The Atlantic recently examined the growing environmental toll of artificial intelligence data centers, joining a wave of scrutiny over how the industry’s electricity appetite is straining power grids and worsening air pollution. The piece arrives as federal agencies and international bodies have released overlapping assessments that put hard numbers on the problem. U.S. data centers consumed 176 terawatt-hours of electricity in 2023, roughly 4.4% of the nation’s total, and federal projections show that figure could triple by 2028. The question is no longer whether AI is an energy problem but how large the health and climate costs will be.
Federal Numbers Show a Demand Surge
A federal model-based estimate from the U.S. Department of Energy pegged national data center electricity consumption at 176 TWh in 2023, equal to about 4.4% of all U.S. electricity that year. The same DOE analysis projects demand will climb to between 325 and 580 TWh by 2028, which would represent 6.7% to 12% of the country’s electricity supply. The wide range reflects uncertainty about how quickly AI workloads will scale and whether efficiency gains in chip design and cooling can keep pace.
That spread matters because even the low end of the projection nearly doubles current consumption in just five years. At the high end, data centers alone would rival the total electricity use of some mid-sized countries. Most of the growth traces directly to AI training runs and inference workloads, which require far more computing power per query than traditional cloud services. Each new large language model or image generator demands clusters of thousands of GPUs running around the clock, and the companies building them have shown little appetite for slowing down.
Federal planners are also trying to anticipate where this new load will land. Tools such as the Department of Energy’s Genesis platform are being used to explore how new generation projects, transmission corridors, and large industrial customers like data centers might interact on regional grids. That kind of scenario work is becoming essential as utilities warn they may not be able to serve all the requested capacity on current timelines.
Global Demand Could Double by 2030
The International Energy Agency reached similar conclusions from a global vantage point. In its Energy and AI assessment, the IEA projected that worldwide data center electricity demand could exceed approximately 945 TWh by 2030 in the base case, more than double current levels. The agency’s executive summary notes that this expansion is driven heavily by AI-specific computing, from model training to real-time inference embedded in everyday applications.
Demand from AI-optimized data centers specifically is expected to more than quadruple over the same period, according to the IEA’s accompanying news release. IEA Executive Director Fatih Birol has framed the tension bluntly: AI promises to reshape the energy sector, but its own consumption is growing so fast that it risks undermining the grid decarbonization efforts it could theoretically support.
In the United States, data centers are on course to account for a large share of electricity demand growth through 2030, the IEA found. That trajectory collides with a grid that is already strained in many regions. Utilities in Virginia, Texas, and the Midwest have reported that data center interconnection requests now dwarf all other categories of new load. When those facilities plug into grids that still rely heavily on natural gas or coal, every additional megawatt-hour carries a direct emissions cost.
Mapping 2,132 Facilities and Their Emissions
A preprint study posted on arXiv attempted to quantify exactly that cost. Researchers compiled detailed information on 2,132 U.S. data centers covering September 2023 through August 2024, estimating each facility’s electricity consumption, electricity sources, and attributable CO2-equivalent emissions. The paper reports comparisons of emissions intensity and location-specific impacts, highlighting how a data center in a coal-heavy grid region can produce several times the carbon footprint of an identically sized facility powered by hydroelectric or wind resources.
This geographic variability is the crux of a problem that aggregate national statistics tend to obscure. A single number like 176 TWh tells you the total load but nothing about where the pollution lands. The preprint’s facility-level approach reveals that emissions are concentrated in regions where cheap land and favorable tax incentives have attracted clusters of data centers, often the same areas where fossil fuels still dominate the generation mix. The result is a pattern in which the communities least responsible for AI development bear a disproportionate share of the air quality burden.
Because the study spans more than two thousand facilities, it also underscores how quickly AI and cloud infrastructure have become embedded in the broader industrial landscape. Data centers now sit alongside refineries, steel plants, and petrochemical complexes as major point sources of indirect emissions, mediated through the power plants that serve them.
Translating Emissions Into Health Costs
Connecting power plant emissions to public health outcomes requires a specific kind of modeling. The U.S. Environmental Protection Agency maintains a screening tool called COBRA, short for CO-Benefits Risk Assessment, designed to estimate health impacts and monetized benefits from changes in air pollution. COBRA addresses fine particulate matter (PM2.5) and, in newer versions, ozone, two pollutants closely linked to respiratory disease, cardiovascular events, and premature death.
The tool takes inputs of sector-level changes in emissions of PM2.5 and its precursors, including nitrogen oxides (NOx), sulfur dioxide (SO2), and volatile organic compounds, then translates those changes into estimated health effects and dollar-value impacts, as the EPA’s documentation explains. When data center electricity demand rises in a region served by gas- or coal-fired plants, COBRA can model the downstream consequences: additional asthma cases, hospital admissions, lost workdays, and mortality. The dollar figures that emerge from such screening are not precise invoices but rather order-of-magnitude estimates that help policymakers weigh tradeoffs between economic development and public health protection.
No official EPA analysis has yet applied COBRA specifically to AI-driven electricity growth. That gap matters. Without a formal federal assessment linking new data center capacity to health outcomes in specific communities, the debate over siting and permitting proceeds largely without public health data on the table. Local officials may hear promises of jobs and tax revenue but see little quantified information about potential increases in particulate exposure or ozone days.
Why Location Decisions Carry Outsized Weight
Much of the current coverage of AI energy use treats the problem as a simple supply question: build more power plants, add more transmission lines, and let the chips keep spinning. But the emerging research suggests that where data centers are built may be at least as important as how many are built. Locating a new facility in a region with abundant wind, solar, or hydropower can cut its effective emissions dramatically compared with siting the same building in a coal-reliant grid.
Location also shapes who bears the health burden. Many of the cheapest parcels of land with ready access to high-voltage transmission sit on the outskirts of metropolitan areas or in rural counties that already host other industrial facilities. Adding another large, constant electricity load can push nearby fossil plants to run more hours, increasing local pollution even if system-wide emissions appear modest. Without tools that explicitly connect these siting choices to health outcomes, communities are left to argue from principle rather than evidence.
Regulators and planners are not powerless in the face of this trend. They can require more granular impact assessments that incorporate facility-level emissions estimates, prioritize interconnection for projects that pair data centers with new renewable generation, or set conditions that encourage flexible loads to align with clean energy availability. As the DOE and IEA numbers make clear, AI’s electricity demand is on track to reshape power systems. Whether that transformation deepens existing environmental inequities or helps accelerate a cleaner grid will depend on decisions being made now about where to plug the next wave of servers into the wires.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
