Asking an AI chat assistant a question takes the user a few seconds. But behind that answer are huge data centres where thousands of powerful chips run and must be constantly cooled. The spread of AI has pushed the energy needs of these facilities to the top of the global agenda.
What do the figures say?
According to the International Energy Agency’s (IEA) 2025 report “Energy and AI”, data centres consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5 per cent of global electricity use. In the report’s base case this is expected to rise to about 945 terawatt-hours by 2030, more than double. The main driver of the increase is accelerated servers used for AI.
Although the global share looks small, consumption is concentrated in certain places. The US and China account for most of the growth, and in some areas where data centres cluster, local power grids are under serious pressure.
Training and use
Energy is spent in two stages: training a large model requires thousands of chips running for weeks or months. Once the model is released, every request from millions of users (inference) also consumes energy. As use spreads, inference’s share of total consumption grows. Generating images and video usually needs far more computing power than generating text.
Water footprint
Some data centres use evaporative cooling for their servers, which consumes significant amounts of water. Power plants generating the electricity also have an indirect water footprint. New data centre projects in drought-hit regions can therefore become controversial.
The search for solutions
- Efficient hardware: Each new chip generation aims to do the same work with less energy.
- Smaller models: For many tasks, smaller and more efficient models can be enough instead of large ones.
- Clean energy: Tech companies are turning to renewable and nuclear power deals.
- Smart cooling and location: Liquid cooling, reusing waste heat and building in cool climates improve efficiency.
What can I do as a user?
The impact of individual use is small on its own, but not generating large numbers of images or videos unnecessarily, using a search engine for simple questions and asking for a task in one clear prompt save both time and energy.
How is efficiency measured? PUE
Not all the electricity a data centre uses goes to the computers; a significant part is used by support systems such as cooling, power conversion and lighting. The most widely used indicator for measuring this difference is PUE (Power Usage Effectiveness).
- Definition: PUE is the ratio of a facility's total energy consumption to the energy consumed by its IT equipment; it was proposed in 2007 by the industry consortium The Green Grid.
- Ideal value: A value of 1.0 is the theoretical limit at which all energy goes to IT equipment; the higher the value, the larger the share going to support systems.
- Industry average: In the Uptime Institute's global data centre surveys, the average PUE has hovered around 1.5–1.6 in recent years.
PUE is a useful measure but not sufficient on its own: it does not show the source of the electricity used, how water is used or how efficiently the hardware runs.
Local impacts and transparency
Although data centres' share of electricity consumption is still limited on a global scale, their impact is much more pronounced in certain regions. For example, according to Ireland's Central Statistics Office, data centres accounted for about 21 per cent of total metered electricity consumption in the country in 2023.
Such concentrations are leading to local debates about grid capacity, electricity prices and water resources. Under its recast Energy Efficiency Directive, the European Union has begun requiring data centres above a certain size to report their energy and water use. Experts stress that, to understand the true environmental cost of artificial intelligence, it is important for companies to share more detailed and comparable data at model and facility level.
Sources and further reading
- International Energy Agency (IEA), “Energy and AI” report (2025).
- IEA, “Electricity 2025” report.
- Li, P. et al., “Making AI Less ‘Thirsty’”, Communications of the ACM (2025).
- The Green Grid – PUE: A Comprehensive Examination of the Metric
- Uptime Institute – Global Data Center Survey 2024
- CSO Ireland – Data Centres Metered Electricity Consumption 2023
Leave a comment
Your email address will not be published. Required fields are marked *