Artificial intelligence affects the environment in two directions at once: running it consumes electricity, and applying it can reduce waste or improve warnings elsewhere. The International Energy Agency (IEA) estimates that data centers used about 415 terawatt-hours (TWh) in 2024, around 1.5 percent of world electricity, and that the total could reach about 945 TWh by 2030. The same agency estimates that wide use of existing AI-led solutions could cut energy-related emissions by around five percent in 2035, but only if barriers are overcome and rebound effects stay small.
Key takeaways
- Data centers used roughly 1.5 percent of global electricity in 2024, and the IEA base case roughly doubles that by 2030.
- In the United States, a federal laboratory estimated data centers at 4.4 percent of electricity in 2023, with a 2028 range of 6.7 to 12 percent.
- Canada's energy regulator models anywhere from 1.5 gigawatts to 12 gigawatts of data center load, depending on the scenario.
- Benefits are real but narrower than headlines suggest: a peer-reviewed flood model covers over 80 countries, while the IEA calls AI "not a silver bullet" for emissions.
The electricity side of the ledger
The IEA's report Energy and AI puts data center use at around 415 TWh in 2024. The United States accounted for 45 percent of that, China for 25 percent, and Europe for 15 percent. Use has grown about 12 percent a year since 2017. In the base case, consumption rises to roughly 945 TWh by 2030, and servers built for AI workloads account for almost half of the net increase.
The agency stresses uncertainty. It models a "Lift-Off" case above 1,700 TWh in 2035, a "High Efficiency" case near 970 TWh, and a "Headwinds" case that levels off near 700 TWh. Data center emissions in the base case rise from about 180 million tonnes today to about 300 million tonnes by 2035, still under 1.5 percent of energy-sector emissions.
North American figures
Estimates for North America come from different institutions and different assumptions, so the table keeps them separate. Gigawatts measure capacity, while terawatt-hours measure energy over a year, so the rows are not directly comparable.
| Source | Scope | Figure |
|---|---|---|
| Lawrence Berkeley National Laboratory (December 2024) | U.S. data centers, 2023 | 176 TWh, about 4.4 percent of U.S. electricity |
| Lawrence Berkeley National Laboratory | U.S. data centers, 2028 | 325 to 580 TWh, 6.7 to 12 percent |
| Canada Energy Regulator, Current Measures scenario | Canada, added data center load | 1.5 GW by 2030 and 3.5 GW by 2050 |
| Canada Energy Regulator, Higher scenario | Canada, data center load | 12 GW by 2050, about one third of demand growth |
The regulator's 2026 outlook has end-use electricity in Canada rising 44 percent from 2023 to 2050 in the Current Measures scenario and 73 percent in the Higher scenario. Those are modeled scenarios, not forecasts, and the fact sheets carry a modification date of March 17, 2026.
What the benefits look like when measured
Much writing on AI and sustainability lists possible uses without evidence. A narrower claim is easier to check. A 2024 paper in the journal Nature, Global prediction of extreme floods in ungauged watersheds, reports that an AI forecasting model reached reliability at up to a five-day lead time that is similar to or better than a leading global system's same-day nowcasts. The model feeds a public early warning service in over 80 countries. Flood warnings help communities adapt rather than cut emissions, and the authors note that better hydrological data would improve results further.
On emissions, the IEA estimate of around five percent by 2035 assumes broad adoption of solutions that already exist, in buildings, transport, and industry. The agency warns that rebound effects, such as a shift from public transport to autonomous cars, could undercut some of the savings.
Questions that separate evidence from hype
- Does the claim name a measured result, a modeled scenario, or only a possibility?
- Does it compare against a realistic baseline, such as an existing forecasting method?
- Does it count the electricity used to train and run the model?
- Would the benefit require adoption that has not happened yet?
- Who produced the number: an independent agency, a university, or the company selling the tool?
FAQ
How much electricity do data centers use?
The IEA estimates about 415 TWh in 2024, around 1.5 percent of global electricity. In the United States, Lawrence Berkeley National Laboratory estimated 4.4 percent for 2023.
Will AI cut emissions overall?
It might, but the evidence is conditional. The IEA estimates around five percent of energy-related emissions in 2035 from widely applied existing solutions, with barriers and rebound effects as stated risks.
How much power could data centers need in Canada?
The Canada Energy Regulator's 2026 scenarios range from 1.5 GW added by 2030 and 3.5 GW by 2050 to 12 GW by 2050. The gap reflects different assumptions about growth and investment.
Is every AI model equally power-hungry?
No. Energy use depends on model size, hardware, and how often the model runs. The sources cited here report totals for data centers, not per-model figures, so no per-query number is given.
Related reading
- AI technology aimed at environmental protection
- AI and farming productivity
- How AI affects modern society
- AI and robotics in society
More in AI and Society.
Last verified: October 2026
Sources
- Energy and AI, International Energy Agency, 2025
- 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, December 2024
- DOE releases new report evaluating increase in electricity demand from data centers, U.S. Department of Energy, December 20, 2024
- Canada's Energy Futures 2026 fact sheets (Current Measures and Higher scenarios), Canada Energy Regulator, 2026
- Nearing et al., Global prediction of extreme floods in ungauged watersheds, Nature, March 20, 2024