THIRSTY AI OPEN DATA, version 1.0.0 (2026-10-22)
Kaushik, P. (2026). ORCID 0000-0002-0767-3983. Licence: CC BY 4.0
(country-locations.csv: ODbL 1.0, as it is derived from OpenStreetMap).
Status: not yet peer reviewed. Corrections are welcome and are logged by version.

FILES
  aquifers.csv                 37 rows
  countries.csv                68 rows
  country-locations.csv      3636 rows
  data-centre-shares.csv       12 rows
  frontier-ai-sites.csv        76 rows
  gauge-check.csv            2559 rows
  lakes-rivers.csv             28 rows
  people-on-falling-water.csv    146 rows
  reported-water.csv           31 rows
  sites.csv                    28 rows
  stored-water-by-country.csv    145 rows
  stored-water-by-year.csv    644 rows
  us-states.csv                47 rows
  water-budget.csv             16 rows
  world-stored-water-trend.csv  30714 rows
  CITATION.cff, zenodo.json   citation metadata

WHAT EACH INDICATOR MEASURES
Stored water (sites.csv grace_*): NASA GRACE/GRACE-FO CSR mascons, mean over land
  within 253 km of a site, April 2002 to the latest month. Linear trend with a 95%
  range adjusted for autocorrelation. A regional signal: it cannot isolate one site.
Basin water stress: WRI Aqueduct 4.0 baseline, withdrawals divided by renewable
  supply in the sub-basin. Modelled, not measured.
Water screen (water_screen columns): "overuse" when Aqueduct withdrawals exceed
  renewable supply (ratio >= 1); "depletion" when the GRACE trend is clearly
  below zero; "both" when both hold; "unassessable" when local renewable supply
  is about zero. A screen of where data centres sit, after Madani (2026). It does
  not measure a data centre's own use or effect.
Daytime heat: Landsat 8-9 surface temperature, campus (<=1.5 km) minus land
  3-6 km away in the same scene, summer scenes before construction and in the
  latest summers. Welch's t-test.
Night heat: NASA MODIS Aqua MYD11A2 1 km night surface temperature, same design.
Land built over: Impact Observatory 10 m annual land cover within 3 km; new
  built land = not built in 2017-2018 and built in 2022-2023. All construction.
Reported water: figures published by operators, utilities or authorities, with
  the kind of figure (measured, permitted, projected, aggregate) and its link.
Country locations: United States from the IM3 Open Source Data Center Atlas;
  elsewhere OpenStreetMap features tagged telecom=data_center or
  building=data_center, centres merged within 150 m. Coverage differs by
  country; compare shares, not counts.

Stored water by year (stored-water-by-year.csv): the annual GRACE anomalies
  behind each site's trend, so the trend can be recomputed.
Gauge check (gauge-check.csv): Sentinel-2 water area of gauged reservoirs
  against the gauge on the same day: surface area for Texas reservoirs (Texas
  Water Development Board) and storage for Arizona reservoirs (USGS).
  failed_scene marks scenes left out of the statistics because they show
  under 60% of the area of the other scenes within 30 days.
World stored-water trend (world-stored-water-trend.csv): the GRACE trend for
  land within 253 km of every 1-degree land centre, as used by "Your place".
  "uneven_loss" means an average loss of at least 1 cm a year whose 95% range
  still reaches zero.
Stored water by country (stored-water-by-country.csv): the mean of those
  1-degree centres inside each country, each weighted by the cosine of its
  latitude, with the same mean of the yearly values (cm_<year>, centimetres of
  water compared with the average over the whole record). share_falling is the share of
  the country's centres where the fall is clear. Countries with fewer than three
  centres are left out. mostly_ice marks Greenland, Iceland, Norway, Antarctica
  and Svalbard, where the fall is mostly melting ice. Because each centre covers
  land within 253 km, small countries also carry their neighbours' signal.

Aquifers (aquifers.csv): the 37 largest aquifer systems (WHYMAP, BGR and
  UNESCO-IHP; numbered as in Richey et al. 2015, Water Resources Research 51).
  people_2025 is WorldPop 2025 (1 km, constrained, UN-adjusted, CC BY 4.0)
  inside each outline; the trend is GRACE over the land inside it, with the
  same statistics as the sites. GRACE weighs all water, not groundwater alone;
  below_grace_size marks systems under about 200,000 km2.
People on falling water (people-on-falling-water.csv): WorldPop 2025 people
  living where the stored-water trend of land within 253 km is clearly falling
  (95% range below zero) or falling unevenly (at least 1 cm a year on average,
  range reaching zero), for the world and each country.
Data-centre shares (data-centre-shares.csv): each mapped data centre (IM3,
  Epoch AI, OpenStreetMap) takes the verdict of its 1-degree land centre. The
  95% range resamples 5-degree blocks of data centres 2,000 times, so a cluster
  of neighbouring sites counts once. Land area excludes Antarctica and Greenland
  and weights each centre by the cosine of its latitude. Basin stress is the
  Aqueduct 4.0 sub-basin polygon that contains each point (categories 3 and 4).

LIMITS
1. Correlation, not cause. Regional water loss is driven mainly by irrigation,
   cities and climate.
2. GRACE resolution is a few hundred kilometres.
3. About 150 statistical tests are run; a few clear results are expected by chance.
4. Most operators publish no site-level water figures.
5. OpenStreetMap and IM3 do not list every data centre.

HOW TO CHECK A RESULT
Each row carries the inputs needed to recompute it: periods, sample sizes,
p-values and 95% ranges. Source datasets are public and listed on the site
under Method and in ATTRIBUTIONS.md.
