Chips, compute and data centres
AI energy use
In short
AI energy use is the electricity used to train and run AI models, plus the power for the servers, storage, networks and cooling around them. One query can be small, while the data centres that host the work draw power at industrial scale.
- Also called
- AI electricity use, AI energy consumption, Energy use of AI
- Topic
- Chips, compute and data centres
- Level
- Beginner
Training a model and running it both happen mainly in data centres, which the International Energy Agency calls the infrastructure for that work. In a modern data centre, servers account for about 60% of electricity demand on average, storage for about 5% and networking for up to 5%. Cooling ranges from about 7% in efficient hyperscale sites to over 30% in less efficient enterprise sites13.
One query has no single electricity cost. Energy use per AI task has fallen by at least an order of magnitude a year, and a simple text query now typically uses less electricity than a television left on for the same time. Video generation, reasoning and agentic tasks can use hundreds or thousands of times more per query. There are no comprehensive statistics on how often or how heavily AI is used3.
Published totals are for data centres, which include more than AI. A US Department of Energy release on a Lawrence Berkeley report says US data centres used 176 terawatt-hours in 2023, about 4.4% of US electricity, and projects 325 to 580 terawatt-hours by 2028. The IEA's 2026 update puts global data-centre electricity at 485 terawatt-hours in 2025, roughly doubling to about 950 terawatt-hours in 2030, around 3% of world electricity, with AI-focused data centres tripling23.
Within AI, the balance of electricity use has already shifted from training to inference, the day-to-day running of trained models3. Water sits beside the power bill. The IEA estimates that data centres use water directly for cooling and indirectly for energy supply and chip manufacturing, varying with cooling technology, local climate and the electricity source. Global data-centre water consumption is about 560 billion litres a year, and about 1,200 billion litres in 2030 in the base case1.
Example
A simple text query now typically uses less electricity than a television running for the same stretch of time, the IEA says. Generating video, running a reasoning model or letting an agent take many steps can use hundreds or thousands of times more energy per query3.
Why it matters
Data-centre electricity is projected to roughly double by 2030, and the IEA says concerns about electricity prices and environmental impacts have already risen around these projects. Energy per task is falling quickly, while heavier tasks and wider use push the total up3.
Questions people ask
Does training AI use more energy than using it?
For the sector, the IEA says the balance has already shifted from training to inference, the day-to-day use of a trained model. A single query is not fixed: video, reasoning and agentic tasks can use hundreds or thousands of times more energy than simple text3.
How much electricity does AI use?
There is no single global total for AI alone. The IEA says comprehensive statistics on AI use do not exist. It does publish data-centre figures: 485 terawatt-hours in 2025, about 950 in 2030. A 2024 US Department of Energy release put US data centres at 176 terawatt-hours in 2023, a total that includes more than AI23.
Can AI become more energy efficient?
Per task, the IEA says it already has: energy use per AI task has fallen by at least an order of magnitude a year. Total data-centre electricity is still projected to roughly double by 2030, because more people use AI and some tasks are far heavier3.
Does AI use water as well as electricity?
The data centres that run it do. The IEA estimates direct water use for cooling and indirect use through energy supply and chip making. It puts global data-centre water consumption at about 560 billion litres a year, and about 1,200 billion litres in 2030 in its base case. Cooling design, climate and the electricity source change the figure1.