How much electricity is consumed by the largest US data centres, and what emissions are allocated to them? A study published on 16 November 2026 provides a detailed estimate for 403 installations. Its central scenario is about 82 terawatt hours over 12 months, or 1.8 per cent of US electricity consumption, and 25.7 million tonnes of CO2. These figures describe a precise perimeter between May 2024 and April 2025, not the global balance sheet of artificial intelligence in 2026.
Data centres: one sample, several scenarios
The Harvard-based team has identified facilities using site-specific information and satellite imagery. It then estimated their electrical activity. The result does not come from an exhaustive collection of the meters at each centre: it depends on load scenarios, i.e. assumptions about their level of operation.
This distinction explains the published range. Based on assumptions, the consumption of 403 plants ranges from 68 to 99 terawatt hours. The central scenario is a reference point between these assumptions, not an exact measure applicable to each building. An installed power, which describes a capacity, must also be distinguished from the energy actually used over a period.
The location counts as much as the volume consumed
Emissions are calculated by assigning the electricity mix of their region to the plants. This method uses annual averages. It does not follow, hour by hour, the physical origin of the current arriving at each server.
The scientific release details this mixture: about 54% of the electricity allocated comes from fossil fuels, 21% from nuclear and 25% from renewables. Virginia concentrates 142 sample centers, ahead of Oregon with 56 and Ohio with 38. With about 21 terawatt hours, it alone accounts for almost a quarter of the estimated consumption.
At comparable volume of calculation, the location and the surrounding electrical system can therefore change the balance sheet assigned. But this accounting does not automatically deduce which additional plant would have operated, or closed, in the absence of a centre. An average attribution and a measure of the effects of a settlement decision do not answer the same question.
Revised figures, not all versions combined
The public deposit of the researchers explains an essential methodological precision: the main calculation uses all the electricity production in its denominator, including nuclear and renewables. The calculation limited to combustion production is retained as a diagnosis, not as a basis for the main results.
In this version, the weighted average carbon intensity of the centres studied reaches about 314 grams of CO2 per kilowatt hour, compared with a national reference of about 348 grams. The higher figures in a June pre-publication should therefore not be mixed with the final November results.
The public files give a central scenario of 81.8 terawatt hours and 25.7 million tonnes of CO2. They are used to verify regional totals and calculation assumptions. On the other hand, the detailed addresses, contact details and commercial characteristics of the facilities remain subject to broadcasting restrictions. The availability of the code does not mean that all original data are public.
What the study does not allow to attribute to AI
These infrastructures have multiple uses: remote computing services, storage and calculation, including artificial intelligence. Harvard notes that the available data do not separate consumption from consumption of generalist uses from that of activities specialized in AI. Attributing the 82 terawatt hours to conversational robots alone would therefore be incorrect.
The sample also does not cover all US installations. And the period ended in April 2025. The researchers estimate that the expansion since then has probably increased needs, but their study does not provide a measured total for November 2026.
Finally, work does not directly quantify health effects. It provides a basis for examining the relationships between computing sites, power grids and emissions, not a count of diseases attributable to the centres. To judge a concrete project, it remains necessary to know its activity, supply and local effects, rather than mechanically applying the national average of this sample.
Illustration photograph: computer code on a screen, May 5, 2016. Markus Spiske / Pexels. It does not represent any of the centres studied.





Comments