Yuang Chen and colleagues at the Massachusetts Institute of Technology (MIT) published a paper on November 19, 2026, in the journal Environmental Research Letters estimating the health impacts of air pollution tied to data center electricity demand. The research team estimates that in 2023, air pollution from power generation serving data centers in the contiguous United States caused about 650 premature deaths. Electricity demand was concentrated in Virginia and Texas, but the health burden per capita was higher in parts of the Midwest. The reason is that the places where electricity is consumed, where power is generated to meet that demand, and where residents are exposed to the resulting pollution are not the same.
Which Power Plants Ramp Up When Data Centers Need More Electricity?
Looking at annual total generation by source shows which sources a grid typically relies on. But when new demand appears, which generators raise their output depends on supply and demand conditions at that hour and on available transmission capacity. If power can be supplied from other regions, generation may increase outside the state where demand grew, and emissions of nitrogen oxides (NOx) and sulfur dioxide (SO₂) may rise there.
According to the authors' abstract of the paper, the team compared hourly output from each generator with and without data center electricity demand. For data center demand, they used state-level estimates from the Electric Power Research Institute (EPRI), allocating them to grid regions that do not necessarily coincide with state borders. The data center load in each region was assumed to be constant every hour throughout the year. Using the power model US-EGO, they then calculated the generation needed to meet demand in each time period, taking transmission between regions into account.
The model also incorporated constraints for coal-fired and some gas-fired plants, such as the minimum output at which they can keep running and how much their output can change by the next time period. Even when the annual generation mix is the same, these operating conditions can change which generators respond to new demand. The team derived the additional emissions from each generator based on the difference in generation with and without data center demand.
Secondary Formation of Particulate Matter Downwind of Power Plants
Gases emitted from power plants undergo chemical reactions in the atmosphere as the wind carries them. The research team used the atmospheric chemistry model GEOS-Chem to calculate how changes in emissions affect pollutant concentrations near the surface. In the Midwest, ammonia from sources such as fertilizer is abundant, and nitrate formed in the atmosphere accounted for much of the increase in fine particulate matter (PM2.5), according to the study. The authors believe that increased output from coal-fired plants with high NOx and SO₂ emission rates in Nebraska and Iowa also contributed to higher concentrations. According to the authors' public preprint, annual average PM2.5 concentrations rose by 0.02 to 0.08 µg/m³ across a wide area from the central to the eastern United States, with local increases of about 0.2 µg/m³.
Ground-level ozone is also produced through chemical reactions involving nitrogen oxides, but the estimated concentration changes showed seasonal differences. According to the abstract of the final paper, concentrations tended to fall in winter and rise in summer, depending on the region. Ozone exposure was assessed over the long term using the maximum daily 8-hour moving average concentration as the indicator.
The team applied the resulting changes in pollutant concentrations to relationships between concentration and mortality risk drawn from epidemiological studies, and estimated premature deaths while accounting for population distribution. Of the roughly 650 deaths, 513 were attributed to PM2.5 and 137 to ozone. The 95% confidence intervals were 469–557 and 90–183, respectively, and the interval for the total shown in the preprint is 559–740. These confidence intervals reflect uncertainty in the relationship between pollutant concentrations and mortality risk; the emissions and atmospheric models carry their own separate uncertainties.
The estimate of about 650 deaths corresponds to 5.3% of the roughly 12,247 premature deaths from air pollution across the entire power sector, calculated in the same study. Both figures are population-level estimates that link power sector emissions to residents' exposure through models.
Data Center Electricity Demand and Health Burden Are Distributed Differently
Comparing data center electricity demand with the health burden from air pollution by state reveals major differences in distribution. Supplementary Table S1 in the authors' public version shows each state's 2023 data center electricity demand, along with estimates of premature deaths among each state's residents caused by data center demand across the contiguous United States. According to the preprint, Virginia's electricity demand was 33.9 TWh and Missouri's was 1.0 TWh, a large difference, yet premature deaths among residents were 25 and 23, respectively. The 95% confidence intervals were 21–29 for Virginia and 21–26 for Missouri, and the death counts compared here are point estimates.
The electricity demand in the supplementary table shows the amount consumed in each state, while the death counts show the impact on the people living there. Virginia's additional demand was also met by plants outside the state through the wide-area grid PJM Interconnection, which includes Virginia. Texas's electricity demand, by contrast, was 21.8 TWh and was met mainly by increased output from plants in the ERCOT grid, which operates largely within the state. Premature deaths among the state's residents were estimated at 69 (95% confidence interval: 59–79), exceeding Virginia, which had the largest electricity demand. These state-by-state comparisons show that the health burden reaches regions with little data center demand. However, determining how much data centers in one state contributed to the health burden in another would require more detailed analysis of power plant operations and the movement of pollutants in the atmosphere.
Death counts and per-capita mortality rates also need to be considered separately. Death counts show the scale of a state's overall health burden, while per-capita mortality rates are an indicator for comparing risk among states of different population sizes. The preprint estimated that per-capita health burden was high even in Kansas and South Dakota, where data center electricity demand is low. Maps showing the concentration of data centers alone make it difficult to grasp the impact on such regions.
How Should Data Center Siting and Power Sources Be Evaluated?
According to the preprint, the atmospheric model used has a horizontal resolution of about 50 km over the contiguous United States, a setting suited to examining broad-scale changes in pollutant concentrations. Assessing impacts near individual data centers or power plants would require higher spatial resolution. The model also covered only electricity supplied from the grid, and emissions from backup generators on site were not included in the assessment.
Where the effects of air pollution appear also depends on the atmospheric model used. When the authors compared a detailed atmospheric chemistry model with a simplified model based on past emissions and atmospheric conditions, the distribution of concentration changes differed, particularly in regions where nitrate forms readily. Even when the emissions of each generator can be calculated accurately, assessing how much of the emitted material turns into particulate matter downwind requires taking local atmospheric conditions into account.
In MIT's announcement, the research team listed two topics for future study: the air quality impacts of on-site power generation at data centers, and operating methods that shift computing to different times of day. In evaluating data center siting and power sources, it matters not only how much electricity is used annually but also which plants ramp up at which times and how many residents live downwind of them. To examine the effect of shifting computing times, researchers would need to combine the differences in which plants operate before and after the change with the atmospheric conditions in each period.
