Walk past a modern artificial intelligence data center and you will see rows of windowless halls and cooling units, without a single smokestack in sight. Yet a new study from researchers at the Massachusetts Institute of Technology (MIT), the University of California at Irvine, and Yale University finds that the grid-supplied electricity running US data centers in 2023 was linked to an estimated 650 premature deaths across the country. Those modeled deaths are not proved individual cases; they are statistical counterfactual estimates of mortality attributable to power-plant pollution that would not have occurred without data center electricity demand.
What surprised the researchers most was where that health burden fell. While server farms cluster heavily in states such as Virginia and Texas, the highest per-capita mortality rates showed up hundreds of miles away in Midwest and Great Plains farming states, including Nebraska, Iowa, Kansas, and South Dakota, as well as Nevada. Because regional power grids draw on older coal-fired units and natural gas plants when demand climbs, and prevailing winds push those exhaust plumes across state lines, living far from a server campus offers little protection from data center air pollution.
“When people think about the negative impacts of data centers, they usually focus on the neighbors. But our study indicates that the air quality impacts of data centers are very widespread across a large region. Even if you don’t think you are affected because there is not a data center near you, think again,” said Noelle Selin, a professor in MIT’s Institute for Data, Systems, and Society (IDSS) and the Department of Earth, Atmospheric, and Planetary Sciences, and senior author of the MIT News study announcement.
What the MIT Team Actually Modeled
To trace how electricity use inside a server hall translates into downwind smog and fine particles, lead author Yuang (Albert) Chen, an MIT graduate student, and his colleagues built a two-stage model of the 2023 US power grid and atmosphere. Their peer-reviewed Environmental Research Letters study, published on November 19, 2026, goes beyond earlier estimates that relied on simple grid-wide emission averages.
First, the team used a chronological unit-level dispatch model called US-EGO to simulate hourly operations across more than 10,000 US generators with and without data center electricity demand. Unlike average-emission methods, US-EGO accounts for minimum stable output levels and hour-to-hour ramping constraints on coal and combined-cycle natural gas plants, along with storage and hydropower links. Because high-resolution hourly profiles for individual server campuses are rarely public, the model assumed data centers drew a uniform, round-the-clock baseline load throughout 2023 and focused strictly on grid-supplied electricity.
“It was a challenge to gather all the data,” Chen noted. “There are more than 10,000 generators in the U.S. grid, and we needed to make sure the data provided for each generator is valid and validated across other data sources.”
Second, the researchers fed those hourly smokestack emissions into GEOS-Chem High Performance (version 14.4.0), a three-dimensional chemical transport model running across 72 vertical atmospheric layers and stretched to roughly 50-kilometer horizontal resolution over the contiguous United States. Full atmospheric chemistry matters because power plants do not just release soot directly. They emit gases that cook in sunlight and react with farm emissions downwind, a chemical pathway that simpler reduced-form models such as the InMAP Source-Receptor Matrix (ISRM) tend to underrepresent, particularly for Midwest nitrate particles.
That detailed chain explains why the MIT team’s estimate of 650 modeled premature deaths in 2023 is roughly 40 to 60 percent higher than prior studies that used average grid mixes. Earlier work by Han and colleagues estimated about 401 deaths for 2023, while an economic analysis by Nicholas Muller estimated $25 billion in total external costs for 2025, including $4.6 billion in local air pollution damage (which corresponds to roughly 459 premature deaths at a $10 million value of a statistical life).
How Data Centers Pollute Without a Smokestack
What are data centers from an energy perspective? They are energy-intensive buildings packed with servers, storage drives, and cooling systems that power cloud computing, web services, and artificial intelligence training. According to a Lawrence Berkeley National Laboratory energy analysis cited by the MIT team, US data centers consumed between 150 and 200 terawatt-hours (TWh) of electricity in 2023, representing about 4 percent of total US power demand and roughly 4.4 percent of power-sector greenhouse gas emissions. (One terawatt-hour is 1 billion kilowatt-hours, enough electricity to power roughly 90,000 average US homes for a year.)
Why would a sector that uses 3.81 percent of grid electricity in the MIT model (171.0 TWh) be responsible for 5.3 percent of all power-sector air-pollution mortality? The answer lies in marginal emissions, meaning the extra pollution released by the specific standby generators that ramp up when a big new customer plugs into the grid. Renewable wind and solar farms generally run whenever the weather allows, and nuclear reactors run flat-out as baseload power. When a data center adds continuous 24/7 demand on top of existing load, grid operators must call on spare fossil-fuel capacity that would otherwise sit idle or run at lower output.
“Different regions have different mixes of power. If you turn on a major new demand source, by plugging a data center into the grid, you might have to fire up your really dirty power plant to meet that new demand,” Selin explained.
Because those marginal generators burn dirtier fuel or lack the newest emission controls, data center electricity demand punched far above its weight in smokestack pollutants during 2023:
- Grid electricity share: 171.0 TWh attributable to data centers, or 3.81 percent of total US grid generation.
- Carbon dioxide (CO2) share: 103.5 teragrams (million metric tons) of CO2, or 6.69 percent of US power-sector CO2 emissions.
- Nitrogen oxides (NOx) share: 61.4 kilotons (thousand metric tons) of NOx, or 7.32 percent of power-sector NOx emissions.
- Sulfur dioxide (SO2) share: 39.1 kilotons of SO2, or 5.63 percent of power-sector SO2 emissions.
- Peaker plant impact: High-emitting peaking generators (units larger than 10 megawatts that run less than 15 percent of the year in baseline conditions) supplied only about 4 percent of data center electricity, yet produced 7 percent of the attributable NOx pollution.
“Generators that respond to additional electricity demand are often natural gas plants, although the response varies by place and hour,” Chen said. In the Midwest and parts of the eastern grid, older coal-fired units also respond heavily on the margin, releasing both SO2 and NOx that transform in the atmosphere into fine particulate matter (PM2.5, particles 2.5 micrometers or smaller that can lodge deep in the lungs and bloodstream) and ground-level ozone, the primary lung irritant in smog.
Why the Midwest Breathes Virginia’s Server Load
If you map where US data centers sit, a handful of states dominate. In 2023, Virginia led the nation with 33.9 TWh of data center electricity consumption (25.6 percent of all power used in the state), followed by Texas at 21.8 TWh, California at 9.3 TWh, Illinois at 7.5 TWh, and Oregon at 6.4 TWh. Data centers also took up a large slice of state electricity in Iowa (11.4 percent) and Nebraska (11.7 percent), compared with the 3.8 percent national average.
Yet when the MIT team mapped where premature mortality rates per capita ran highest, the picture flipped. Virginia and Texas experienced per-capita mortality rates roughly half those of most Midwest states. Meanwhile, the highest per-capita rates appeared in Iowa, Nevada, South Dakota, Kansas, and Nebraska, where attributable deaths reached 6.5 per million residents. Elevated mortality rates stretched across Kansas, South Dakota, Missouri, and Arkansas, states with relatively small server footprints of their own.
“While it is not surprising that any additional source of air pollution will result in mortalities, with data centers, when and where these deaths are occurring are different than you might expect,” Selin said.
Three connected forces drive this geographic mismatch between where servers hum and where people breathe the pollution:
- Cross-state grid dispatch: Texas operates largely on its own grid (ERCOT), so its data center demand is met mostly by in-state power plants. Virginia, however, sits inside the sprawling PJM Interconnection. When Northern Virginia server campuses ramp up demand, PJM frequently dispatches older, higher-emitting coal-fired units and natural gas plants westward across the Ohio Valley.
- Downwind atmospheric transport: Prevailing winds carry NOx and SO2 plumes over hundreds of miles as they convert into secondary particles and ozone, spreading exposure across entire regions rather than confining it to a neat radius around a facility.
- Midwest agricultural ammonia: As power-plant NOx and SO2 plumes drift over corn and livestock country in the Midwest and Great Plains, they mix with abundant ammonia gas rising from fertilized fields and animal operations. That ammonia reacts readily with nitric and sulfuric acids to form ammonium nitrate and ammonium sulfate aerosols, supercharging secondary PM2.5 concentrations over rural farming communities.
| State / Region | 2023 Data Center Electricity Load | Share of State Electricity | Modeled Per-Capita Mortality Burden and Grid Mechanism |
|---|---|---|---|
| Virginia | 33.9 TWh | 25.6% | Moderate (about half Midwest rates); PJM grid dispatches Ohio Valley plants outside the state |
| Texas | 21.8 TWh | Below VA share | Moderate (about half Midwest rates); ERCOT grid meets demand mostly with in-state generation |
| Nebraska | Moderate absolute load | 11.7% | 6.5 deaths per million residents (among highest in US); downwind transport plus farm ammonia chemistry |
| Iowa | Moderate absolute load | 11.4% | Among highest per-capita mortality rates in US; Midwest coal and gas dispatch and agricultural ammonia |
| Kansas, South Dakota, Missouri, Arkansas | Low data center load | Near or below 3.8% US average | Elevated per-capita mortality rates driven by imported downwind PM2.5 and ground-level ozone |
Because electricity grids and wind plumes span hundreds of kilometers, the study found almost no difference in per-capita health risk between counties that host data centers, neighboring counties, and counties much farther away. Across the country, the distance gradient was a negligible -0.192 deaths per 100,000 people for every 100 kilometers moved away from a data center. In short, moving a few towns over does not spare residents from grid-level data center air pollution.
In fact, rural communities bore a heavier relative burden than cities. Modeled mortality averaged 0.27 deaths per 100,000 residents in rural areas (95 percent confidence interval: 0.24 to 0.31), compared with 0.21 per 100,000 in urban areas (95 percent confidence interval: 0.19 to 0.25). Even within individual states, non-urban residents faced a median exposure 6.3 percent higher than the regional average. Across racial and income groups, national averages were 0.24 per 100,000 for White and American Indian or Alaska Native populations, 0.19 for Asian populations, and 0.18 for Hispanic populations, with within-state racial gaps near zero and only minor differences across income quintiles (0.21 to 0.24 per 100,000).

PM2.5 and Ozone Behave Differently
Of the 650 modeled premature deaths attributable to 2023 data center grid power (95 percent confidence interval: 559 to 740), fine particulate matter accounted for the largest share, with 513 deaths (95 percent confidence interval: 469 to 557). Ground-level ozone accounted for the remaining 137 deaths (95 percent confidence interval: 90 to 183), based on maximum daily 8-hour average (MDA8) ozone exposure. When the authors tested alternative US-specific concentration-response functions in a sensitivity analysis, modeled mortality was 565 deaths from PM2.5 (95 percent confidence interval: 551 to 579) and 51 from ozone (95 percent confidence interval: 1 to 107).
In raw concentration terms, data center electricity demand added a median of 0.017 micrograms per cubic meter of annual average PM2.5 across the contiguous United States. Across much of the central and eastern states, the increase ranged from 0.02 to 0.08 micrograms per cubic meter, with local hotspots reaching about 0.2 micrograms per cubic meter. For everyday scale, typical annual baseline PM2.5 concentrations in the United States range from 5 to 12 micrograms per cubic meter. An addition of 0.02 to 0.08 micrograms per cubic meter is a fraction of one percent of total background air pollution, undetectable to an individual breath, yet large enough across tens of millions of downwind residents to produce hundreds of statistical premature deaths over a year.
Ground-level ozone showed an even broader footprint and a distinct seasonal quirk. Annual MDA8 ozone rose by a median of 0.042 parts per billion (ppb), with net annual increases in 99.2 percent of US land grid cells. During the winter, however, fresh nitric oxide (NO) plumes near dense power-plant corridors in the Mid-Atlantic and along the Colorado Front Range actually caused localized ozone dips, a well-known chemical effect called NOx titration where fresh NO temporarily scavenges ozone molecules in low-sunlight conditions. Come summer, strong sunlight and warm temperatures flip that chemistry into a NOx-limited regime across vast downwind stretches of the country, driving widespread summer smog increases that more than outweigh the winter dip.
What This Means for Siting and Cleaner Power
As technology companies race to build larger server campuses for generative AI, siting decisions have historically revolved around cheap land, fiber-optic trunk lines, tax incentives, and local electricity rates, alongside engineering efforts to find the wildest places data centers stay cool to trim cooling energy. Chen noted that developers and grid planners should treat air quality as a core variable alongside land and energy costs, using unit-level models as a baseline when deciding where to build and how to supply power.
Better visibility into where server loads are plugging in is a necessary first step. Federal trackers have only recently begun to catch up, with the EIA starting to count data center power separately as states like Maine weigh pauses on large 20-megawatt facilities over grid and ratepayer strains. Adding regional air-quality modeling to permitting and grid transmission planning could help regulators see when a proposed server cluster will end up calling on high-emitting coal or peaker units two states away.
The study points toward three practical levers to shrink data center air pollution:
- Procuring clean firm power: Matching server demand with 24/7 low-emission baseload or firm generation, rather than relying on annual renewable credits in grids where coal and gas still serve the margin, directly prevents dirty standby units from ramping up.
- Siting in regions with cleaner marginal grids: Placing new load where the marginal generator is cleaner, or where downwind ammonia and population exposure are lower, cuts health damages per terawatt-hour consumed.
- Shifting flexible computing workloads: Scheduling non-urgent AI training jobs away from peak demand hours can keep high-NOx peaker plants from firing up, an approach the MIT team plans to evaluate in future research.

Limits to Keep in View
The authors emphasize several boundaries around their findings so policymakers and communities interpret the numbers accurately. First, the study captures a 2023 baseline snapshot under a uniform 24/7 load assumption, and its roughly 50-kilometer atmospheric grid cells are designed to capture regional transport rather than block-by-block pollution gradients right next to an individual smokestack or server building.
Second, the model looks strictly at grid-supplied electricity and excludes on-site power generation. Across the United States, behind-the-meter generation at data centers totaled less than 3 gigawatts in the study’s baseline, and prior research under typical operating assumptions estimated that on-site diesel backup generators cause roughly 30 premature deaths per year nationally (compared with hundreds from grid electricity). However, as grid interconnection queues lengthen, developers are increasingly turning to behind-the-meter gas turbines at AI factories, creating local permitting gaps and on-site exhaust plumes that are not included in the 650-death total and that the MIT researchers flag for future study.
Third, the health calculation counts only secondary PM2.5 (formed from NOx and SO2) and ground-level ozone. It excludes primary PM2.5 emitted directly from smokestacks as well as the direct health effects of breathing nitrogen dioxide (NO2) gas. Leaving out primary particles, direct NO2 exposure, and on-site diesel or gas turbines means the study likely underestimates total air quality health damages, especially in neighborhoods close to power plants and server campuses. Furthermore, the confidence intervals reflect statistical uncertainty in the epidemiological concentration-response functions only, and assuming equal baseline vulnerability across populations may underestimate disparities for lower-income or medically vulnerable communities. Because the model tracks incremental annual averages, it cannot be used on its own to determine whether data centers push specific counties out of compliance with federal National Ambient Air Quality Standards, nor does a state-level rural tilt rule out localized environmental justice burdens near individual facilities.
Frequently Asked Questions
Does a data center pollute if it has no smokestack?
Yes, indirectly. While a server hall itself emits no exhaust during normal grid-powered operation, its continuous 24/7 electricity demand forces regional power grids to ramp up marginal natural gas plants, coal-fired units, and high-emitting peaker plants. In 2023, US data centers used 3.81 percent of grid electricity (171.0 TWh) but accounted for 7.32 percent of power-sector NOx and 5.63 percent of SO2 emissions.
How far away can data center air pollution reach?
Data center air pollution travels across entire multi-state regions rather than stopping at a fixed radius. Because electricity grids dispatch power plants far from the servers and prevailing winds carry NOx and SO2 plumes hundreds of miles downwind, the MIT study found almost no change in per-capita mortality risk with distance from a data center (-0.192 deaths per 100,000 people per 100 kilometers). Living far from a server hub does not eliminate exposure.
Are data centers bad for health?
At the local level, a single grid-connected data center adds only a tiny fraction of a microgram per cubic meter to regional PM2.5 levels. Stacked across the entire US grid and population, however, 2023 data center electricity use was linked to an estimated 650 modeled premature deaths from PM2.5 and ozone, with rural and Midwest communities experiencing the highest per-capita rates. These regional grid emissions are completely separate from localized on-site concerns such as cooling fan noise, backup diesel generator exhaust, or local water use, which the MIT grid study did not evaluate.
Will cleaner power or moving data centers fix it?
Simply moving a data center to a rural county does not solve the problem if it still plugs into a regional grid where coal or older gas plants respond on the margin. Reducing the health toll requires factoring downwind air quality into siting decisions, shifting flexible computing loads away from dirty peak hours, and backing new server demand with clean firm power that keeps fossil-fuel peakers and coal units from ramping up.
