Is AI drinking our water?

Sites. Locations, operators, capacities and local impacts come from company announcements, regulators and news reporting. Each claim on a site dossier links to its source. Twenty-six sites have campus coordinates taken from an official address, geocoder, permit or published building outline and checked against Sentinel-2 imagery; for clusters the measurements centre on one named campus, shown on the dossier with its location source. The proposed Cerrillos site marks its announced plot, and the Reliance–Meta Jamnagar site, whose location has not been published, marks the city.

Water bankruptcy screen. Madani (2026) defines water bankruptcy as persistent over-withdrawal from surface and groundwater relative to renewable inflows and safe levels of depletion, and the resulting irreversible or prohibitively costly loss of water-related natural capital. The two-signal screen is this atlas's construction, not part of that definition. The over-withdrawal signal is Aqueduct 4.0 baseline withdrawals at or above the basin's renewable supply (raw water stress ≥ 1); the 40% high-stress level is shown separately because stress is reversible and distinct from bankruptcy. Where Aqueduct's modelled local supply is effectively zero, demand is met by transfers or desalination it does not count, so the site is listed as not assessable. The depletion signal is a GRACE/GRACE-FO stored-water trend whose 95% range stays below zero. GRACE series are deseasonalised by removing each calendar month's mean before annual means are taken (years with at least six months), and the trend's standard error is inflated for lag-1 autocorrelation and tested against the t distribution on the effective degrees of freedom. Lag-1 autocorrelation from about 23 years, across the 2017–2018 gap, is itself uncertain. Volume rates are trend × land area; total volume is trend × years × land area; the endpoint change shown on the evidence board is the 2024–26 mean minus the 2002–04 mean. All ± values are statistical only and exclude GRACE solution, leakage and glacial isostatic adjustment errors. The screen cannot establish irreversibility or cause.

Surface heat. For sites with campus coordinates, Landsat 8 and 9 Collection 2 surface temperature (thermal band, delivered at 30 m) is read for clear summer scenes (June–August, or December–February south of the equator) in 2013–2016 and 2024–2026. In each scene, clear pixels within 1 km of the site are compared with clear pixels 3–6 km away, using the QA band to drop cloud, dilated cloud, cirrus, cloud shadow and water. The same acquisition delivered in two overlapping tiles is counted once, and at sites where work began in 2024 or later the later period starts the year after, so no scene predates construction. Up to 14 scenes per period are used.

Night heat. For the same sites, MODIS/Aqua 8-day night-time land surface temperature (MYD11A2 v6.1, 1 km, overpass about 01:30 local time) is read from the ORNL DAAC for every summer since 2003. In each composite, pixels centred within 1.5 km of the site are compared with pixels 3–6 km away, using only pixels whose quality flags give an LST error of 2 K or less; the MODIS sinusoidal grid is converted back to latitude and longitude before distances are measured. Yearly values are summer means of at least four composites. Up to ten summers before work began are compared with the summers after it by Welch's t-test.

Land conversion. Impact Observatory/Esri 10 m annual land cover (v2, 2017–2023) is read within 3 km of each campus. A pixel counts as newly built only when it is not built-up in both 2017 and 2018 and is built-up in both 2022 and 2023, which removes most single-year classification flicker; its 2017 class shows what it replaced. All construction is counted, not only data centres, and the maps end in 2023.

Reported water use. Site water figures are taken only from published primary or record-based sources: Google's per-data-centre table (Environmental Report 2026), Meta's withdrawal-by-facility table (2025 Environmental Data Index), Amazon's water-withdrawal map (city or state totals that include colocation and domestic water), and utility, municipal or permit records reported by journalists. Values are shown as published and converted to litres; projections, permits and one-time fills are labelled and never treated as measured use. Where a region shows a clear stored-water loss, annual use is expressed as a share of that loss to show scale, not cause. Microsoft, xAI and most operators publish no per-site figures.

3D ground block. The opening view drapes each campus's Sentinel-2 close-up on the Copernicus GLO-30 surface model (30 m, including buildings), with heights exaggerated by the stated factor. The layers below ground are schematic; the water table moves only in the direction of the region's GRACE result.

Ask Droplet. Droplet, the water guide, runs in your browser and answers only from this atlas's data files, naming the source. It sends nothing anywhere.

Beyond the 28. The same water screen is applied to every location in the IM3 Open Source Data Center Atlas (Mongird et al., PNNL, 2025; 1,382 mapped US data-centre buildings and campuses) and to Epoch AI's Frontier Data Centers (CC BY 4.0), whose addresses are located with OpenStreetMap Nominatim at street or, where needed, town level. Sites already among the 28 use the atlas's checked coordinates.

People and aquifers. People come from WorldPop's 2025 global population at 1 km (constrained, adjusted to United Nations totals; 8.20 billion people). Each 5 km block of people takes the stored-water verdict of the 1-degree centre it lies in, which is the GRACE trend for land within 253 km, and the country whose Natural Earth outline it lies in; "people within 253 km" of a place is the sum of WorldPop inside that circle. The 37 largest aquifer systems are the WHYMAP delineation (BGR and UNESCO-IHP), numbered as in Richey et al. (2015); five systems unnamed in the file are named from that list. For each system, the stored-water trend is the cos-latitude weighted mean of the GRACE mascon cells inside its outline, with the same statistics as everywhere else. GRACE weighs all the water over an aquifer (groundwater, soil moisture, lakes, rivers and snow), so the trend is not groundwater alone, and systems smaller than about 200,000 km² are below the size GRACE measures reliably and are marked. People living over an aquifer do not all draw on it, and a region's water also arrives by rivers and canals; the figures show who lives on the water that is changing, not who uses how much.

Country by country. Outside the United States, data-centre locations come from OpenStreetMap features tagged telecom=data_center or building=data_center, downloaded through the Overpass API. Ways and relations are reduced to their centre, and features within 150 m of each other are counted once. Each location is assigned the country of its Aqueduct sub-basin and given the same water screen as above. Only countries with at least three mapped locations are shown. OpenStreetMap is more complete in some countries than others, so counts describe what is mapped, not every data centre that exists; the shares within a country are the more comparable figures.

Where does the water go? For each country with an atlas site, three independent records are compared. On paper: annual freshwater withdrawals and their split by sector, and internal renewable resources, as reported to FAO AQUASTAT (World Bank indicators ER.H2O.FWTL.K3, FWAG, FWIN, FWDM and INTR; latest year). From space: the GRACE/GRACE-FO trend in all water stored on land, averaged over 0.25° land cells inside Natural Earth 1:50m borders (contiguous United States only), with a 95% range adjusted for autocorrelation; countries under 150,000 km² are flagged as too small. Climate: monthly precipitation and land evaporation from NASA's MERRA-2 reanalysis through NASA POWER, sampled on a regular grid across each country and compared between 2002–2011 and 2016–2025. The page states which record explains a loss only by comparing magnitudes; it does not estimate return flows or consumption, which no public national dataset reports. Countries with large mountain glaciers are flagged, because melting ice lowers the satellite total without appearing in rain or evaporation.

Lakes and rivers. Open water in a 16 × 16 km square centred on each site is mapped from Copernicus Sentinel-2 Level-2A scenes (read through the Microsoft Planetary Computer at 40 m grid spacing). A pixel counts as water when both the NDWI (green against near-infrared) and the MNDWI (green against short-wave infrared) are above zero, the Sentinel-2 scene classification also labels it as water (this second test rejects dark roofs, solar panels and terrain shadow that pass the index test alone), and it is not cloud, snow or missing (cloud-shadow flags are ignored because they often fall on dark open water); tiles from the same date are joined, and each pair of months is compared only over pixels clear in both scenes. For each calendar month the three clearest scenes with at least 85% clear pixels are mapped in 2017–2018 and in 2024–2026 and the one with the middle water share is kept (with two, the larger), because a single scene sometimes misses water under algae, haze or glint; only months present in both periods are compared, and the change is tested with a paired t-test on monthly water area. The method is checked against gauged reservoirs in Texas (surface area, Texas Water Development Board) and Arizona (storage, USGS); the results are in the table below. Stable control waters (the Columbia River at The Dalles, the Middenmeer polder lakes, the Quincy canals) are re-checked at every update. Maps show water seen in at least half of the clear scenes of each period. Area says nothing about depth, so no volume is derived. Long-term context comes from the JRC Global Surface Water transitions layer (Pekel et al. 2016), 1984–1999 against 2000–2020.

Statistics. Changes between periods are tested with Welch's t-test on yearly values (climate, vegetation) or per-scene values (heat), and called clear when p < 0.05. Yearly values carry trends and autocorrelation, and scenes within a summer are not independent, so p-values near 0.05 should be read with care. About 150 tests are run across the atlas; at p < 0.05 roughly 7 clear results would be expected by chance alone, so a single clear result at one site is a lead, not a finding.

Water Risk Score. Five parts add up to 100: basin water stress today from WRI Aqueduct 4.0 (35 points), the GRACE/GRACE-FO stored-water trend since 2002 (25), Aqueduct's business-as-usual (SSP3-7.0) stress projection for 2050 (15), the change in rainfall and root-zone soil wetness from NASA POWER (15), and warming (10). Scores of 65 and above are Severe, 45–64 High, 25–44 Elevated and below 25 Low. The weights are a transparent judgement, published so anyone can change them. Aqueduct gives its top stress score to basins that are arid with low water use; any such basin receives full stress points here.

Climate and soil. NASA POWER provides monthly temperature, rainfall and root-zone soil wetness from 1981 on a grid of roughly 50 km. These numbers describe the surrounding region, so most of their change reflects regional climate, not the campus alone.

Vegetation. NDVI is the MODIS MOD13Q1 250 m product. For each year the highest value in the peak growing window (16-day composites starting late June to mid-August in the north, January to mid-February in the south) is averaged over a box of 9 × 9 MODIS pixels (about 2 × 2 km) at the site and at two reference boxes 20 km east and west. Comparing the site with the reference boxes helps separate local land change from regional weather.

Imagery. The wide view uses NASA Harmonized Landsat and Sentinel-2 (HLS) 30 m scenes from NASA GIBS. The close-up uses Copernicus Sentinel-2 Level-2A 10 m scenes from 2016–2017 and 2025–2026, read through the Microsoft Planetary Computer. In both, the clearest scene in each period is kept and its date is shown. All imagery is openly licensed.

Life. GBIF records are counted inside a ~20 × 20 km box for 2001–2010 and 2016–2025. They show observation effort as much as nature.

Future. The scenario model multiplies power, load and published intensities. Grid carbon intensity is each country's latest annual average from Ember. Water to generate power applies the median consumption factors of Macknick et al. (2012) to the country's generation mix, with tower cooling assumed for thermal plants; hydropower reservoir evaporation can be added with a switch because charging it to electricity is contested. On-site cooling water uses the US average unless the operator has reported a closed-loop, air-cooled or seawater design. The temperature line extends the observed trend; it is not a climate-model projection.

Siting comparison. Each site is compared with every city of at least 300,000 people in its country (Natural Earth), using a Basin Water Score built only from the water-supply parts of the Water Risk Score. The nearest cities scoring at least 15 points lower are listed. It answers only the water question; power, fibre, land and law also decide where a data centre can go.

Verification. Every number comes from the public sources listed below, processed as described on this page. The derived results are open: download them below, with the inputs needed to recompute each one. The analysis code is held by the author; research groups and public bodies may request it for verification.