In short
Data-centre electricity use is growing fast, and AI is the main reason. The International Energy Agency expects it to roughly double between 2024 and 2030, while local grids in some places are already under strain. Energy now limits how fast AI can grow, so efficiency matters for cost and capacity as well as for the climate, and governments are starting to require measurement and reporting.
Key points
- The IEA estimates that data centres used about 415 TWh in 2024, around 1.5% of the world's electricity, and expects around 945–950 TWh by 2030.
- The growth is concentrated in places: data centres used 23% of Ireland's metered electricity in 2025 and 4.4% of US electricity in 2023.
- Energy per AI prompt is falling quickly, but use is growing faster.
- Power and grid connections now limit AI's growth, which makes energy efficiency a commercial priority as well as an environmental one.
- Policy is shifting from voluntary pledges to reporting duties, especially in the EU.
AI's environmental footprint was a niche research topic a few years ago. It is now a question for energy ministers, grid operators and finance directors. This article sets out why, using the most recent public figures.
How fast is AI's energy demand growing?
Fast enough to reshape electricity planning. The International Energy Agency (IEA) estimates that data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's total, and that this will more than double to around 945 TWh by 2030, with AI the most important driver.
The IEA's April 2026 update put data-centre use at 485 TWh in 2025, up 17% in a single year, with use by AI-focused data centres up 50%. It still expects around 950 TWh by 2030, about 3% of global electricity demand.
A terawatt-hour (TWh) is a billion kilowatt-hours. For comparison, the IEA has noted that the 2030 figure is slightly more than Japan's total electricity consumption today.
Why does a few per cent matter?
Because the demand is concentrated in a few places, and arrives quickly. A global average hides very large local effects.
| Where | Data-centre share of electricity | Source |
|---|---|---|
| Ireland, 2025 | 23% of metered electricity | Central Statistics Office |
| United States, 2023 | 4.4% | Lawrence Berkeley National Laboratory |
| United States, 2028 (projected) | 6.7–12% | Lawrence Berkeley National Laboratory |
A large data centre can need as much power as a town, and it needs it on a timescale of a few years, while new power stations and transmission lines take much longer to build. That is why grid connections, not chips, are now often the limiting factor for new AI capacity.
Isn't each prompt getting cheaper?
Yes, but use is growing faster. Providers report large efficiency gains per prompt:
- Google reported in 2025 that its median Gemini text prompt used 0.24 watt-hours of energy and about 0.26 millilitres of water, and that the energy per prompt had fallen 33-fold in twelve months.
- OpenAI's chief executive wrote in 2025 that an average ChatGPT query used about 0.34 watt-hours.
These are self-reported figures, measured in different ways, for text prompts only. They do not cover image or video generation, long reasoning or agents, which can use far more. See measurement is the bottleneck.
Meanwhile, use keeps growing. ChatGPT reached 900 million weekly active users in February 2026, according to figures reported by TechCrunch, and AI is being built into search, office software and phones by default. A small number multiplied by a very large and growing number is still a large number. Efficiency is not sufficiency explains why falling cost per task tends to increase total use.
What is happening to tech companies' emissions?
They are rising, despite large renewable energy purchases. The largest AI providers publish annual environmental reports, and recent ones show the strain:
- Microsoft's 2026 environmental data fact sheet reports total emissions of about 20.3 million tonnes of CO₂ equivalent in its 2025 financial year, up from about 12.9 million tonnes in 2020, the baseline for its climate goals.
- Google's 2026 environmental report shows its emissions continuing to rise as electricity use for data centres grows.
Both companies remain among the world's largest buyers of renewable power. The difficulty is that AI demand is growing faster than clean supply can be added.
Why does energy now matter commercially?
Because power limits growth, so energy efficiency means more AI from the same capacity, and lower costs. When data centres are power-constrained, the useful measure of a system becomes how much work it does per watt.
This changes the case for green AI. Organisations that would not buy something because it is green will buy it because it is cheaper, faster or more reliable. The useful framing is that cost, speed, quality and energy are, in large part, the same requirement. See the guide for business.
How is policy changing?
From voluntary pledges towards measurement and reporting. Some recent steps:
- United Nations. In December 2025 the UN Environment Assembly adopted its first resolution on the environmental sustainability of AI systems, encouraging states and other stakeholders to minimise AI's environmental impacts.
- European Union. Data centres with at least 500 kW of IT power must report energy and sustainability indicators to a common EU scheme under Delegated Regulation (EU) 2024/1364. Under the AI Act, providers of general-purpose AI models must document the known or estimated energy consumption of their models (Annex XI).
- United Kingdom. The evidence annex to the UK Compute Roadmap estimated that emissions from AI compute between 2025 and 2035 could be 34–123 million tonnes of CO₂, around 0.9–3.4% of projected UK emissions over the period, as summarised by GreenCode.
At the same time, many governments see AI as a matter of competitiveness and security and are expanding capacity rather than restraining it. See sovereign AI and sustainability and green AI policy.
Why act now rather than later?
Because the infrastructure being built now will run for decades, and habits set now are hard to change. Data centres, power plants and software platforms all have long lives. Choices about where to build, which models to deploy by default, and what to measure are being made now.
A 2026 UK green AI sandpit of mathematicians, computer scientists and industry framed the question as how the mathematical sciences could reduce AI's energy and resource needs as demand grows across society and industry. Its emphasis on impact within months or a few years reflects the same urgency.
What to do next
- Find out how much AI your organisation uses, and what it costs: see measuring AI energy.
- Choose providers and regions with clean power and published figures: see hosting and data centres.
- Use the smallest model that does the job: see choosing the right model.
- Build energy and reporting into contracts: see sustainable AI procurement.
- Read what is green AI? for the key definitions.
Frequently asked questions
How much electricity does AI use?
There is no exact figure for AI alone. The IEA estimates that all data centres used about 415 TWh in 2024, around 1.5% of global electricity, with AI the fastest-growing part and the main driver of expected growth to around 945 TWh by 2030.
How much energy does one AI prompt use?
Providers' own figures put a typical text prompt at around a third of a watt-hour: Google reported 0.24 Wh for a median Gemini text prompt and OpenAI 0.34 Wh for an average ChatGPT query. These are self-reported, use different methods and exclude image, video and agent workloads.
Will clean energy solve AI's footprint?
It helps, but it is not enough on its own. Clean power takes time to build, demand for it is high across the economy, and AI's growth can delay the retirement of fossil plants or claim renewable supply that would otherwise displace them.
Sources
- Energy and AI, International Energy Agency (2025)
- Key questions on energy and AI, International Energy Agency (2026)
- 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (2024)
- Data centres metered electricity consumption 2025: key findings, Central Statistics Office, Ireland (2026)
- Measuring the environmental impact of delivering AI at Google Scale, Elsworth et al., Google (2025)
- The gentle singularity, Sam Altman (OpenAI) (2025)
- ChatGPT reaches 900M weekly active users, TechCrunch (2026)
- Google 2026 environmental report, Google (2026)
- 2026 environmental data fact sheet, Microsoft (2026)
- Resolution 7/9: Environmental sustainability of artificial intelligence systems, United Nations Environment Assembly (2025)
- Commission Delegated Regulation (EU) 2024/1364 on the first phase of the establishment of a common Union rating scheme for data centres, European Commission (2024)
- Regulation (EU) 2024/1689 (AI Act), Annex XI, European Union (2024)
- UK Compute Roadmap: evidence annex, Department for Science, Innovation and Technology (2025)
- Green AI is a mathematics problem as much as a hardware one, GreenCode (2026)
Related
What is green AI?
A plain-language definition of green AI: AI for green and greener AI, efficiency and sufficiency, and its footprint in energy, carbon, water and materials.
Efficiency is not sufficiency: the rebound effect in AI
Why making AI more efficient does not guarantee it uses less energy overall, what the rebound or Jevons effect is, and what helps keep savings real.
Green AI hosting and data centres
How to choose where AI runs: grid carbon intensity, PUE and WUE, renewable energy claims, water use and siting, with questions to ask hosts.
How to write a green AI policy
A step-by-step guide to writing an organisation's green AI policy, with a template outline covering use, models, measurement, buying, claims and review.