A working paper published by the National Bureau of Economic Research (NBER) puts a concrete figure on what has long been treated as an externality: the public health and environmental damage caused by the electricity consumption of data centers across the United States. The study, authored by Carnegie Mellon economist Nicholas Muller, analyzed 2,800 facilities and concluded that the power demands of AI infrastructure generate an estimated $25 billion in annual costs to the economy — costs borne not by the companies operating the servers, but by the communities living near the power plants that feed them.
The paper arrives at a moment when the data center industry is expanding at an unprecedented pace. Hyperscale operators — the large cloud and AI companies that dominate the sector — have announced wave after wave of new construction, driven by the computational appetite of large language models and generative AI workloads. That expansion translates directly into electricity demand, which in turn translates into emissions, depending on the fuel mix of the local grid.
The mechanics of an externality
The core methodology of the NBER paper relies on a well-established framework in environmental economics: the social cost of carbon and the monetization of health impacts from air pollution. When a data center draws power from a grid that relies partly on natural gas or coal, the generation process produces pollutants — fine particulate matter, sulfur dioxide, nitrogen oxides — that are linked to respiratory disease, cardiovascular illness, and premature mortality. Muller's analysis maps the location of data center facilities against the fuel profiles of the grids that serve them, then applies damage functions to estimate the downstream costs.
The $25 billion figure encompasses both climate-related damages (via the social cost of carbon) and direct health costs, including increased mortality risk for populations living in proximity to fossil-fuel power plants. The distinction matters: climate costs are diffuse and global, while health costs are concentrated and local. A data center built in a region powered largely by renewables imposes a fraction of the damage of one connected to a coal-heavy grid. Geography, in this framework, is destiny.
This is not an abstract accounting exercise. The social cost of carbon is already embedded in U.S. federal regulatory analysis, and health-impact assessments are standard tools in environmental policy. What the paper does is apply those tools specifically to an industry that has, until recently, escaped granular scrutiny of its environmental footprint. The tech sector has historically emphasized its own operational efficiency — power usage effectiveness ratios, renewable energy procurement — while the upstream emissions from grid electricity have received less attention.
Who bears the cost
The political and regulatory implications are significant. Data centers are often welcomed by local governments for the tax revenue and jobs they bring, but the health costs identified in the study fall on the same communities. That asymmetry — private gains for operators, socialized costs for residents — is the textbook definition of a negative externality, and it tends to attract regulatory attention once it is quantified.
Several jurisdictions have already begun to push back on data center development, citing strain on local power grids, water consumption for cooling, and noise. The NBER paper adds a new dimension to that resistance by attaching a dollar figure to health outcomes. Whether that figure shifts the policy conversation depends on how state and federal regulators weigh economic development against environmental and public health costs — a tension that has no clean resolution.
It is also worth noting what the study does not resolve. The damage estimate is sensitive to assumptions about grid composition, and grids are changing. Utilities across the country are adding renewable capacity, and some hyperscale operators are signing long-term power purchase agreements for clean energy. If the grid decarbonizes faster than data center demand grows, the health costs could shrink. If demand outpaces the energy transition — a scenario many grid operators consider plausible — the costs could rise.
The $25 billion figure, then, is not a fixed verdict but a snapshot of a system in motion. The forces pulling in opposite directions — accelerating AI demand on one side, grid decarbonization on the other — will determine whether the hidden toll of data center emissions grows or diminishes. The question is which force moves faster, and who pays in the interim.
With reporting from Fortune.
Source · Fortune
