The International Energy Agency's report Energy and AI, released on 10 April 2025, is the agency's most detailed attempt so far to quantify how the growth of artificial intelligence affects electricity demand and emissions. On the emissions side, its headline numbers are smaller than some public discussion has suggested, but the trend runs in the opposite direction to most other sectors.
According to the IEA, data centres used about 415 terawatt-hours of electricity in 2024, about 1.5% of global electricity consumption. The indirect carbon dioxide emissions associated with that electricity, which depend on the generation mix of the grids supplying the data centres, were about 180 million tonnes. That compares with global fuel combustion emissions of about 35,000 million tonnes in 2024, putting data centres at about 0.5% of the total. The figure excludes emissions from backup generators, and it covers all data centre workloads, of which AI is a subset.
How emissions grow in the IEA's cases
The IEA models several scenarios. In its Base Case, indirect emissions from data centres grow by almost 80% over the rest of the decade, reaching about 1% of global combustion emissions, and rise to about 300 million tonnes by 2035. In its Lift-Off Case, which assumes faster AI adoption and data centre construction, emissions grow two and a half times by 2030, reaching 1.4% of combustion emissions, and rise to about 500 million tonnes by 2035.
The IEA's Base Case projects data centre electricity demand reaching about 1,200 TWh by 2035. Emissions grow more slowly than electricity demand because the IEA expects a rising share of data centre power to come from renewables and nuclear, and because grid emissions intensity is falling in most major markets.
The agency stresses that data centres are among the few sectors, along with road transport and aviation, whose combined direct and indirect emissions rise through 2030 in its projections. In the Lift-Off Case, data centres show the largest emissions growth of any sector.
Where the growth is concentrated
The report finds that data centre growth is concentrated in a small number of countries, led by the United States and China. That concentration matters for emissions because the carbon intensity of electricity differs widely between grids. Data centres in regions with high shares of coal generation produce far more indirect emissions per unit of electricity than those in regions dominated by hydro, nuclear or renewables.
Within countries, data centres also cluster in particular regions, such as Northern Virginia in the United States, and their local grid effects can be much larger than national averages suggest. The IEA notes that in some regions data centres already account for a large share of electricity demand.
The offsetting effect of AI applications
The IEA's report also examines how AI might reduce emissions elsewhere in the energy system. It lists examples including faster detection of methane leaks in oil and gas operations using satellite data, improved efficiency at power plants, optimised industrial processes such as cement production, more efficient vehicle routing, and better control of heating, ventilation and air conditioning in buildings. The IEA gives illustrative efficiency gains: more than 2% for an improved fuel mix in cement production, 5% to 10% for better route choice or driving, and around 10% for optimised building HVAC control.
The agency estimates that adoption of existing AI applications in end-use sectors could reduce emissions by about 1,400 million tonnes of CO2 in 2035 in what it calls a Widespread Adoption Case. That would be about three times the total data centre emissions in the Lift-Off Case and about four times those in the Base Case.
The IEA is explicit that this outcome is not assured. It says there is currently no momentum that could ensure widespread adoption of these applications, and that their aggregate effect could be marginal without the right conditions. Barriers it lists include limited access to data, lack of digital infrastructure and skills, regulatory and security restrictions, and social obstacles. Rebound effects, such as a shift away from public transport toward autonomous cars, could offset some gains.
How the numbers compare
The 180 million tonne estimate for data centre emissions is comparable to the annual energy-related emissions of a mid-sized industrial economy. It is small relative to emissions from steel, cement or road transport, each of which is measured in billions of tonnes.
The IEA's figures also differ from emissions reported by technology companies themselves. Companies typically report market-based Scope 2 emissions, which account for renewable energy purchases and certificates, and these can be considerably lower than location-based emissions calculated from grid averages. The IEA's estimates are based on the physical generation mix of the grids supplying data centres, which is closer to a location-based approach.
Implications for power markets
For electricity markets, the emissions question is linked to the question of which generation will serve new data centre load. In the United States, utilities in several regions have proposed new gas-fired plants to meet data centre demand, alongside renewable and storage procurement and agreements to restart or extend nuclear plants. In China, data centre growth adds to demand on grids with high shares of coal, though China is also adding large amounts of wind and solar capacity.
The IEA's report suggests that the emissions trajectory of data centres will depend less on the data centres themselves than on how fast grids decarbonise. That puts the focus on transmission investment, interconnection queues and the pace at which low-carbon generation can be connected.
Backup generation and local air quality
The IEA's 180 million tonne figure excludes backup generators. Most large data centres keep diesel generators on site to cover grid outages, and they are tested regularly. Their annual emissions are small relative to grid electricity use, but they can matter for local air permits. Some operators are trialling batteries, gas engines or fuel cells as alternatives, and some grid operators are exploring whether backup generators could be called on during grid emergencies, which raises its own permitting questions.
What to watch
Technology companies will publish their annual sustainability reports over the coming months, giving updated figures on electricity use and reported emissions. Grid operators' load forecasts and utilities' integrated resource plans will show how much new generation is being proposed to serve data centres, and what type. Those plans, more than the global averages in the IEA report, will determine where the emissions from AI end up.