US Data Center Electricity Caused 650 Premature Deaths from Dirty Air in 2023
A data center plugged into the grid forces distant coal and gas plants to run, and the wind carries their smoke into other states. MIT estimates this cost the US about 650 premature deaths in 2023, or 5.3% of all deaths from electricity generation. The risk is highest in the Midwest, even though the load is concentrated in Virginia and Texas. When choosing a site for a cluster, consider not only the price of land but also how clean the grid is.
AI-processed from MIT News; edited by Hamidun News
Imagine a new AI cluster coming online in northern Virginia. Its electricity doesn't come out of thin air. To cover the added load, the grid ramps up output at whichever plants can do so right now, and these are often older gas and coal power plants in the Ohio Valley.
The wind carries their emissions into neighboring states, so the harmful air settles far from the data center itself. According to calculations by MIT researchers, this is how electricity for data centers led to roughly 650 premature deaths in the US in 2023. That is about 5.
3% of all deaths in the country linked to electricity generation. The main cause of harm is PM2.5 fine particulate matter: the particles are so small that they penetrate the lungs and bloodstream and cause heart attacks, strokes, and lung disease.
The practical takeaway for anyone building or choosing sites: land prices and electricity tariffs are not the whole cost, because the same amount of power affects people's health differently at different points on the grid.
How researchers counted who breathes for someone else's computing
The team, led by Professor Noelle Selin and graduate student Yuan (Albert) Chen, abandoned a common simplification. Many earlier studies take the average "dirtiness" of the electric grid and multiply it by data center consumption. But new load is not covered by all plants equally. It is covered by the ones that can increase output in that particular hour. Often these are gas plants, although the picture differs across regions and at different times. As in earlier studies, the authors assumed that data centers consume power around the clock.
The researchers built a model of the US power system that calculates, hour by hour, how much electricity each of more than 10,000 generators produces and how it flows between regions. The output was emissions of nitrogen oxides (NOx) and sulfur dioxide (SO2) for each plant. Chen acknowledges that collecting the data and cross-checking it against other sources for so many generators was not easy.
The emissions were then fed into an atmospheric chemistry model, which shows how pollutants form and travel through the air. NOx contributes to the formation of ground-level ozone (smog), and NOx together with SO2 produces PM2.5. Finally, the added pollution was converted into premature mortality using data from epidemiological studies.
Why the harm is not where the servers are
Of the 650 deaths, 513 are linked to PM2.5 and 137 to smog. For scale: according to Lawrence Berkeley National Laboratory, data centers consumed about 4% of US electricity in 2023 and accounted for about 4.4% of total emissions. That share has probably grown since then, as new sites are being built very fast.
The most electricity goes to northern Virginia and Texas, but the highest mortality per capita is in Iowa, Nevada, South Dakota, and Kansas. In Virginia and Texas it is roughly half the level in most Midwestern states. "If you connect a large consumer to the grid, you may have to fire up a very dirty plant," says Selin.
There are several reasons. The power grid is interconnected, so the load in Virginia is probably covered by high-emission plants in the Ohio Valley. The wind carries the pollution into neighboring states. And in agricultural Nebraska and Iowa, the air contains a lot of ammonia from fertilizers, which boosts PM2.5 formation. The effect reaches Kansas, South Dakota, Missouri, and Arkansas, where there are relatively few data centers. Rural residents face noticeably higher risk than urban ones.
What to do in practice
Chen believes that when choosing where to place data centers and how to operate them, people should consider not only land and energy prices but also air quality. In his words, this is a good starting point for a conversation about where to build such facilities and what to power them with. For engineers and companies, this means: ask your provider for data on which plants actually cover your load, and compare sites by health impact, not just by the price per kilowatt-hour.
This is a model-based estimate, not a direct measurement: the result depends on assumptions about around-the-clock operation, on generator data, and on epidemiological coefficients. Going forward, the team wants to study on-site power plants at data centers, a growing trend, and the flexibility of the data centers themselves: how air quality would change if load were shifted adaptively.
The work was published in Environmental Research Letters. Researchers from MIT, the University of California, Irvine, and Yale University took part. The study was partly funded by the MIT Energy Initiative Society of Energy Scholars and MIT Martin Family Society of Fellows for Sustainability programs.
- Feb 12, 2026Anthropic will pay for power grid upgrades so household bills don't rise
- Nov 8, 2026Thailand froze 49 data centers and delayed decisions on another 117 projects
- Mar 1, 2026Billion-Dollar Deals: The Infrastructure Holding Up the AI Boom
- Apr 30, 2026Nvidia to invest $2 billion in Nebius, helping the former Yandex build a global AI cloud
Key Questions Answered
How researchers counted who breathes for someone else's computing?
The team, led by Professor Noelle Selin and graduate student Yuan (Albert) Chen, abandoned a common simplification. Many earlier studies take the average "dirtiness" of the electric grid and multiply it by data center consumption. But new load is not covered by all plants equally. It is covered by the ones that can increase output in that particular hour. Often these are gas plants, although the picture differs across regions and at different times.
Why the harm is not where the servers are?
Of the 650 deaths, 513 are linked to PM2.5 and 137 to smog. For scale: according to Lawrence Berkeley National Laboratory, data centers consumed about 4% of US electricity in 2023 and accounted for about 4.4% of total emissions. That share has probably grown since then, as new sites are being built very fast. The most electricity goes to northern Virginia and Texas, but the highest mortality per capita is in Iowa, Nevada, South Dakota, and Kansas.
What to do in practice?
Chen believes that when choosing where to place data centers and how to operate them, people should consider not only land and energy prices but also air quality. In his words, this is a good starting point for a conversation about where to build such facilities and what to power them with. For engineers and companies, this means: ask your provider for data on which plants actually cover your load, and compare sites by health impact, not just by the price per kilowatt-hour.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.