From 600,000 New Asthma Cases Annually To A Public Health Burden Twice That Of The US Steel Industry, From PM2.5 Particles That Travel Hundreds Of Miles From The Nearest Server Farm To A Single AI Model Training Run That Generates The Pollution Equivalent Of 10,000 New York-To-LA Car Trips. The Unpaid Toll Of AI Data Centers Is Being Counted By Researchers Before It Is Paid By Communities
According to Caltech, Interesting Engineering, U.S. News and Fortune, researchers led by Shaolei Ren, an associate professor of electrical and computer engineering at UC Riverside, and Adam Wierman, the Carl F Braun Professor of Computing and Mathematical Sciences at Caltech, have published what they describe as the first comprehensive study of its kind assessing the public health impact of AI-related air pollution. The paper, titled “The Unpaid Toll: Quantifying the Public Health Impact of AI,” estimates that US data centers will generate more than $20 billion in annual public health costs by 2030 and contribute to approximately 1,300 premature deaths and 600,000 asthma symptom cases per year — a figure that would represent more than one third of all asthma-related deaths in the United States.
The energy chain that creates the health crisis. The immediate cause of the pollution is electricity. Data centres are among the fastest-growing sources of electricity demand in the United States, and the majority of that electricity is generated by fossil fuel-burning power plants — coal, natural gas, and oil — whose combustion produces nitrogen dioxide, sulphur dioxide, and fine particulate matter with a diameter of 2.5 micrometres or less, known as PM2.5. Data centres also rely on diesel backup generators that activate during power outages or grid stress events, adding a second, more locally concentrated pollution source. The researchers used a tool provided by the US Environmental Protection Agency to model how these emissions travel through the atmosphere and how much health damage they cause at different distances from data centre clusters.
The finding that surprised many observers is not the existence of the pollution but its geographic reach. “Public health impacts are direct and tangible impacts on people, and these impacts are substantial and not limited to a small radius of where data centres operate,” Ren told Interesting Engineering. Fine particulate matter can remain suspended in the atmosphere and travel hundreds of miles from its emission source, meaning that the communities bearing the health cost of data centre pollution are frequently not the communities that receive any economic benefit from the data centres producing it.
The numbers and what they mean. The study projects that by 2030, the public health costs associated with data centre air pollution will exceed $20 billion annually — covering cancers, asthma and other respiratory diseases, cardiovascular conditions, and missed workdays and school days. For comparative context, this figure is projected to be twice the public health cost of the entire US steel-making industry. It may also rival the health impact of all cars, buses, and trucks in California — the most populous state in the country and the one with the strictest vehicle emissions standards in the world.
The 1,300 premature deaths figure requires its own context. These are statistical deaths — a methodology used routinely in environmental health research that calculates, across a large population, the number of deaths that can be attributed to a given pollutant exposure above baseline. They are not 1,300 identifiable individuals. They are the aggregated endpoint of a pollution exposure distributed across millions of people in communities downwind of data centre electricity generation. This is the same methodology used to calculate the health toll of vehicle emissions, industrial smokestacks, and wildfire smoke — and it is the standard used by the EPA in all regulatory impact assessments.
The 1,300 figure also represents a 36% increase over current annual asthma-related deaths in the United States — a specific, measurable consequence of a specific, measurable increase in power grid emissions driven specifically by data centre demand growth.
Training one AI model generates the equivalent of 10,000 cross-country drives. The researchers provided a concrete illustration of the scale of emissions involved. The electricity required to train a single large language model at the scale of Meta’s Llama 3.1 — one commercially deployed AI model among dozens now in production — would generate enough air pollutants to be equivalent in health impact to a passenger car completing 10,000 round trips between New York City and Los Angeles. That is not a lifecycle calculation. It is not an estimate of the model’s total operational footprint over its deployed lifetime. It is the emissions from a single training run of a single model.
This figure does not include inference — the ongoing computational process of running the model in response to queries, which occurs billions of times per day across commercially deployed models and which, in aggregate, consumes significantly more electricity than the initial training run.
Virginia as a case study. The state of Virginia hosts the largest concentration of data centres in the world by market cap, accounting for roughly a third of all US data centre capacity. The state’s electricity grid relies heavily on natural gas and, in some regions, coal, meaning that its data centre density translates directly into proportionally high pollution emissions. The study estimates that backup diesel generators in Virginia alone could be responsible for approximately 190 air pollution-related deaths — a figure Ren and his colleagues use to illustrate that even a single component of data centre operations at a single geographic cluster can generate a meaningful public health burden.
What tech companies are not reporting. The study’s most pointed institutional critique is directed at corporate sustainability reporting. Despite the scale of the health costs the research documents, the vast majority of tech companies’ sustainability reports make no mention of the emission of unhealthful air pollutants. They focus instead on carbon emissions — which are real and important — and, increasingly, water usage for cooling systems. “If you look at sustainability reports by tech companies, they only focus on carbon emissions, and some of them include water as well, but there’s absolutely no mention of unhealthful air pollutants and these pollutants are already creating a public health burden,” Ren said.
Adam Wierman made the same point from a community impact perspective. “When we talk about the costs of AI, there has been a lot of focus on measurements of things like carbon and water usage. And while those costs are really important, they are not what’s going to impact the local communities where data centres are being built. Health is a way of focusing on the local impact these data centres are having on their communities and understanding, quantifying, and managing those impacts, which are significant.”
The implication is that current sustainability reporting frameworks are structurally incomplete. They capture the global externalities of data centre operations — carbon’s contribution to climate change — but not the local externalities — the PM2.5 that the child with asthma three counties downwind of a natural gas plant is breathing today because that plant is running at higher capacity to serve a data centre contract.
A separate April 2026 analysis from Carnegie Mellon extends the picture. A separate study published in April 2026, authored by economist Nicholas Muller of Carnegie Mellon University and reported by Fortune, analysed approximately 2,800 operational data centres and found that their environmental damage last year cost the economy $25 billion in total — of which $3.7 billion was directly attributable to AI activities. Muller’s methodology assigned an economic value to the shortened life expectancy caused by PM2.5 exposure, producing a social cost figure that is comparable to but distinct from the UC Riverside-Caltech public health burden calculation.
The convergence of two independently conducted studies on two different datasets at two different institutions arriving at figures of the same order of magnitude — $20 billion and $25 billion respectively — is the kind of scientific convergence that typically precedes regulatory attention.
The disproportionate burden on low-income communities. The health burden is not distributed equally. Environmental Health Sciences notes that the UC Riverside-Caltech study found that the health impacts of data centre pollution disproportionately affect certain low-income communities. This is consistent with decades of environmental justice research showing that fossil fuel infrastructure — power plants, highways, industrial facilities — is systematically sited in proximity to communities with less political and economic power to resist it. Data centre electricity demand does not alter the location of the existing power plants that generate the pollution; it increases their utilisation rates, and the communities already living with the baseline pollution from those plants absorb the incremental health cost.
The verdict. Ren captured the stakes directly: “If you have family members with asthma or other health conditions, the air pollution from these data centres could be affecting them right now. It’s a public health issue we need to address urgently.” The growth of AI infrastructure is accelerating, not slowing. The projections in the UC Riverside-Caltech paper were made before the most recent wave of data centre expansion announcements — including the commitments from Microsoft, Google, Amazon, and Meta to spend hundreds of billions of dollars on new AI infrastructure through 2028.
The question the study poses is not whether AI is useful. It is who pays for the costs that AI’s energy consumption currently externalises onto communities that have no say in the contracts that place power plant demand on their local grid.
To check out our previous coverage on AI, data centres, energy infrastructure, and public health, read our articles here.
