The server farms powering the artificial intelligence boom are warming the land around them by roughly 2 degrees Celsius on average, with the effect reaching towns and residential areas up to 10 kilometers (about 6.2 miles) away, according to a working paper led by University of Cambridge researchers. At the most extreme sites, the temperature spike hit 9 degrees Celsius, a figure that rivals the heat island effect measured in dense urban cores. The authors estimate that more than 340 million people worldwide live within range of existing or planned data center clusters, though the study focuses primarily on U.S. locations where construction has surged since 2020.
The paper, now in its third version as of early 2026, has not yet passed formal peer review. But its reliance on two decades of NASA satellite data and well-established gap-filling techniques gives the findings more weight than a typical preprint, and the results land at a moment when communities from northern Virginia to central Texas are grappling with proposals for massive new facilities.
What the satellite data shows
The Cambridge team built its analysis on land-surface temperature records derived from NASA’s MODIS instruments aboard the Terra and Aqua satellites. A peer-reviewed dataset published in Scientific Data describes a global MODIS product covering 2002 through 2020 at 0.05-degree resolution. That product uses a statistical method called DINEOF, combined with ERA5-Land reanalysis data, to fill gaps caused by cloud cover, a persistent obstacle in satellite temperature measurement.
Because MODIS provides a consistent, long-term record, researchers can control for seasonal cycles and broader climate trends when assessing local changes. By comparing temperature patterns in the years before a data center was built to those after it became operational, the team isolated abrupt shifts that coincided with facility activity. The warming signal appeared most strongly in daytime summer readings, when solar radiation and waste heat from cooling systems reinforce each other.
Reconstruction techniques for clear-sky land-surface temperatures under cloudy conditions have a documented scientific lineage. A methods paper in Computers and Geosciences details approaches that build on earlier work demonstrating high-resolution temperature time series at 250-meter resolution across continental scales. These methods form the technical backbone that makes before-and-after comparisons around data center sites possible at all.
Where the uncertainties lie
Several gaps separate these findings from settled science. The most significant: the MODIS baseline covers 2002 through 2020, but the largest wave of AI-driven data center construction in the United States began after 2020, fueled by demand for generative AI training and inference capacity. That timing mismatch means post-construction temperature comparisons for the newest and largest facilities may rely on a narrower window of satellite data or on extrapolation methods not yet fully detailed in public summaries of the work.
It is also unclear how many facilities were included, how the authors categorized them as “AI clusters” rather than conventional cloud or colocation sites, or how they accounted for nearby industrial development that could independently raise surface temperatures. If the selected sites skew larger or more power-hungry than the broader universe of data centers, the heat estimates may not generalize.
The relationship between land-surface temperature and the air temperature people actually feel adds another layer of complexity. Satellite instruments measure thermal energy radiated by ground, rooftops, and pavement, which can diverge significantly from air temperature at weather stations. A 2-degree Celsius rise in surface temperature does not necessarily translate into the same increase in the air a resident breathes, though it can contribute to localized warming, particularly during summer nights when surfaces retain and re-radiate heat.
No direct statements from U.S. energy regulators, local planning boards, or data center operators appear in the available evidence to confirm or dispute the measured impacts. The 340-million figure is presented as a global order-of-magnitude estimate, not a precise census count, and the paper does not break it down by country or metro area. No population-level health or economic impact studies tied specifically to data center heat islands have been published.
Why the findings are plausible
Even with those caveats, the underlying physics supports the direction of the results. Data centers reject enormous amounts of waste heat, whether through air-cooled systems that vent hot exhaust into the surrounding environment or water-cooled systems that transfer heat to local water bodies. As AI workloads demand more power per rack than traditional cloud computing, the thermal output per square foot of facility space has climbed. That combination of higher power density and rapid build-out in concentrated clusters makes detectable local warming a reasonable expectation, even if the exact magnitude remains under debate.
Independent cross-checks are possible. The European Space Agency operates a separate land-surface temperature program using multi-sensor merged products over long time spans. Researchers outside the Cambridge group could use those records to test whether the same warming signatures appear around the same sites, providing a valuable second opinion on the satellite evidence.
What communities and regulators should watch
The practical takeaway is not that every AI facility will automatically add 2 degrees to local air temperatures. It is that large clusters can measurably alter the thermal landscape around them, and that this effect deserves a seat at the table alongside water use, noise, and grid strain when new projects are evaluated.
Urban planners weighing data center proposals in regions already prone to dangerous summer heat may want to incorporate thermal modeling into environmental impact assessments. Requiring waste-heat mitigation, whether through district heating reuse, reflective building materials, or setback requirements, could blunt the effect. Coordinating siting decisions with broader climate adaptation plans is another step suggested by the emerging evidence.
For researchers, the path forward is straightforward: replicate the Cambridge results with alternative satellite products, extend the analysis beyond the United States, and pair satellite observations with ground-based measurements of air temperature, humidity, and wind near data center clusters. That combination could turn a provocative working paper into a definitive account of how the physical footprint of AI reshapes local climates, giving communities the hard numbers they need before the next wave of construction breaks ground.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
