Analysis projects Amazon’s data-center gas plants could top the dirtiest U.S. coal plant’s carbon emissions

Two planned gas projects serving Amazon data centers could together produce more carbon pollution than America’s dirtiest coal plant, according to an analysis described by Harvard Chan School on November 5. The finding puts the power supply—not just the computing equipment—at the center of the emissions outlook for the AI and cloud-computing buildout.

Gianluca Guidi, a Harvard Chan postdoctoral research fellow in biostatistics, conducted the analysis for The Washington Post. He used air permits, corporate disclosures, community presentations and other materials to estimate emissions from gas projects associated with data centers. Harvard’s account summarizes findings from a Post article it dates November 2; these are projections for the projects, rather than measurements of pollution already emitted.

Power projects measured against whole companies

The Amazon comparison concerns two gas projects planned to power data centers in Texas and Pennsylvania. Guidi’s estimate puts their combined carbon pollution above that of the country’s dirtiest coal plant, located in Alabama. Against Amazon’s own greenhouse-gas footprint, the projected pollution would be equivalent to a 28.5% increase.

For Meta, the comparison is more stark: gas plants serving a nearly six-square-mile data center under construction in rural Louisiana could produce more carbon emissions than the entire corporation currently generates. The campus is described as consuming as much power as three cities the size of New Orleans—a measure of the computing site’s demand, distinct from the estimate of its associated carbon pollution.

AI may feel completely digital, immaterial, but its infrastructure is very physical. It needs land, power plants, transmission, water and fuel, and these choices can last for decades.

Gianluca Guidi, Harvard Chan postdoctoral research fellow

Different comparisons reveal different kinds of scale

The analysis also examines a Microsoft project in West Texas and a Google project in Wyoming, both involving gas plants. Together, their projected electricity consumption would be five times that of every household in Wyoming.

An OpenAI data center in Pike County, Ohio, provides a separate carbon comparison within the same analysis. Its projected carbon pollution would exceed that of the Microsoft and Google projects combined.

The plants could outlast the boom

Guidi’s warning extends beyond the size of any individual campus. Once built, he expects the gas plants to keep operating for at least a generation, even if the AI boom fades. The distinction is between the life of the computing equipment and the life of the infrastructure supplying it: decisions made to power today’s expansion could shape emissions long after today’s generation of AI hardware becomes obsolete.