AI Data Center Health Risks Draw a Warning From Former EPA Officials

The Environmental Protection Network has identified 30 federal actions that it says intensify AI data center health risks while weakening oversight of a historic infrastructure buildout.

The nonpartisan network represents more than 800 former Environmental Protection Agency employees. Its September 10 report connects data center growth with pollution from diesel generators, gas turbines, distant power plants, and fossil fuel waste.

The warning challenges the usual framing of the AI infrastructure race. Electricity costs, water consumption, and carbon emissions have dominated the debate. Former EPA officials argue that illness, local air quality, and public participation deserve equal attention.

Their argument also creates a direct conflict with the current EPA. The agency has promoted faster construction and more flexible permitting as tools for making the United States an AI leader. It says development can accelerate without abandoning environmental protection.

That promise now faces a demanding test. Data centers require reliable electricity, but the sources supplying that electricity determine where pollution appears and who absorbs its consequences.

The Warning Extends Beyond the Data Center Fence

The central change is not a newly discovered pollutant. It is a new attempt to measure the AI boom as a connected public health system.

The Environmental Protection Network, or EPN, released a 27-page report titled Hidden Health Costs. It examines the facilities, power plants, generators, permits, and enforcement systems supporting data center expansion.

The report documents at least 30 federal actions since January 2025 that its authors believe can increase pollution-related health risks. Seventeen actions explicitly cite or target AI or data centers. The remaining 13 affect their electricity supplies, pollution controls, approvals, or enforcement.

This approach matters because a data center does not need a smokestack to shift pollution into another community. A campus can buy electricity from a regional grid supplied partly by coal or gas plants located many miles away.

Backup generators create a second pathway. Data centers install fleets of diesel engines to maintain service during an outage. Operators must also test those engines, producing emissions even when the wider grid remains available.

Some developers are pursuing dedicated gas turbines or engine plants instead. These systems can provide continuous power without waiting for a utility to complete major transmission or generation projects.

Each route creates a different pollution map. Diesel engines can concentrate nitrogen oxides, soot, and other pollutants near a campus. Utility power can transfer much of the burden to communities near distant generating stations.

Nitrogen oxides contribute to ground-level ozone, commonly called smog. Fine particulate matter, often described as soot, can penetrate deeply into the lungs and enter the bloodstream.

Exposure to these pollutants is associated with asthma attacks, heart and lung disease, stroke, and premature death. Coal combustion can also release mercury and create ash or wastewater containing toxic metals.

The report does not claim that every data center produces the same exposure. Facility size, operating schedules, power contracts, grid conditions, weather, and pollution controls all influence the result.

That distinction is essential. A national estimate can reveal the scale of a problem, but it cannot replace site-specific monitoring or tell one neighborhood exactly what it will breathe.

EPN’s broader point is that regulators need better information while projects are still being designed. Once construction, utility investments, and power contracts are locked in, cleaner alternatives become harder to substitute.

The warning therefore reaches beyond conventional emissions accounting. It asks whether the institutions responsible for measuring and limiting harm are expanding as quickly as AI infrastructure itself.

AI Data Center Health Risks Now Carry a Measurable Cost

A national modeling study gives the warning numerical weight, but its projections remain scenarios rather than observed outcomes.

Researchers from the University of California, Riverside, Caltech, and Rochester Institute of Technology modeled emissions associated with data center operations and electricity use. Their public health model uses EPA tools to estimate exposure and economic damage.

Across the study’s 2028 scenarios, data center air pollution produced estimated annual public health costs between $11.7 billion and $20.9 billion. The high-growth scenario included roughly 600,000 asthma symptom cases and about 1,300 premature deaths.

Those figures combine estimated exposure with established relationships between pollution and health outcomes. The cost calculation incorporates consequences such as medical treatment, missed work, and premature mortality.

It is not a medical bill already issued to the industry. It is a modeled estimate of harm under defined assumptions about electricity demand, generation, generator use, and pollution.

The distinction matters because some headlines have presented the upper estimate as an inevitable outcome. The underlying study offers multiple scenarios, and the actual path will depend on what gets built and how operators power it.

EPN also notes that the model does not incorporate the combined effect of the 30 federal actions in its new report. No public analysis has yet calculated how those policy changes would alter the total health burden.

That creates uncertainty in both directions. Cleaner generation, stronger controls, improved efficiency, or slower construction could reduce the impact. More fossil generation and weaker safeguards could push it higher.

One finding is especially important for understanding AI data center health risks. More than 90 percent of the modeled national health impact came from power plants supplying electricity, not generators located at data centers.

Under the model’s high-growth case, electricity generation accounted for about $19.3 billion of the $20.9 billion estimate. It also accounted for 1,165 of the 1,262 modeled premature deaths.

That result changes the geographic frame. A community hosting a data center is not necessarily the community bearing most of the associated air pollution.

Regional power markets dispatch plants according to demand, availability, transmission limits, and operating costs. Added electricity consumption can increase output from generators elsewhere, including across state lines.

The burden is also uneven. The study estimated that per-household health costs in the most affected counties could reach about seven times the national average.

On-site equipment still matters locally. The researchers modeled Virginia data center generators operating at only 10 percent of their permitted emissions levels. Even then, they estimated 14,000 asthma symptom cases and 13 to 19 premature deaths.

Estimated annual public health costs in that scenario ranged from $220 million to $300 million. A hypothetical maximum-permitted case during a prolonged regional outage produced a much larger estimate.

Those calculations should not be read as predictions for every outage. They illustrate why permitted capacity, testing schedules, actual operating hours, fuel type, and nearby populations all matter.

The report also places children near the center of the risk analysis. Developing lungs and nervous systems can make children more vulnerable to soot, ozone, mercury, and other pollutants.

EPN identified at least 188 schools within three miles of proposed gas plants dedicated to data centers. Proximity does not establish exposure or prove harm, but it highlights where detailed assessment is needed.

The numbers make the health concern harder to dismiss. They also strengthen the case for continuous monitoring, transparent operating data, and independent evaluation of the assumptions behind each permit.

The Real Conflict Is Speed Versus Preventive Oversight

The primary dispute is whether faster AI infrastructure approvals can preserve meaningful health protections before companies commit billions to construction.

The Trump administration has made domestic AI capacity a national priority. EPA Administrator Lee Zeldin has presented regulatory flexibility as part of the effort to expand reliable power and compete with China.

The agency argues that streamlining does not require abandoning environmental safeguards. Industry groups make a similar case, saying faster infrastructure development and environmental protection can coexist.

That position deserves serious consideration. Data center demand is rising faster than many utilities expected, while new transmission lines and generating resources often require years of planning and review.

Reliable electricity is not optional for cloud platforms, financial systems, hospitals, communications services, and AI products. A poorly managed transition can create grid congestion or reliability problems.

However, EPN argues that the administration is not merely processing the same protections more efficiently. Its report describes changes affecting pollution limits, public notice, environmental review, scientific capacity, and enforcement.

One proposed EPA rule would allow more construction activity before developers receive final air permits. The agency says its construction proposal would remove unnecessary barriers while maintaining required controls.

Critics see a sequencing problem. Developers gain leverage once foundations, roads, buildings, or supporting systems are underway. Regulators and communities then face greater pressure to approve the remaining pieces.

Another proposal would remove federal minimum public-participation requirements for certain minor-source permits. “Minor” refers to emissions below major-source thresholds, not to harmless facilities.

EPA says its minor-source plan would reduce duplication and give state or local agencies more control. Public notice requirements could continue where state rules provide them.

Former officials and community advocates warn that protection would become inconsistent. Residents in one state might receive detailed notice and a comment period, while residents elsewhere might learn less before approval.

The issue becomes more complicated when a large campus is developed in phases. Separate permits for generators, turbines, or buildings can obscure the combined scale if the public sees each decision in isolation.

Public participation does not guarantee a project will be rejected. It can identify nearby schools, existing pollution burdens, emergency operating risks, missing monitoring plans, or feasible control technologies.

The administration has also issued guidance stating that off-grid power plants generally fall outside the Acid Rain Program. These are plants that serve a campus directly instead of sending electricity through the wider grid.

EPA describes that off-grid guidance as clarification of existing law. Critics argue that the interpretation can leave similar turbines subject to different requirements based on their electrical connection.

EPN adds an institutional concern. Rules and permits only work when trained staff can review data, inspect equipment, verify emissions, and pursue violations.

Larry Starfield, a former senior EPA enforcement official, argues that independent oversight becomes more necessary as infrastructure expands. He warns against assuming companies can police themselves.

This is the core tradeoff. Accelerated approval can reduce delays, but preventive safeguards lose value when questions are postponed until after construction.

The choice is not simply development or no development. It is whether developers must resolve pollution controls, monitoring, public disclosure, and cumulative effects before their projects become difficult to change.

Cleaner Contracts Do Not Always Mean Cleaner Local Air

Corporate climate commitments can reduce emissions over time without guaranteeing cleaner electricity at the hours and locations serving a particular data center.

Technology companies have become major purchasers of renewable electricity. Long-term contracts can help finance wind and solar projects and support broader corporate climate goals.

Yet most clean energy claims rely on annual matching. A company compares its yearly electricity consumption with an equivalent amount of renewable generation or energy certificates.

Pollution exposure does not operate on an annual balance sheet. It depends on which plants run at a particular hour, where those plants sit, and where their emissions travel.

A data center can therefore satisfy an annual renewable-energy target while drawing grid power during hours when coal or gas plants supply the marginal demand. The accounting commitment remains meaningful, but it does not eliminate local exposure.

EPN’s report argues that this gap is poorly represented in corporate disclosures. Technology companies commonly publish energy use and greenhouse gas totals, while site-specific soot and smog impacts receive less attention.

That gap also complicates comparisons among projects. Two campuses with similar computing capacity can produce different health burdens because they connect to different grids or use different backup systems.

A campus paired with new clean generation, storage, and flexible workloads can reduce pressure during polluting hours. Another campus might depend on dedicated gas engines or extend the operating life of an older coal plant.

The report cites delayed plans to retire or convert nearly 10 gigawatts of coal capacity as electricity demand increased. It does not attribute every delay solely to data centers.

Manufacturing growth, electrification, weather, and reliability requirements also shape utility decisions. Treating AI as the only cause would overstate the evidence.

Still, data centers represent a large new load that utilities must serve continuously. The report says electricity generation and power-sector emissions increased in 2025 as data center and manufacturing demand contributed to record generation.

EPN also cites an analysis identifying at least 74 proposed gas plants dedicated to data centers. If built, those projects would add a substantial new fleet outside the conventional model of shared utility supply.

The proposal pipeline is not the same as completed capacity. Projects can be canceled, redesigned, delayed, or replaced by different power strategies.

That uncertainty is exactly why permitting choices matter now. Rules established during the planning wave can shape infrastructure that operates for decades.

Industry representatives argue that gas generation can arrive faster than some alternatives and provide steady output. They also contend that modern equipment can operate more efficiently than older power plants.

Critics respond that new gas capacity creates long-lived fuel demand and pollution sources. They want developers to demonstrate why cleaner combinations cannot meet reliability needs before receiving exceptions or accelerated approvals.

Renewables alone do not solve every hour of the reliability problem. Storage, transmission, geothermal power, nuclear generation, efficiency, and demand flexibility can each contribute to a broader portfolio.

Workload flexibility deserves particular attention. Some computing jobs can move across locations or times, although latency-sensitive services and continuous operations allow less freedom.

AI companies also influence demand through hardware selection, cooling design, model architecture, and utilization. A server that performs more useful work per unit of electricity reduces the amount of generation required.

These choices turn data center pollution into a technology design issue, not only an environmental compliance issue. Efficiency improvements can lower both operating costs and the pressure to build new power plants.

For enterprise buyers, this expands the questions worth asking cloud and AI providers. Carbon totals remain relevant, but they reveal little about hourly generation, backup fuel, local pollutants, or community exposure.

The next generation of infrastructure reporting will need greater geographic and temporal detail. Without it, companies can meet broad climate targets while leaving local health consequences difficult to see.

What the Warning Still Cannot Prove

The report establishes a credible risk pathway, but it does not prove that every federal change will cause a specific illness or quantified increase in mortality.

EPN is an advocacy organization composed of former agency professionals. Its members bring deep regulatory and scientific experience, but the group has a clear policy position favoring stronger environmental safeguards.

Its count of 30 actions combines final decisions, proposals, guidance, enforcement changes, and broader power-sector policies. These actions do not all have the same legal status or measurable effect.

Some proposed rules might change before adoption. Courts could block others, states could impose stricter requirements, and developers could voluntarily exceed federal minimums.

The national health model also relies on assumptions about data center growth, power generation, permitted emissions, and population exposure. Different assumptions would produce different totals.

EPN openly acknowledges a major evidence gap. No analysis has quantified the combined health effects of the 30 federal actions it identified.

That means the report’s directional argument is stronger than any claim about the precise incremental toll of deregulation. We know how the pollution pathways work, but not the final magnitude.

It would also be inaccurate to attribute all fossil generation growth to AI. Electricity demand is changing because of factories, vehicles, buildings, population shifts, weather, and other economic activity.

Likewise, the existence of a nearby generator does not show that residents experienced harmful exposure. Actual risk depends on operations, controls, dispersion, background pollution, and human activity.

The current EPA maintains that it can remove unnecessary permitting burdens while protecting communities. That claim should be evaluated through emissions data, compliance records, and measurable outcomes rather than rhetoric alone.

Industry’s best response would be comparable transparency. Developers could disclose hourly electricity sources, generator testing schedules, actual emissions, water discharges, and monitoring results in consistent formats.

Independent researchers would then have better evidence for testing the projected AI data center health risks. Communities could distinguish responsibly designed projects from those shifting costs onto neighbors.

Transparency would also protect companies from exaggerated claims. When operating data remains unavailable, critics and supporters both rely more heavily on scenarios and permit limits.

The same principle applies to public notice. Early participation can generate opposition, but secrecy often deepens distrust and makes later conflict more expensive.

The warning should therefore be read as a demand for measurement before irreversible commitments. It is not proof that AI development must stop.

The strongest conclusion is narrower and more defensible. Rapid infrastructure growth increases the value of science, disclosure, monitoring, and enforcement because the consequences of weak decisions multiply with scale.

Three Signals Will Show Whether Health Protections Hold

The decisive evidence will come from final EPA rules, actual power choices, and public emissions data rather than another round of competing promises.

The first signal is what happens to federal permitting and notice requirements. Proposed changes involving early construction and minor-source participation will show how the administration balances speed with preventive review.

If final rules preserve early pollution analysis, meaningful notice, and enforceable monitoring, they would weaken the claim that faster approval necessarily leaves communities less protected.

If those elements disappear, EPN’s warning gains strength. State rules would become more important, and protections could vary sharply across the country.

The second signal is the generating capacity that actually reaches construction. Proposed gas plants matter, but canceled projects do not emit pollution.

New gas approvals, delayed coal retirements, and expanded diesel fleets would reinforce concerns about health costs. Clean generation, storage, transmission, and flexible demand would point toward a lower-pollution buildout.

Companies should be judged by operational power supplies, not only procurement announcements. Hourly data can reveal whether new clean resources serve the periods when data centers increase grid demand.

The third signal is whether regulators and operators publish credible local measurements. Permit limits describe what a facility may emit, while monitoring shows what it actually releases.

Useful disclosure would include generator operating hours, fuel consumption, nitrogen oxide emissions, particulate pollution, equipment failures, and emergency events. Comparable reporting would make cumulative analysis possible across campuses.

Public opinion adds political pressure to all three signals. A 2026 national data center poll found substantial opposition to local projects, with environmental and resource concerns leading the objections.

That resistance is not simply an anti-technology response. Communities are being asked to accept large facilities, new power infrastructure, tax arrangements, noise, water demand, and uncertain health effects.

Developers that address those concerns early can create a more durable path to construction. Those that rely on secrecy or regulatory gaps invite lawsuits, moratoriums, and political backlash.

Federal leaders face a similar choice. They can treat public health review as an obstacle, or as infrastructure that keeps a rapid industrial expansion from producing preventable damage.

The EPN report puts a measurable challenge before government and industry. Protecting customers from higher electricity bills addresses only one cost of the AI boom.

Readers should now watch the permits, power plants, and monitoring systems behind each major announcement. Ask providers where the electricity comes from, what operates during an outage, and whether nearby communities can inspect the evidence.

AI data center health risks are not fixed outcomes. They are consequences shaped by engineering, energy procurement, regulation, and public accountability. The next decisions will determine whether AI capacity expands with cleaner power and stronger evidence, or transfers hidden costs to people who never chose to bear them.