Data centres used about 1.5% of global electricity in 2024, and AI only a fraction of that. The load is growing fast and set to roughly double by 2030 — yet even then it would add less new demand than cooling or electric cars. The same pattern holds for the carbon and water that come with it: individually tiny per query, modest in the global total, but concentrated enough to bite locally.
In 2024, the world’s data centres consumed roughly 415 terawatt-hours of electricity — about 1.5% of global supply.1 Servers running AI accounted for an estimated 15% of that, so AI itself was on the order of a fifth of a percent of world electricity.1 The load is also concentrated: the United States, China and Europe together made up about 85% of it, with the US alone at 45% (some 180 TWh).1
It is rising quickly. Data-centre electricity use has grown around 12% a year since 2017, several times faster than total electricity demand,1 and the IEA’s central projection has it more than doubling to about 945 TWh by 2030 and near 1,200 TWh by 2035.1 Those later figures are forecasts, not measurements — and the spread across the IEA’s scenarios is wide.
The most useful way to size the “cost to the grid” is not the data-centre total but how much new demand it adds against other growing uses. On the IEA’s central case, data centres account for under 10% of the growth in global electricity demand between 2024 and 2030.1 Their roughly 530 TWh of added demand is less than the growth expected from electric cars or from cooling, and far below industry.7
None of this means the build-out is costless. Because data centres cluster — nearly half of US capacity sits in five regional hubs1 — they can dominate a local grid even while barely moving the global figure. In the US, they are set to drive close to half of all electricity-demand growth to 2030,1 against a backdrop where total demand is climbing again after years of stagnation: the IEA expects world consumption to rise from about 28,200 TWh in 2025 to 33,600 TWh by 2030.2
Generating the electricity for data centres produced roughly 180 million tonnes of CO₂ in 2024 — about 0.5% of energy-related emissions.37 On the IEA’s projections that share climbs to around 1% by 2030 in the central case, or about 1.4% under faster growth.7 Modest in absolute terms — but notable because data centres are one of the few sectors where emissions are still expected to grow, even as power grids decarbonise.7
The pressure is concentrated in a handful of firms. Google’s total greenhouse-gas emissions rose by about half between 2019 and 2024 as AI and data-centre demand grew;4 Google and Meta each now report several million tonnes of scope 1 and 2 emissions from their operations.6 Company reporting is also contested: a Guardian analysis argued that big-tech data-centre emissions in 2020–2022 were several times higher than officially reported, largely a function of how clean-energy purchases are counted.10
Data centres also use water, mostly for cooling. Globally the IEA puts data-centre water consumption at about 560 billion litres in 2023.3 At the level of a single request, Google’s own production measurement finds a median Gemini text prompt uses about 0.24 watt-hours of electricity, emits 0.03 g of CO₂, and consumes 0.26 millilitres of water — roughly five drops, and less energy than nine seconds of television.5
That number is unusually well-measured, but it counts only on-site cooling water, not the water used to generate the electricity itself; nor does it include model training, networks or your device.5 Draw the boundary differently and the per-query figure moves by orders of magnitude — earlier public estimates ran to 45–50 mL a prompt.5 As with electricity, the individual number is tiny; the aggregate and the location are what matter. Two-thirds of US data centres built or under construction since 2022 sit in areas of high water stress, and US data-centre water use is projected to reach 150–280 billion litres by 2028.6
| Per… | Electricity | CO₂ | Water |
|---|---|---|---|
| One median text prompt | 0.24 Wh | 0.03 g | 0.26 mL |
| One million prompts | 240 kWh | 30 kg | 260 L |
| All data centres, 2024 | 415 TWh | ~180 Mt | ~560 bn L |
Today’s figures are themselves uncertain: definitions differ, and the IEA’s own baseline has shifted between reports as boundaries (for example, whether crypto mining is counted) changed. AI’s slice is disputed — the IEA’s ~15% share for 2024 is judged an underestimate by some researchers.8 Per-query footprints depend heavily on what is counted: cooling water versus power-generation water, market-based versus location-based carbon, and whether training and networks are included.9 The biggest unknown is efficiency: whether gains in chips, software and cooling offset rising use is the difference between the IEA’s ~700 TWh and ~1,700 TWh outcomes for 2035.1
Credible sources genuinely disagree — not about values, but about the measurement.
Measured in production, a median text prompt uses about five drops of water and less energy than nine seconds of TV, and Google reports the per-prompt energy and carbon of that query fell 33-fold and 44-fold in a single year as models and hardware improved.5
Both can be true at once: a single query is genuinely small and getting smaller, while the total across billions of queries and non-AI workloads is large and concentrated. The unresolved part is the accounting boundary — and, until disclosure is standardised, reasonable estimates will differ.
Two threads left out to hold length.