency: about 970 TWh, about 2.6%
- Headwinds: a plateau near 700 TWh, under 2%
The scenarios differ in how quickly AI is adopted, how many chips can be built and delivered, and how much more work each unit of electricity can do as hardware and software improve. None of these can be known in advance.
A national total hides local effects
The IEA notes that data centre growth is under 10% of the growth in global electricity demand from 2024 to 2030. In other words, at world level it is a small slice. But data centres cluster in particular places, so a single region can see a large jump in demand. The IEA says this can make connecting them to the grid harder than the global figures suggest.
What the figures do not tell you
- They do not say how much power a single chatbot question uses. Neither report gives that figure, and estimates vary with the model, the hardware and the length of the answer
- They do not separate AI from other cloud work such as video streaming and storage, except in the broad split above
- They are projections, not measurements of the future
To follow the story as it develops, look for updated editions of the two reports, which restate the numbers as new data arrives.
Sources
Share
Keep reading
Related articles
What Is a Context Window? Why AI Chatbots Forget Things
A context window is the amount of text an AI model can look at in one go. Learn what counts toward it, how tokens relate to words, and why long chats can lose track of earlier details.
How Speech Recognition Turns Your Voice Into Text
From the 1952 digit recognizer to today's neural networks: how speech recognition converts sound into words, and how its accuracy is measured.
Open-Weight vs Open-Source AI Models Explained
Open-weight models share their trained numbers; open-source AI, by one formal definition, shares much more. Here is what each term means and why the difference matters.