The footprint does not stop at the site

A US data centre can add electricity demand in one region while the resulting air-quality burden falls elsewhere. A study published on 19 November in Environmental Research Letters models how that happens through the interconnected power grid and the atmosphere.

The MIT-led team, including Yuang Chen and Noelle Selin, examines grid-supplied data-centre electricity demand in 2023. It estimates about 650 attributable premature deaths across the contiguous United States: 513 associated with fine particles and 137 with ozone. These are population-level estimates derived from modelled exposure, not a list of deaths observed and individually traced to facilities.

Following the extra demand

The researchers compare hourly electricity generation in scenarios with and without data-centre demand. A chemical transport model then estimates how the changed emissions form and move through the atmosphere. Published relationships between long-term pollution exposure and mortality translate those differences into health estimates.

The results show a geographic mismatch. States hosting large loads do not necessarily have the highest estimated mortality rates. The authors identify the Midwest as particularly affected by the combination of power-generation emissions and atmospheric conditions that encourage fine-particle formation. Proximity to a data centre alone is therefore an incomplete guide to this grid-mediated burden.

A bounded estimate, not an AI death toll

The analysis covers data centres broadly; it does not separate electricity used for AI from other computing. It is also a 2023 baseline, not a forecast of present-day or future deaths.

The earlier author manuscript assumes constant hourly data-centre demand and excludes on-site backup generators. Its stated uncertainty ranges cover the exposure-response relationships, not all uncertainty in emissions and atmospheric modelling. Model Current has not rerun the analysis. The useful planning question is wider than which community hosts a facility: which generators respond to its demand, and where their pollution goes.

Sources

  1. MIT News: Data centres and distant air-quality effects19 November account: linked grid and atmospheric modelling, 2023 estimates and the mismatch between facility locations and affected regions.
  2. MIT CS3: Research record and published abstractAuthors, journal DOI 10.1088/1748-9326/aebbd4, counterfactual hourly dispatch method, 513 particulate-related and 137 ozone-related estimated deaths, and spatial scope.
  3. Chen and colleagues: Earlier author manuscriptFull 46-page author PDF linked by MIT’s research record, read locally. Earlier submitted version, not the final publisher PDF. Methods sections 2.1–2.4 and discussion checked: constant hourly load, excluded on-site backup generation, and uncertainty limits.