1,300 Deaths, 600,000 Asthma Cases, $20 Billion a Year: Ex-EPA Officials Publish the Hidden Health Bill of the AI Boom

The most pointed critique of the AI infrastructure boom this week did not come from a rival lab, a stranded investor, or a senator. It came from the people who used to run America’s pollution police. On September 10, the Environmental Protection Network (EPN) — a nonpartisan group of more than 800 former EPA scientists, engineers, toxicologists, and enforcement attorneys — released “The Hidden Health Costs of AI Data Centers,” a report documenting at least 30 federal actions since January 2025 that, taken together, weaken the protections standing between the data center buildout and the people breathing its exhaust.

The report’s core argument is about direction of travel, not just magnitude. The United States is adding gigawatts of AI-driven electricity demand at a pace with no modern precedent, and in the same window the federal government has been loosening soot, mercury, and air-toxics rules, slowing coal retirements, trimming public review, and cutting the science and enforcement staff needed to know what any of it emits. “You would think a buildout of this magnitude would mean more vigilance from EPA,” said Dr. Lynn Goldman, a pediatrician and former EPA Assistant Administrator, in the report’s launch materials. “Instead, EPA and other federal agencies are doing the opposite.”

The numbers: a $20 billion-a-year externality

To size the stakes, EPN leans on a national study from researchers at UC Riverside, Caltech, and the Rochester Institute of Technology. Their central estimate: data-center-related air pollution could impose roughly $11.7 billion to $20.9 billion in annual public-health costs in the United States by 2028. Under the study’s high-growth scenario, that burden translates into roughly 600,000 asthma symptom cases and about 1,300 premature deaths every year. The New York Times, which covered the report on September 10, led with the same figures.

Two structural details in that study matter more than the headline number. First, more than 90 percent of the estimated health impact comes not from the data centers themselves but from the power plants supplying them — often many miles away. The pollution is diffuse, regional, and disconnected in the public mind from the AI economy driving it. “You do not have to live near a data center to be affected by it,” Goldman said. Second, the electricity system is bending around the new demand in visible ways: plans to retire or switch nearly 10 gigawatts of coal capacity have been delayed as demand rose, and an Environmental Integrity Project analysis counts at least 74 proposed gas plants dedicated specifically to data centers.

Crucially, EPN is explicit about what the numbers do not capture. The academic study models growth and emissions scenarios — it does not estimate the incremental health effect of the 30 policy actions EPN documents. Dr. Dan Costa, who spent about 35 years at EPA and led its national research program on air quality, climate, and energy, calls that missing analysis exactly what the agency should be producing now: measuring what these facilities actually emit, modeling local and regional air quality, and translating the difference into health impacts. That work, the report implies, is not happening.

Thirty actions, three failures

The heart of the report is an inventory: 30 federal actions since the start of 2025, of which 17 explicitly cite or target AI or data centers, with the remaining 13 shaping how data centers are powered, permitted, and policed. EPN sorts them into three buckets.

Pollution safeguards. Federal agencies are keeping fossil-fired generation online longer while weakening or reconsidering standards for soot, mercury and air toxics, interstate smog pollution, oil and gas pollution, coal wastewater, and coal ash. Each rollback is individually technical; cumulatively they raise the emissions baseline against which the largest demand surge in a generation lands.

Permitting and public review. Agencies are allowing projects to move further ahead before key pollution requirements are resolved, and reducing — or proposing to reduce — the public’s opportunity to learn about projects and object before money is committed. This extends a story this blog covered earlier in the month: EPA’s proposal to end mandatory public notice for “minor source” air permits, the category covering many data center backup generators. EPA’s own justification was that such participation is not generally “necessary.”

Science, monitoring, and enforcement. The same agencies have lost scientific and enforcement capacity — the people and instruments that measure emissions, assess risk, inspect facilities, and refer violations. Larry Starfield, former Principal Deputy Assistant Administrator for EPA’s enforcement office, puts the point bluntly: “Passing a law is not effective unless you enforce it. And issuing a permit is not the same thing as ensuring compliance with it. The AI buildout makes the enforcement function more necessary, not less.”

The community dimension threads through all three. Data center projects routinely proceed under nondisclosure agreements and accelerated review timelines that leave residents learning the details late. Federal actions now also seek broader authority to end Clean Air Act citizen suits when regulators decline to act — closing the courthouse door as the agency door narrows. Walter Mugdan, a former EPA Deputy Regional Administrator with 48 years at the agency, frames the trade-off in a sentence built for a poster: “Faster permitting shouldn’t mean ‘build first and ask the pollution questions later.’”

A pledge, and a campaign

EPN’s ask is deliberately modest in form and sharp in politics. The White House has promoted a Ratepayer Protection Pledge — asking AI developers and utilities to keep families from shouldering higher electricity costs from the boom. EPN, under its “Safer, Not Sicker” campaign, is calling for a Data Center Health Protection Pledge on the same logic: “If families shouldn’t have to pay for the AI boom through their electricity bills, they shouldn’t have to pay for it with their health either.” The report also asks EPA to do the unglamorous things it used to do: measure how the buildout is changing pollution, identify who is exposed, publish the results, and rebuild the monitoring and enforcement capacity that measurement presupposes.

The honest caveats

Three caveats keep this from being a clean morality tale. First, the mortality and cost figures come from a growth-scenario model, not from observed deaths; they are projections with wide uncertainty bands, and EPN did not itself quantify the rollbacks’ added harm. Second, the counterargument — that faster permitting and firm power are what a nationally strategic technology requires, and that gas plants dispatched for data centers are at least newer and cleaner per megawatt-hour than the aging fleet they extend — gets a hearing in policy circles even if the report gives it none. Third, EPN is an organization of former officials whose institutional posture has been consistently critical of the current administration; its inventory of actions is documented, but its synthesis is advocacy.

Even discounted for all of that, the report’s central arithmetic stands: the AI economy’s power draw is on a path to impose a public-health bill in the tens of billions of dollars annually, the burden falls most heavily on people downwind rather than downstream of the profits, and the federal government is simultaneously reducing its ability to measure, permit, and enforce against exactly that pollution. Whatever one thinks of the pace of AI progress, the emissions are not hypothetical — and as of this month, nobody in Washington is formally assigned to count them.

For now, the loudest accounting of the AI boom’s health costs comes from the people who used to do that accounting for a living — and who now have to publish it from outside the building.