Research project · Columbia University · Northern Virginia · updated November 2026

Can wetlands cool the water footprint of AI data centers?

Northern Virginia is the world's largest data center cluster, and its facilities used about 1.85 billion gallons of water in 2023. I study whether nearby wetlands lower the heat that drives that water use, using satellite data and state water records.

Sole author Remote sensing Econometrics Working paper, preliminary

Key results so far

−3.3 °C

Wetlands run cooler in summer

Daytime surface temperature near Occoquan Bay wetland vs. Data Center Alley, 72 months (p < 0.001).

+43,600 gal

Heat raises water use

Extra monthly groundwater withdrawal per +1 °C at one data center in the pilot model (p = 0.002).

−1.2 M gal

Cooling technology matters

Monthly drop in withdrawal after the facility's 2023 cooling retrofit (p < 0.001).

Monthly daytime surface temperature 2019 to 2024: the urban data center area runs hotter than the wetland every summer and the two meet in winter.
Monthly surface temperature, wetland vs. data center core, 2019 to 2024.
Seasonal wetland minus urban temperature: winter −0.06, spring −2.88, summer −3.29, fall −1.67 °C, with 95% confidence intervals.
The cooling gap is largest in spring and summer, when cooling demand peaks.
Monthly groundwater withdrawal rises with air temperature, before and after the April 2023 cooling retrofit.
Water withdrawal rises with temperature in both cooling regimes.
Temperature coefficient under five model specifications; significant with season controls, not significant with month-of-year fixed effects.
My own robustness check: the temperature effect weakens once month-of-year effects are controlled, which motivates the multi-facility study.

Scope of these results

The pilot compares one wetland with one urban area and one data center, so the results are suggestive, not final. The next phase scales the study to hundreds of wetlands and every reporting facility in the region.

What I did

Research design and analysis

  • Framed a two-part study linking an ecological signal to an infrastructure outcome
  • ANOVA with Tukey HSD; OLS with structural breaks, fixed effects and HAC errors
  • Stress-tested my own findings and redesigned the study around the weak points

Data and tools

  • Google Earth Engine: MODIS, Landsat, Sentinel-2, ERA5-Land, National Wetlands Inventory
  • Virginia DEQ withdrawal records, NOAA climate data, public-records requests
  • Python (pandas, statsmodels, matplotlib), JavaScript, Excel

Next steps

Pilot study and working draft

Two-site temperature comparison and single-facility water model

Done

Earth Engine pipeline

Hundreds of wetland patches at 30 m, day and night temperature 2015 to 2025, wet-bulb weather at each facility

In progress

Multi-facility water panel

Monthly withdrawals for every reporting facility in Northern Virginia

In progress

Regional exposure model

Wetland cooling, wet-bulb temperature and water use across the data center fleet, with uncertainty

Planned

Journal submission

Peer-reviewed article

Planned

Code for the figures and the Earth Engine pipeline is on GitHub. The manuscript and data are available on request.