A new preprint from researchers at Cornell University estimates that AI data centers raise nearby land surface temperatures by an average of about 3.6 degrees Fahrenheit once they begin operating, with extreme cases reaching as high as 16 degrees Fahrenheit. The findings, drawn from nearly two decades of satellite data, suggest that more than 340 million people worldwide live close enough to these facilities to feel the thermal effects. As the AI industry drives a historic surge in data center construction, the study introduces a concept it calls the “data heat island” and frames it as an emerging environmental cost that planners and policymakers have largely ignored.
What the Satellite Data Actually Shows
The research team analyzed land surface temperature (LST) readings captured by NASA’s MODIS instruments aboard the Aqua satellite, covering the period from 2002 to 2020. Rather than relying on raw satellite passes, which are often obscured by cloud cover, the researchers used a reconstructed dataset that fills gaps caused by clouds using a mathematical technique called DINEOF combined with ERA5-Land reanalysis data. That reconstruction method, published as a peer-reviewed paper in Nature Scientific Data, allowed the team to build a spatiotemporally continuous temperature record spanning nearly two decades and covering the entire globe.
By comparing surface temperatures in areas surrounding data centers before and after those facilities became operational, the researchers calculated an average LST increase of roughly 2 degrees Celsius, or about 3.6 degrees Fahrenheit. The maximum warming detected at individual sites reached approximately 16 degrees Fahrenheit, though the preprint’s abstract does not specify whether that figure represents a single outlier location or a broader subset of high-emission facilities. The full text of the paper, hosted on arXiv, has not yet undergone formal peer review, so the details of site selection, baseline periods, and statistical controls are still open to scrutiny.
Why Waste Heat From Servers Creates Local Hot Spots
Data centers concentrate thousands of servers in a single building, and every watt of electricity those servers consume is eventually released as heat. Cooling systems, which can account for a significant share of a facility’s total energy use, expel that thermal energy into the surrounding air and ground. In dense clusters or arid regions where natural heat dissipation is slow, the cumulative effect can measurably alter local surface temperatures in much the same way that pavement, rooftops, and vehicle exhaust create urban heat islands in cities.
The “data heat island” framing in this study is distinct because it isolates a single industrial category rather than measuring the combined effect of all urban infrastructure. That specificity matters: it means the warming signal the researchers detected is attributable to the data center itself and not to broader urbanization trends. If the methodology holds up under peer review, it would offer regulators a way to quantify the thermal footprint of individual facilities and compare them against zoning or environmental standards. It could also inform siting decisions, such as avoiding locations next to vulnerable neighborhoods or critical infrastructure that is already stressed by heat.
More Than 340 Million People in the Thermal Shadow
The preprint estimates that more than 340 million people live within the zones affected by data center heat output. That population figure, drawn from the study’s spatial analysis, reflects the global spread of large-scale computing facilities across North America, Europe, and Asia. The estimate does not distinguish between people who experience the full 16-degree maximum and those exposed to the smaller average increase, which limits how precisely the health or comfort burden can be assessed from the abstract alone.
Still, even the average 3.6-degree increase carries real consequences. Research on urban heat islands has consistently linked small sustained temperature increases to higher rates of heat-related illness, greater energy demand for air conditioning, and reduced outdoor labor productivity. If data center heat islands produce similar effects, communities near these facilities could face compounding thermal stress, especially during summer heat waves when background temperatures are already elevated. For low-income residents or those without reliable access to cooling, even modest additional warming can translate into higher health risks and energy bills.
How the Study Used NASA Earth Observation Tools
The core temperature measurements came from NASA’s MODIS/Aqua product, a standard dataset that records land surface temperature and emissivity at regular intervals from orbit. NASA distributes this data through its Earthdata portal, where researchers can search, subset, and download global observations for detailed analysis. Visualization tools such as Worldview and other GIS platforms then allow those measurements to be mapped against infrastructure like data centers, power plants, and urban areas.
The choice of satellite-based LST rather than ground-station air temperature readings is both a strength and a limitation. Satellites provide consistent global coverage and avoid the patchy distribution of weather stations, but surface temperature can diverge from the air temperature that people actually experience. The reconstructed dataset addresses cloud contamination, yet any reconstruction introduces modeling assumptions that could affect results near the margins. Independent ground-truth measurements at data center sites (using local meteorological sensors or thermal cameras) would strengthen the findings considerably and help clarify how much of the LST signal translates into lived heat exposure.
A Preprint, Not a Settled Verdict
The study was posted on arXiv, the open-access preprint repository hosted by Cornell University and supported by a consortium of institutional members. Preprints allow researchers to share findings quickly and invite community feedback, but they have not passed through the formal peer-review process that journals require. That distinction matters here because the 16-degree maximum figure, in particular, needs scrutiny: without access to the full methodology, it is unclear whether that reading reflects a persistent warming pattern or a short duration spike at a single facility under unusual conditions.
The average 3.6-degree finding is easier to evaluate because it aggregates across many sites and time periods, reducing the influence of outliers. But even that number depends on the accuracy of the reconstructed LST dataset and the researchers’ method for isolating data center heat from other land use changes. Independent replication using different temperature products or direct sensor deployments would go a long way toward confirming or narrowing the range. As with other work hosted on arXiv, readers can consult its guidance materials to better understand how preprints fit into the broader scientific publishing ecosystem and what caveats apply when interpreting early-stage results.
What Current Coverage Gets Wrong
Early reporting on this study has tended to lead with the 16-degree figure, which is the most dramatic number but also the least representative of typical conditions. The average increase of 3.6 degrees Fahrenheit is the more policy-relevant finding because it reflects the typical warming effect across many facilities rather than a single extreme case. Focusing on the outlier risks overstating the threat in most locations while potentially undermining public trust if later peer-reviewed work revises that maximum downward.
Another common misstep is to treat the preprint as a definitive verdict on the safety or unsafety of data centers. The work is better understood as an early warning signal that a rapidly expanding industry carries a localized heat burden that has not been fully accounted for in permitting or climate planning. That signal should prompt more targeted measurements and modeling, rather than immediate blanket restrictions. It should also encourage technology companies to invest in mitigation strategies such as improved waste-heat recovery, district heating integration, or siting in cooler regions where added warmth has less impact on public health.
Where the Research Goes Next
As AI and cloud computing continue to scale, the concept of a data heat island is likely to become more central to debates over infrastructure and climate adaptation. Future studies could refine the global picture by distinguishing between different cooling technologies, building designs, and climate zones, or by examining how data center clusters interact with existing urban heat islands. Policymakers may eventually need to incorporate thermal impact assessments into environmental reviews, alongside more familiar metrics like water use and carbon emissions.
For now, the Cornell preprint underscores that the environmental footprint of AI is not limited to distant power plants or abstract carbon ledgers. It is also etched into the temperatures of the neighborhoods and landscapes that host the physical machinery of the digital world. Sustaining open, transparent research on those impacts will depend in part on community-backed infrastructure, including services like arXiv’s funding model, and on continued public access to high-quality Earth observation data. As more evidence accumulates, the challenge will be turning these emerging insights about data-driven heat into concrete standards that protect people while still enabling technological progress.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
