EPN Data Center Pollution Warning Puts AI’s Health Costs in Focus

The Environmental Protection Network has issued an EPN data center pollution warning built around 30 federal actions and a growing public health concern. The group argues that Washington is accelerating AI infrastructure while weakening safeguards governing its power supply. That combination creates a conflict between faster construction and the health protections surrounding new electricity demand.

The warning, published September 10, does not claim that every data center directly sickens nearby residents. It makes a broader case about the power plants, generators, turbines, and waste systems supporting rapid development. Its central challenge is aimed at the Environmental Protection Agency, which has made faster AI infrastructure development a stated priority.

EPA maintains that permitting flexibility can support grid reliability, restrain electricity costs, and protect communities. Former agency officials say that position overlooks cumulative pollution and shrinking oversight capacity. The disagreement has moved beyond carbon accounting into asthma, heart disease, public participation, and who absorbs AI’s physical costs.

The Warning Connects AI Growth to 30 Federal Actions

The new report treats data center pollution as a policy system, not a collection of isolated projects.

The Environmental Protection Network, or EPN, represents more than 800 former EPA specialists and policy professionals. Its members include scientists, toxicologists, engineers, enforcement officials, and former agency leaders. On September 10, the organization released a 27-page assessment titled Hidden Health Costs.

The report identifies at least 30 federal actions taken since January 2025. Seventeen explicitly concern AI or data centers. The remaining 13 affect electricity generation, pollution limits, permitting, public review, scientific capacity, or enforcement.

EPN groups those actions into three connected areas. The first covers weaker or reconsidered pollution safeguards for fossil fuel production and power generation. The second covers faster approvals with fewer opportunities for public participation. The third covers reduced federal capacity to measure emissions and enforce existing requirements.

That framing matters because a data center building produces only part of the relevant environmental footprint. Servers consume electricity continuously, while cooling and backup systems add local demands. The generation needed to support those operations can produce pollution many miles from the computing campus.

On-site diesel generators and gas turbines can emit nitrogen oxides, soot, and volatile organic compounds. Nitrogen oxides and volatile compounds help form ground-level ozone, commonly called smog. Fine particulate matter can penetrate deep into the lungs and enter the bloodstream.

Coal generation adds other risks, including mercury, toxic metals, ash, and contaminated wastewater. Those effects appear near power plants and waste sites rather than beside the data center itself. A narrow review of the campus can therefore miss much of the exposure associated with its electricity.

Former EPA Assistant Administrator Lynn Goldman summarized the concern in unusually direct terms. She said a buildout of this magnitude should bring stronger safeguards, more science, and more monitoring. Instead, she argued, federal agencies are moving in the opposite direction.

EPN also stresses that its list does not constitute a complete health-impact model. No public analysis has calculated the combined effect of all 30 actions. The report establishes policy direction and plausible exposure pathways, but it does not prove a specific national disease total caused by those decisions.

That limitation does not make the warning empty. It reveals the central information gap. Federal policy is changing faster than regulators are measuring the cumulative consequences.

The EPN data center pollution warning therefore changes the debate’s unit of analysis. Individual permits remain important, but the larger question concerns simultaneous changes across energy, pollution controls, enforcement, and public access.

The Largest Health Burden May Originate Far from Data Centers

The most consequential pollution source is often the electric grid, not the row of backup generators visible at a data center.

EPN relies heavily on a 2026 research paper from scholars at UC Riverside, Caltech, and Rochester Institute of Technology. Their health-impact model estimates pollution connected to data center electricity and backup generation under several growth scenarios.

The researchers estimate annual public health costs between $11.7 billion and $20.9 billion by 2028. Under their high-growth scenario, the modeled pollution contributes to roughly 600,000 asthma symptom cases and about 1,300 premature deaths.

These figures are projections, not a count of diagnosed patients linked to identified facilities. They depend on assumptions about electricity growth, grid composition, emissions, atmospheric movement, exposure, and established health-response relationships. Different assumptions would produce different totals.

The distinction is essential because EPN’s report arrived with a forceful political argument. Readers should not confuse that advocacy with the underlying study’s uncertainty ranges. The model estimates potential health effects under defined scenarios.

Yet one result changes how the issue should be understood. More than 90 percent of the modeled health burden comes from power plants supplying electricity to data centers. On-site generators account for a smaller national share, although they can create concentrated local risks.

In the high-growth case, electricity generation represents approximately $19.3 billion of the estimated $20.9 billion annual burden. It also accounts for 1,165 of the model’s 1,262 premature deaths. That geographic separation complicates local accountability.

A community hosting a data center might experience generator exhaust, turbine emissions, construction traffic, water pressure, and noise. Another community could experience pollution from the coal or gas plant supplying its electricity. Downwind communities can face additional exposure across county or state lines.

This mechanism also weakens simple corporate claims about renewable procurement. A technology company can buy enough renewable energy annually to match its total consumption. However, annual matching does not guarantee clean electricity at the location and hour when a facility operates.

The grid must balance supply and demand continuously. When local renewable production falls, another generator responds. If that marginal source burns coal or gas, the associated emissions occur despite the company’s annual clean-energy accounting.

EPN says data centers consumed about 4.7 percent of American electricity in 2024. It cites a federal laboratory projection placing their 2030 share near 11.8 percent, with scenarios ranging from 9.5 percent to 15.3 percent.

Those percentages include uncertainty about chip efficiency, server utilization, model development, inference demand, and construction schedules. Still, every scenario represents a major new industrial load. The public health outcome depends heavily on which resources meet it.

Plans to retire or convert nearly 10 gigawatts of coal capacity were delayed as electricity demand increased, according to research cited by EPN. Data centers were not the only cause. Manufacturing growth, electrification, reliability concerns, and extreme weather also influenced those decisions.

The report also points to at least 74 proposed gas plants dedicated to data center projects. If built at the described scale, these facilities would total 143 gigawatts of generation capacity. Their locations and operating patterns would determine the real exposure.

The result is an indirect but measurable chain: AI use drives computing demand, computing drives electricity demand, and generation decisions shape pollution. Each link gives companies and regulators room to deny sole responsibility. Together, the links define the health question.

EPN Data Center Pollution Claims Challenge EPA’s Permitting Strategy

The primary conflict is between EPA’s promise of faster, flexible development and EPN’s demand for stronger oversight before construction becomes difficult to reverse.

EPA presents its approach as a response to two immediate pressures. The United States wants more domestic AI capacity, while utilities must serve large new loads without destabilizing the grid. Developers also argue that lengthy approvals can delay generation needed for completed computing facilities.

The agency’s data center resources say streamlined Clean Air Act reviews can help developers build with greater confidence and speed. EPA also says flexibility can preserve reliability and reduce costs while protecting health.

EPN does not argue that permitting must remain slow. Its position is that faster decisions should retain emissions analysis, enforceable conditions, monitoring, disclosure, and meaningful public review. Speed becomes dangerous when major commitments precede answers about pollution.

One flashpoint involves islanded power generation, meaning a power facility that does not connect to the public electricity grid. These plants can supply a dedicated data center campus without selling electricity to outside customers.

In July, EPA issued islanded-power guidance stating that such facilities generally fall outside the Acid Rain Program. The program controls sulfur dioxide and nitrogen oxide emissions from covered power generators.

EPA bases that conclusion on statutory and regulatory definitions tied to electricity sales and federal energy reporting. It says the interpretation expands development options while reducing strain on public grids. A facility that later connects to the grid can become subject to the program.

EPN sees a troubling mismatch. A dedicated plant can produce pollution even when its electricity never reaches public customers. The physical effect of emissions does not depend on whether a wire connects the plant to a regional market.

That does not mean islanded facilities escape every environmental requirement. Stationary engines and turbines can remain subject to other federal, state, or local rules. Their obligations depend on equipment, fuel, location, emissions, operating status, and permit design.

Another dispute concerns backup generators. During specific grid emergencies, federal orders can direct large customers to operate their own generators. EPA says qualifying operation under those orders counts as emergency use and does not consume certain non-emergency operating-hour allowances.

The policy supports reliability during extreme demand. However, simultaneous generator use can concentrate diesel emissions in regions with many data centers. Health advocates want regulators to model the combined effect instead of evaluating each engine fleet separately.

EPA has also proposed changing federal minimum requirements for public participation in some minor-source air permits. A minor source falls below a legal threshold for major-source permitting. The label does not mean its pollution is harmless.

That distinction becomes important when a large campus contains separate generators, turbines, phases, or affiliated facilities. Critics argue that fragmented permits can hide the project’s cumulative emissions. Developers respond that permits should reflect legally distinct equipment and realistic operations.

The Data Center Coalition supports faster infrastructure approvals while saying environmental safeguards and acceleration can coexist. That position captures the industry’s strongest answer to EPN. Faster review does not automatically require weaker pollution control.

However, the burden now falls on regulators and developers to show how both goals survive implementation. They need transparent emissions estimates, enforceable operating limits, monitoring data, and clear explanations of cumulative impacts.

Without those elements, “streamlining” becomes difficult to distinguish from reduced scrutiny. With them, agencies can shorten timelines while preserving evidence and accountability. That practical distinction will determine whether EPA’s promise withstands the EPN data center pollution challenge.

Louisiana Shows Why Local Disclosure Matters

Louisiana illustrates how a national infrastructure strategy can leave residents negotiating health, utility, and land-use questions with incomplete information.

The Louisiana Illuminator’s coverage gives the national report a particularly relevant local setting. The state has pursued major data center investment while already carrying a large concentration of petrochemical, refining, and power infrastructure.

Data centers can offer construction work, tax revenue, infrastructure spending, and demand for specialized services. State and local officials often view a large project as a route to long-term economic development. Those benefits help explain the political urgency surrounding approvals.

The employment picture can be less straightforward. Large campuses require substantial labor during construction, but their permanent staffing can be modest relative to their size. Communities must compare those jobs against tax incentives, grid investments, water demands, and environmental burdens.

Disclosure shapes whether that comparison is possible. EPN cites reporting that at least five Louisiana state senators signed nondisclosure agreements connected to one data center project. The local assessor and every member of a public development board involved in negotiations also signed agreements.

Nondisclosure agreements can protect a company’s commercially sensitive site search. They can also prevent land speculation before a project becomes public. Yet broad secrecy can delay public understanding until officials have already committed political and financial support.

That timing matters for pollution review. Residents need project capacity, power sources, generator counts, expected operating hours, water plans, and construction phases. Without those details, public meetings become arguments over trust instead of evaluations of measurable impacts.

Louisiana’s exposure also cannot be assessed through a single statewide average. Existing pollution, health conditions, income, age, access to medical care, and distance from emission sources vary by community. A new project can impose different risks across neighboring areas.

The same facility might draw grid power from sources located elsewhere. Its electricity could extend operation at an existing fossil plant or support construction of new gas generation. Residents near that power source would share the burden without receiving the host community’s tax revenue.

This separation creates an accountability problem for technology companies. A developer can accurately say its servers meet local permit requirements. A utility can accurately say it serves a growing customer base. A state can accurately say the project attracts capital.

None of those statements answers whether the combined system raises regional pollution. Each institution controls only one part of the chain. Effective oversight must connect the campus, dedicated generation, grid dispatch, fuel supply, and affected communities.

Louisiana also shows why public participation is more than a procedural delay. Residents often contribute local knowledge about flooding, existing odors, respiratory conditions, road access, school locations, and previous industrial exposure. Agencies can test those concerns against technical evidence.

Developers have legitimate concerns about unpredictable hearings and shifting requirements. A clear process can address those concerns with published timelines, standardized data requests, and early community engagement. Eliminating information does not make conflict disappear.

Secrecy can instead increase opposition after a project becomes public. Residents who discover a negotiated deal late may challenge permits, subsidies, utility arrangements, or construction approvals. A faster private negotiation can produce a slower public dispute.

For Louisiana, the EPN data center pollution warning is therefore not only about Washington policy. It asks whether state and local institutions have enough information and capacity to fill gaps left by federal changes.

The Health Estimates Are Serious but Not a Project-Level Verdict

The strongest reading of the evidence is neither dismissal nor certainty, but a demand for measurement before exposure becomes an accomplished fact.

EPN’s report combines peer-reviewed methods, government records, outside analyses, media investigations, and advocacy judgments. That makes it useful for identifying a policy pattern. It does not make every cited risk equally certain at every proposed site.

The national health model carries several uncertainties. Future data center electricity demand could exceed or fall below projections. Grid generation could become cleaner, remain similar, or shift toward additional fossil capacity. Equipment use and emissions controls will vary.

Atmospheric models also estimate how pollutants move and transform. Health-response functions then connect exposure changes to population outcomes. These are established tools in regulatory analysis, but their results remain estimates rather than individual medical diagnoses.

The study’s authors note that the national burden remains modest compared with pollution from all American fuel combustion. EPN emphasizes a different comparison. In the high-growth scenario, data-center-related health costs rise 213 percent between 2023 and 2028.

Both observations can be true. The total remains a subset of a larger national problem, while its growth rate warrants attention. Policymakers should not inflate the estimates, but they should not ignore a rapidly expanding source category.

Location also changes the outcome. A data center supplied by low-emission generation presents a different air pollution profile from one served by coal or dedicated gas. A campus with tightly controlled backup generators differs from one using turbines for regular power.

Project-specific permits therefore matter. Regulators need realistic operating scenarios, cumulative emissions modeling, control requirements, monitoring plans, and enforceable limits. They also need to consider nearby sources rather than treating each permit as an isolated stack.

EPN highlights children because developing lungs and nervous systems can be more vulnerable to pollution. The report says 188 schools sit within three miles of proposed gas plants dedicated to data centers. Proximity alone does not establish exposure or illness.

Wind, terrain, stack height, pollution controls, operating schedules, and background air quality all affect exposure. The school count identifies locations needing closer review. It should not be presented as proof that those students have suffered harm.

Industry promises also require careful treatment. Technology companies have become major purchasers of renewable electricity and have supported nuclear, geothermal, and storage projects. Those investments can reduce long-term emissions and add firm capacity.

Yet a contract signed today does not instantly change local hourly generation. New transmission and generation can take years to complete. The near-term gap between server deployment and cleaner supply can keep fossil plants operating longer.

Regulators face a similar credibility test. EPA says flexibility and environmental protection can coexist. That claim needs verification through emissions data, inspection records, enforcement outcomes, and public access to permits.

EPN’s critics should also separate disagreements about policy from disagreements about science. One can support faster AI development while accepting that soot and ozone harm health. The real dispute concerns how much added pollution will occur and which protections are justified.

Likewise, supporting stronger review does not require opposing data centers altogether. Cleaner power, energy efficiency, better siting, strict generator limits, storage, demand flexibility, and transparent monitoring can reduce the burden. The outcome is shaped by design choices.

This is the report’s most important restraint. Pollution is not an unavoidable property of artificial intelligence. It results from specific decisions about energy, equipment, location, permitting, and enforcement.

Three Signals Will Show Whether the Warning Was Justified

The next test will come from regulatory text, operating data, and the actual power plants built for AI demand.

The first signal is EPA’s final approach to minor-source public participation and early construction. Proposed language can change after comments, legal review, and interagency work. The final rules will show whether federal minimum protections meaningfully shrink.

If EPA preserves accessible permit records, cumulative emissions analysis, and effective public comment, the agency strengthens its claim that speed and safeguards can coexist. Removing those elements would reinforce EPN’s argument about approvals advancing before answers.

Legal challenges will also shape implementation. Courts can reject an agency interpretation, narrow it, or allow it to proceed. Litigation involving temporary turbines and dedicated power systems will help define which equipment counts as a stationary pollution source.

The second signal is measured operation of generators and turbines. Permit limits describe what a facility can emit, but monitoring reveals how equipment actually runs. Emergency orders, grid congestion, testing, maintenance, and on-site generation can change annual emissions.

Communities need generator hours, fuel consumption, emissions testing, malfunction reports, and ambient-air measurements. Data should cover entire campuses where practical. Separate filings should not prevent reviewers from understanding combined operations.

Transparent operating data could weaken the EPN data center pollution warning if emissions remain low and controls perform reliably. Repeated emergency use or unexpectedly high emissions would strengthen it, especially in regions with dense data center development.

The third signal is the generation mix serving new computing loads. Announcements alone will not settle the issue. Analysts should track which coal retirements are delayed, which gas plants receive permits, and which cleaner projects reach commercial operation.

The Department of Energy has acknowledged that large new loads are affecting reliability planning. Its emergency-order record shows how federal authorities can intervene when supply becomes tight. Those interventions deserve health analysis alongside reliability analysis.

A cleaner expansion would weaken EPN’s highest-burden scenario. That means new demand is matched with additional low-emission generation, storage, transmission, efficiency, and flexible computing schedules. Annual corporate procurement claims are not enough.

A fossil-heavy expansion would support the report’s central mechanism. Dedicated gas plants, prolonged coal operation, and frequent generator use would connect AI growth to higher pollution. Weaker monitoring would then make the resulting burden harder to quantify.

Technology buyers and AI users should care because computing choices now carry infrastructure consequences. Model size, inference volume, latency expectations, and geographic deployment influence electricity demand. Efficiency improvements can reduce pressure, but rebound growth can offset those gains.

Developers should also expect environmental performance to become part of procurement and reputation. Enterprise customers already ask about security, privacy, and carbon emissions. Local pollution, hourly power sourcing, and community disclosure are likely to join that list.

For policymakers, the immediate task is narrower than resolving every dispute about AI. They can require comparable emissions disclosures, preserve meaningful participation, measure cumulative effects, and enforce permit conditions. Those steps would produce evidence for better decisions.

For readers, the useful question is not whether every data center is harmful. Ask what powers a specific facility, where that power is generated, which pollution controls apply, and whether operating data will be public.

The EPN data center pollution warning will be vindicated or weakened by those answers. Watch the final federal rules, measured generator use, and new generation entering service. Together, they will reveal whether AI expansion is shifting hidden health costs onto communities.