Data centres worldwide consumed 485 TWh in 2025, a little over 1.5% of global electricity, according to the International Energy Agency. The 2030 projection is 945 TWh, around 3% of world demand.
Except that it leaves out cryptocurrency mining, and does not separate AI from the rest of computing. And behind the global average, Ireland sits at 23% of its metered electricity.
The figure, and exactly what it covers
The reference report changed in April 2026, and the two publications circulate together without giving the same values for the same year.
| Quantity | Value | Source, date |
|---|---|---|
| Data centres, 2024 | 415 TWh, 1.5% of global electricity | IEA, Energy and AI, April 2025 |
| Data centres, 2024 (revised) | 416 TWh | IEA, Key Questions on Energy and AI, 16 April 2026 |
| Data centres, 2025 | 485 TWh | same, Table A.1 |
| 2030 projection | 945 TWh, around 3% | same, Base Case |
| Installed capacity 2025 | 114 GW, of which 82 GW IT | same |
| Global average PUE 2025 | 1.38 | same, Tables A.1 and A.3 |
2025 split: United States 224 TWh, China 117, Europe 72.
Total energy consumed by the data centre divided by the energy that actually reaches the servers. A PUE of 1.38 means that for 1 kWh of compute, you spend 0.38 kWh on cooling, UPS and losses.
For a carbon assessment, it is the multiplier to apply to any IT consumption. PUE says nothing about how efficient the servers themselves are: a PUE of 1.1 on machines running at 10% load is still a bad data centre.
Why 485 and 790 TWh describe the same year
A second figure circulates for 2025: around 790 TWh, or 2.5% of global generation, attributed to the Energy Institute.
The 305 TWh gap is not an error. The IEA excludes cryptocurrency mining, estimated at 150 to 200 TWh, which accounts for roughly two thirds of the difference. The rest comes from different boundary choices.
Question
A client quotes you "790 TWh for data centres in 2025" while your notes say 485 TWh. What do you do?
Choisissez une réponse pour voir l'explication.
The IEA does not say what AI consumes
The most widespread misreading on the subject. The 2025 report does not separate AI from the rest of data centres. Any sentence of the form "according to the IEA, AI consumed X TWh in 2024" is false.
The 2026 report introduces an "AI-focused data centres" category and publishes a projection: around 465 TWh in 2030, more than triple 2025, with growth of 50% in 2025. But the 2025 value itself is written nowhere in the text, it only appears on a figure. Our World in Data reads it at around 155 TWh, or roughly 0.5% of global electricity, and that is a reading off a chart, not a published number.
The growth, and the nuance that puts it in its place
Data centres grew 17% in 2025 while global electricity demand grew 3%. Growth ratio: around 5.7.
And yet they account for only 9% of global electricity growth over the year: the extra 70 TWh sit against a total rise of around 800 TWh. Air conditioning, electric vehicles and industry weigh more in absolute terms.
An honest framing gives both numbers. Taken alone, the first alarms and the second reassures.
One detail shows how fragile the data is: the same IEA publication announces that growth three different ways, "17%" in its press release, "more than 15%" in chapter 1, and 16.6% if you divide its own tables. Quote whichever you like, but say which.
What the global average hides
The IEA raises the real issue itself: data centres were 1.5% of global electricity in 2025, but 20 to 30% of demand in the states, metro areas and regions where the clusters sit.
Ireland provides the cleanest series, published by its national statistics office.
| Year | GWh | Share of metered consumption |
|---|---|---|
| 2015 | 1,240 | 5% |
| 2019 | 2,490 | 9% |
| 2021 | 4,012 | 14% |
| 2023 | 6,339 | 21% |
| 2024 | 6,973 | 22% |
| 2025 | 7,663 | 23% |
2025 growth: +10%, against +2% for every other user, residential included. And the number that lands: all of Irish residential is 28%, data centres 23%.
In the Netherlands, the share is 4.6% in 2024. The gap with Ireland is not only physical: the Dutch CBS excludes server rooms in universities, hospitals and companies, where the Irish CSO excludes nothing that is declared. Comparing 4.6% with 23% without saying so is a mistake.
Contracted is not consumed
A profitable trap to know about: it defuses half the alarmist projections without denying the problem.
| Territory | Queued or contracted | Measured reality |
|---|---|---|
| Dominion, Virginia | 47 GW contracted in July 2025 | 4 GW actual peak in 2025 |
| ERCOT, Texas | over 230 GW in January 2026 | state peak record: 85 GW |
| Denmark | around 60 GW queued | national peak around 7 GW |
| Ireland | around 2,400 MVA contracted | 959 MVA of real demand in 2024 |
Dominion writes it in black and white in its 6 January 2026 letter to grid operator PJM: "The Company is forecasting 16.6 GW of demand by 2046, not the 47 GW of capacity. This difference highlights the difference between capacity and demand." And the operator classes 30.1 GW of its 47 GW contracted as "not firm load".
Announced projects show the same gap. Stargate advertises $500 billion and more than 5.5 GW, for 0.3 GW actually operational, or 3% of what is announced for 2029. Meta Hyperion announces 5 GW and has nothing in service.
An announced project is a communications intention. Contracted capacity is a grid connection reservation, often non-firm, and free or nearly free to file. Built capacity is steel and silicon in service.
All three travel under the same word, "gigawatt". The observed ratio between announced and built runs from 3% to 0% on the world's two largest projects. Always ask which of the three you are being quoted.
The emissions, and what the figure leaves out
The IEA puts data centre emissions at around 180 Mt CO2 in 2024, or 0.47% of the 38.4 Gt of global energy emissions. Revised trajectory: around 350 Mt in 2035, roughly 2% of power sector emissions.
That figure covers one component, not a footprint: the indirect emissions from electricity generation, the equivalent of a scope 2. Checked against the full text of the 2026 report: it contains no occurrence of "embodied", "life cycle", "scope 2" or "scope 3". Hardware manufacturing, building construction, networks and end-user devices are out of scope.
For scale, commercial aviation emitted 942 Mt CO2 in 2024. Data centres are around 19% of that. Here too the boundaries do not overlap: aviation is measured as direct combustion, data centres as indirect electricity with no manufacturing. The ratio gives an order of magnitude, nothing more.
What to take away
Data centres are around 1.5% of global electricity and 9% of its annual growth. Nobody publishes AI's share of that properly. Behind the average, some territories are at 23%. And between what is announced and what runs, there is an order of magnitude.
The next chapter drops down to the level of the request: what a per-request footprint figure really covers, and what it leaves outside.
FAQ
How much electricity do data centres use worldwide?
485 TWh in 2025, a little over 1.5% of global electricity, according to the International Energy Agency in its Key Questions on Energy and AI report of 16 April 2026. The 2030 projection is 945 TWh, around 3% of world demand. That scope excludes cryptocurrency mining.
How much of that consumption comes from artificial intelligence?
The IEA publishes no written value for 2025. Its 2025 report did not separate AI from the rest of data centres, and the 2026 one introduces an "AI-focused data centres" category with a projection of around 465 TWh in 2030, but with no 2025 figure written in the text. A chart reading by Our World in Data puts it at around 155 TWh, roughly 0.5% of global electricity.
Why does one source say 485 TWh and another 790 TWh?
Because the IEA excludes cryptocurrency mining, estimated at 150 to 200 TWh, while the Energy Institute includes it. That single boundary choice explains two thirds of the 305 TWh gap. Both figures are correct within their own framing.
What share of a country's electricity do data centres take?
It varies by an order of magnitude. In Ireland, data centres were 23% of metered electricity consumption in 2025, more than urban residential use. In the Netherlands the share is 4.6% in 2024, on a narrower scope that excludes server rooms in companies and hospitals. The IEA records 20 to 30% in regions where data centre clusters concentrate.
Are the gigawatts announced by big projects already being consumed?
No, and the gap is enormous. Stargate announces more than 5.5 GW for 0.3 GW operational, or 3%. Dominion Energy reports 47 GW of contracted capacity in Virginia against 4 GW of actual peak in 2025, and itself classes 30.1 GW of that capacity as non-firm load. Announced, contracted and built are three different quantities.