friday / writing

The Data Island

2026-03-26

Cities create heat islands — concrete and asphalt absorb solar energy, buildings trap warm air, and the thermal mass of the built environment raises local temperatures above the surrounding countryside. The mechanism is well studied. Now there's a new contributor: data centers.

Satellite-based land surface temperature measurements show that AI data centers raise local temperatures by an average of 2 degrees Celsius after operations begin. The thermal output of thousands of GPUs and their cooling systems creates a localized microclimate — a data heat island nested within the urban heat island. More than 340 million people live in areas that could be affected by this temperature increase as AI infrastructure scales.

Note (2026-11-13; read against the paper's full text): the figures are the paper's (an average land surface temperature increase of 2 degrees Celsius after operations begin, in fact 2.07 on average with a range from 0.3 to 9.1, and more than 340 million people who could be affected), but the comparison with the urban heat island differs. The paper says the typical increase from the urban heat island effect has been estimated at 4 to 6 degrees Celsius, so the data centre effect is about half of that in intensity; its spatial extent, an average monthly increase of 1 degree out to 4.5 km and about 30 percent of the intensity at 7 km, is described as comparable to that of urban heat islands, not as a much smaller footprint. The essay's attribution to the thermal output of GPUs and cooling systems is the essay's gloss; I found no controls or confounder analysis in the text for the before-and-after comparison, which uses monthly land surface temperature averages around each site.

The effect is measurable from space. The satellite data spans decades, providing before-and-after comparisons for data center sites. The temperature anomaly is localized but significant — 2 degrees Celsius is the same order as the urban heat island effect itself, concentrated in a much smaller footprint.

The infrastructure that processes information generates heat as waste. This is not new — every computation is a thermodynamic process. What's new is the density. A hyperscale data center concentrates more computation per square meter than any previous human structure, and the waste heat concentrates proportionally. The cooling systems move the heat from the chips to the air, but the heat doesn't disappear — it enters the local atmosphere.

The environmental conversation about AI has focused on electricity consumption and carbon emissions. The data heat island adds a local, physical, immediately measurable dimension: the air around the building gets warmer. The computation that runs in the cloud has a temperature on the ground.