As AI and cloud services spread, data center electricity consumption is rising rapidly. A study published in the journal Nature Sustainability on November 16, 2026 estimated that 403 hyperscale data centers operating in the United States have annual electricity demand of about 82 TWh and attributed carbon dioxide (CO₂) emissions of about 25.7 million tons. That is equivalent to roughly 1.8% of total US electricity consumption. However, these are not measured values from each facility's power meter. They were calculated by combining each facility's electrical capacity, an assumed utilization rate, and the power generation mix of its region. The research team built a method for estimating electricity demand and CO₂ emissions at the facility level, and used it to analyze how the environmental impact of data centers varies with location and operating conditions.

AD

403 US hyperscale data centers: about 82 TWh of annual electricity demand and 25.7 million tons of CO₂

The peer-reviewed paper (Nature Sustainability) by Gianluca Guidi and colleagues estimates electricity demand and CO₂ emissions for individual US hyperscale data centers (HDCs). The paper was submitted on June 23, 2025, accepted on October 13, 2026, and published on November 16. The analysis covers 403 facilities that were operating between May 2024 and April 2025.

In the team's central scenario, the 403 facilities had combined annual electricity demand of about 82 TWh. That is equivalent to about 1.8% of total US electricity consumption, and the CO₂ emissions attributed to that demand came to about 25.7 million tons (25.7 Mt). The study covers CO₂ emissions only, not CO₂-equivalent (CO₂e) emissions that include other greenhouse gases.

The team also analyzed several scenarios with different facility load factors. Annual electricity demand ranged from about 68 to 99 TWh, and attributed CO₂ emissions from about 21 to 31 million tons (21–31 Mt). The central figures of 82 TWh and 25.7 million tons are the results when a specific condition is set for average facility load.

Scenario Facility load factor (u) Annual electricity demand (TWh/year) Attributed CO₂ emissions (million tons)
Low-load scenario Value not confirmed in the summarized portion of the text About 68 About 21
Central scenario 0.58 About 82 About 25.7
High-load scenario Value not confirmed in the summarized portion of the text About 99 About 31

Note: Model estimates for 403 US HDCs operating from May 2024 to April 2025. The specific load factor values for the low- and high-load scenarios could not be confirmed in the summarized portion of the paper's text.

In this estimation, the load factor "u" plays a key role. It represents what share of a facility's electrical capacity is consumed on average over the year. Electricity demand and attributed CO₂ emissions vary with this setting. The approach makes it possible to evaluate facilities whose actual power consumption is not public, but the results directly reflect the load-factor assumption.

An estimation method combining satellite imagery and power grid data

The research team first identified 675 candidate facilities in the US. They screened them using operator records, address geocoding, OpenStreetMap building footprint data, visual inspection, and satellite imagery, and ultimately settled on 403 facilities for analysis.

Next, they determined the electrical capacity, meaning how much power each facility can use. For 397 of the 403 facilities, capacity could be identified from data provider records and similar sources. For the remaining six, missing values were estimated with a gradient-boosted regression tree model that used building area, climate zone, and balancing authority (the entity responsible for balancing supply and demand on the power grid) as explanatory variables.

Annual electricity demand was calculated by multiplying a facility's electrical capacity by 8,760 hours per year and by the load factor.

\[ E=P_{\mathrm{capacity}}\times8760\times u \]

Here, is annual electricity demand, P_} is the facility's electrical capacity, and is the facility load factor. The electrical capacity includes not only IT equipment such as servers but also losses from cooling systems and power conversion. The value u=0.58 used in the central scenario assumes that 58% of a facility's electrical capacity is used on average over the year; it is not an actual load factor measured at individual facilities.

CO₂ emissions were calculated using the US Environmental Protection Agency's (EPA) "eGRID2023 Revision 2." For the jurisdiction of the balancing authority to which each facility belongs, the team aggregated plant-level generation and CO₂ emissions, then allocated emissions to data center electricity consumption based on that region's generation mix.

This method is a location-based accounting approach, using a total-output-based carbon intensity with the total generation of power plants as the denominator. It does not measure CO₂ emitted directly by data centers, nor does it calculate how much the construction or operation of a facility increased emissions across the grid as a whole. It is an attributional method that allocates CO₂ emissions generated on regional grids according to each facility's estimated electricity demand.

To ensure reproducibility, the capacity estimation model, calculation procedures, and state-level and balancing-authority-level aggregated data have been published on GitHub. However, identifying information for individual facilities, exact addresses and coordinates, and attributes involving commercially sensitive information are restricted under data use agreements. External researchers can verify the published aggregate results and calculation procedures, but cannot freely access all of the original facility-level data.

AD

About 54% from fossil fuels, and over half of demand concentrated in just four states

Analyzing the regional power mix, the team found that about 54% of the electricity generation attributed to the 403 facilities came from fossil fuels. Natural gas accounted for about 38% and coal for about 15%, showing that fossil fuels still make up a large share of the generation supporting data center demand.

The total-output-based carbon intensity, weighted by the electricity demand of the 403 facilities, was about 314 gCO₂/kWh. This is below the roughly 348 gCO₂/kWh US average that the team presented for comparison. On average, the facilities analyzed are located on grids with lower carbon intensity than the US as a whole.

However, even when carbon intensity is below the national average, large demand increases total CO₂ emissions. Because the 403 facilities' annual electricity demand reaches about 82 TWh, attributed CO₂ emissions come to about 25.7 million tons even though emissions per kWh are lower than the national average.

The study also revealed a geographic concentration of demand. More than half of the 403 facilities' electricity demand is concentrated in four states: Virginia, Oregon, Ohio, and Iowa. In regions where hyperscale data centers are clustering, the local generation mix, along with rising demand, strongly influences the CO₂ emissions attributed to data centers.

Not just AI: how this differs from earlier estimates, and the limits of the method

An important point in interpreting the study is that the 403 facilities do not cover all hyperscale data centers in the US. Some facilities were not included, and the demand of the facilities analyzed varies with the assumed load factor. Facility coverage and load-factor estimation carry different uncertainties, and the team does not position the 403-facility total as a strict lower bound on US electricity demand.

Facility-level data also cannot distinguish general cloud processing from AI-specific processing. The roughly 82 TWh of demand and 25.7 million tons of CO₂ are therefore estimates for the hyperscale data centers analyzed as a whole, not figures for AI-dedicated equipment alone. The study did not directly assess health impacts either; its analysis is limited to electricity demand, attributed generation and fuel mix, carbon intensity, and CO₂ emissions.

There is also a time lag in the data used. The period for which facility operations were examined is May 2024 to April 2025, but the power mix and CO₂ emissions were calculated using eGRID2023, which reflects 2023 power plant data. The facility analysis period and the plant data period therefore do not match.

Comparisons with existing industry studies are also complicated by differences in scope. The 2024 United States Data Center Energy Usage Report published by Lawrence Berkeley National Laboratory (LBNL) in December 2024 used a bottom-up approach based on installed equipment and power consumption. It estimated electricity consumption of all US data centers at about 76 TWh in 2018 and about 176 TWh in 2023, equivalent to roughly 1.9% and 4.4% of total US electricity consumption, respectively.

The LBNL report covers the entire US data center industry, including smaller facilities, whereas this study is limited to 403 confirmed hyperscale data centers. Because the facilities analyzed, the data used, and the calculation methods all differ, the two sets of figures cannot be simply compared to judge increases or decreases in electricity demand.

Gianluca Guidi, the paper's lead author and a postdoctoral researcher in the Department of Biostatistics at the Harvard T.H. Chan School of Public Health, said that the scale that can already be confirmed is striking, but it is only a snapshot of a rapidly expanding industry. The study demonstrates a way to estimate electricity demand and CO₂ emissions from facility electrical capacity and regional power mix for an industry with limited public information. If measured facility-level power data and up-to-date power plant data become available, estimates could become more accurate, and the effect of AI-oriented facility expansion on electricity demand and CO₂ emissions could be analyzed in more detail.