Local thermal externalities of hyperscale datacenters: causal satellite evidence and implied cooling costs

Jorge Colazo

Hyperscale datacenter construction converts agricultural or greenfield land into continuous heat rejection campuses, yet no causal estimate of the resulting local thermal effect exists. This paper applies a Callaway and Sant'Anna (2021) staggered difference-in-differences estimator to spatially differenced Landsat satellite thermal imagery across 59 datacenter sites in the United States, with treatment onset dated independently of the thermal outcome using Visible Infrared Imaging Radiometer Suite (VIIRS) nighttime radiance. The VIIRS dating is validated against 17 independently documented commissioning dates (median absolute error of one month; 94 percent within three months). The estimated change in the daytime land-surface temperature (LST) gradient relative to the 5 to 10 km reference ring following energization peaks at +0.80 K in the 500 m to 1 km band (significant at 1 percent), with +0.38 K within 500 m (SE = 0.19, significant at 10 percent) and +0.24 K at 2 to 5 km. The sign is stable across leave-one-cohort-out specifications and seasonal subsamples, although magnitude and precision vary, and the estimate strengthens when datacenter cluster markets are excluded. Translating the thermal gradient into implied annual residential cooling expenditures under alternative persistence and LST-to-air conversion assumptions produces a sample-wide central estimate of $0.50 million per year across 46 operational sites, with joint scenario bounds between $0.17 and $0.88 million. This is a detectable but modest channel relative to facility scale, equivalent to roughly $52 per megawatt-year in the subsample with documented capacity.