HomeTech NewsData Centres & EnergyMIT Study Estimates 650 Deaths From Data-Centre Air Pollution

MIT Study Estimates 650 Deaths From Data-Centre Air Pollution

A peer-reviewed MIT study of data-centre air pollution estimates that electricity demand from US data centres in 2023 caused about 650 premature deaths. The paper, by Yuang Chen and colleagues, was published on 19 November 2026 in Environmental Research Letters. It puts the total at 650, with a 95% confidence interval of 559 to 740: 513 deaths (469–557) attributed to fine particulate matter, or PM2.5, and 137 (90–183) to ozone. Those are modelled estimates, not a register of named deaths.

MIT News, in a story published the same day, says the total is “about 5.3 percent of the nation’s total electricity-related mortality,” and that “some Midwestern states had among the highest estimated rates of premature deaths per capita attributable to data centers, even though much of the data center electricity demand was concentrated in Virginia and Texas.” The paper itself states the 5.3% figure more tightly: although data centres accounted for 3.8% of US electricity generation, the estimated deaths represent 5.3% of total power-sector air-pollution mortality.

In brief

  • Chen et al. estimate 650 premature deaths in the contiguous US in 2023 (95% CI 559–740) from their model of data-centre air pollution, which couples electricity dispatch with a chemical transport model.
  • 513 of those deaths are attributed to PM2.5 and 137 to ozone. The confidence intervals reflect uncertainty in the concentration-response functions only.
  • Demand was concentrated in Virginia and Texas; MIT News says some Midwestern states had among the highest estimated per-capita death rates.

The model traces data-centre air pollution through the grid and the atmosphere

The authors couple what they call the United States Electricity Generation Optimization model with the GEOS-Chem chemical transport model. They compare hourly, unit-level electricity dispatch with and without data-centre demand, then translate the extra emissions into changes in ambient PM2.5 and ozone, and from there into premature mortality. The senior author is Noelle Selin, a professor in MIT’s Institute for Data, Systems, and Society and the Department of Earth, Atmospheric and Planetary Sciences, and director of the Center for Sustainability Science and Strategy. Co-authors are also listed at MIT, the University of California, Irvine, and Yale.

The paper finds that generation attributable to data-centre demand in 2023 was 171.0 TWh, or 3.81% of total US grid output, and that it was “largely supplied by fossil fuel units across multiple regions.” Additional annual mean PM2.5 concentrations are put at 0.02–0.08 µg m−3, with local peaks of about 0.2 µg m−3. Ozone responses are described as varying by region, with wintertime depletion and summertime enhancement. The impacts, the abstract says, are “most pronounced in the Midwest United States, where marginal emissions from coal-fired units and ammonia-rich atmospheric conditions enhance secondary particulate formation.”

Virginia and Texas, in the paper’s state-level account, account for the largest absolute data-centre loads, 33.9 and 21.8 TWh respectively. In Virginia, data centres are estimated to have made up 25.6% of state electricity consumption, against a national average of 3.8%. That is a statement about where the computers were drawing power, not about where the estimated deaths occurred.

Midwestern rates were high even when the load sat elsewhere

MIT News says “the premature mortality rates were highest in Iowa, Nevada, South Dakota, and Kansas,” while “the bulk of the electricity demand was concentrated in northern Virginia and in Texas.” Premature mortality rates in Virginia and Texas, it adds, “are only about half as much as most Midwestern states.” The paper describes the same geographic mismatch: states that host the largest data-centre loads are not necessarily those with the highest attributable mortality rates, “while some states with little data center demand bear substantial burdens.” It reports that Nebraska and Iowa “experience some of the highest attributable mortality rates, reaching 6.5 deaths per million population in Nebraska.”

Chen, quoted in MIT News, points to the interconnected grid: generators that respond to extra demand in Virginia, he says, are likely to be in the Ohio Valley, “where there is a higher proportion of older power plants.” Pollutants generated in those regions, the same story says, are then carried to downwind states. Midwest states such as Nebraska and Iowa, it notes, are agricultural regions where ammonia from fertiliser can enhance the formation of PM2.5. The estimated health impacts, MIT News adds, also extend to Kansas, South Dakota, Missouri and Arkansas.

Selin’s summary of the result is that the air-quality effects “are very widespread across a large region.” People who assume they are unaffected because there is no data centre nearby, she says, should think again.

That long-distance pattern is why the paper sits next to other recent arguments about where data centres are built and how they are powered, even though those debates were not about this model. Alberta has tabled Bill 42, a proposal for 600-metre setbacks around large data centres, a land-use response to local impacts rather than to downwind air pollution. Oregon’s governor has asked the state’s environmental regulator to pause data-centre permits until July 2027. President Trump has said large data centres should make their own power, without issuing an order. Those are separate policy stories. What the MIT paper adds is a quantified claim that the health burden of data-centre air pollution need not fall in the state that hosts the halls.

A model estimate is not a body count, and the mismatch is the result that matters

The 650 figure from this data-centre air pollution study will travel further than the caveats if it is detached from the method. It is a 2023 model estimate for the contiguous United States, produced by turning extra power-plant emissions into modelled pollution and then into deaths through epidemiological concentration-response functions. The paper is explicit that the reported confidence intervals “are derived from CRF uncertainty” and leave out uncertainties associated with emissions or air-quality modelling. In other words, the 559–740 range is not a full accounting of everything that could be wrong with the chain from a server rack to a mortality estimate.

A sensitivity test in the paper underlines the point. Using alternative, more recent US-specific concentration-response functions, the authors arrived at 565 (551–579) PM2.5-attributable deaths and only 51 (1–107) ozone-attributable deaths. The overall burden was similar and PM2.5 remained the dominant contributor, but the ozone total moved substantially. The researchers, MIT News notes, also assume that data centres draw power around the clock, as earlier studies have done, and they identify which generators respond hour by hour.

None of that makes the central finding disposable. It does mean the right use of the study is the geographic one the authors themselves emphasise. If the modelled deaths cluster in states that are not hosting the load, then siting rules written only as neighbourhood protections, or as setbacks around the campus fence, will miss a large part of the estimated harm. The practical questions the paper leaves for planners are where the marginal plants sit, what they burn, and how the weather and chemistry downwind turn those emissions into PM2.5. Those questions will not be settled by a single year’s model of data-centre air pollution. They are, however, the questions this study is actually built to raise.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.