Water Stress Index for data centers (WSI-DC)
How much water stress a data center faces at its location, today and by 2050, from 0 to 100.
What it is not: A measure of any facility's water use: that depends on its cooling technology (air-cooled / closed-loop sites use little water, evaporative cooling a lot). The index describes the location.
In plain words
- Each data center gets a score from 0 to 100 that describes the water situation where it stands, not how much water it uses.
- Seven things are combined: scarcity today, scarcity expected in 2050, the worst months of the year, drought and groundwater, how much cooling the climate demands, how many residents share the supply, and water disputes already documented nearby.
- A higher score means scarcer water. From 60 the index is called high.
- An air-cooled site in a dry place can use very little water: the score says where water is a constraint, not who is using it.
- A country or hub figure is the median of its data centers. With fewer than 10 of them, one site can move it.
Formula
Index = Σ wᵢ · pillarᵢ / Σ wᵢ over available pillars (at least 3 required). Pressure = the same over the six physical pillars (no social pillar). “High” = index ≥ 60.
Pillars
| Pillar | Weight | Inputs | Normalisation (0–100) | Source |
|---|---|---|---|---|
| Scarcity | 25 % | Baseline water stress, baseline water depletion | mean, each score ÷ 5 × 100 (Aqueduct scores are 0–5) | WRI Aqueduct 4.0 (CC BY 4.0) |
| Future (2050) | 15 % | Water stress 2050, business-as-usual scenario | score ÷ 5 × 100 (Aqueduct scores are 0–5) | WRI Aqueduct 4.0 future projections (CMIP6) |
| Seasonality & peaks | 15 % | Worst-month water stress, number of months at high stress (score ≥ 3), seasonal variability | mean of: peak score ÷ 5 × 100 (Aqueduct scores are 0–5); months ÷ 12 × 100; variability score ÷ 5 × 100 (Aqueduct scores are 0–5) | WRI Aqueduct 4.0 monthly baseline |
| Drought & groundwater | 10 % | Drought risk, interannual variability, groundwater table decline (where modelled, ~19 % of sites) | mean of available, each score ÷ 5 × 100 (Aqueduct scores are 0–5) | WRI Aqueduct 4.0 |
| Cooling demand | 15 % | Cooling degree days (base 18 °C) today (2015–2024) and in 2050; 99th-percentile daily wet-bulb temperature | linear: 0 CDD → 0, 3,500 CDD → 100, clipped for both CDD; linear: 15 °C → 0, 28 °C → 100, clipped for wet-bulb; mean | NASA POWER daily (MERRA-2 cell, ~50 km); 2050 = POWER daily + K × (WorldClim CMIP6 9-model median 2041–2060 − WorldClim 1970–2000), K = 0.475 (remaining warming under SSP2-4.5). CC BY 4.0 |
| Local competition | 10 % | Residents within 5 km | logarithmic: 1,000 residents → 0, 2,000,000 residents → 100, clipped | JRC GHSL GHS-POP 2020 (CC BY 4.0) |
| Social & regulatory | 10 % | Aqueduct reputational risk, electric-power-weighted overall water risk, documented water conflicts within 50 km (L; recent events weigh 1, older 0.5) | mean of: score ÷ 5 × 100 (Aqueduct scores are 0–5); score ÷ 5 × 100 (Aqueduct scores are 0–5); 100 × (1 − e^(−L)) | WRI Aqueduct 4.0, Conflict atlas |
Coverage (sites with the pillar): Scarcity 8,330 · Future (2050) 8,330 · Seasonality & peaks 8,330 · Drought & groundwater 8,336 · Cooling demand 8,357 · Local competition 8,357 · Social & regulatory 8,357.
How the counts fit together: the inventory holds 8,460 facilities, of which 6,790 are data centers (the others are network points of presence or unclassified). The index is computed at the 8,357 facilities located precisely enough; its maps, rankings and medians use data centers only, 6,687 of the 6,790. The rest have a town-level location or fall outside a source's coverage.
Validation
ROC-AUC of the pressure score (no social pillar) at ranking data centers within 25 km of a documented contestation whose issues include water above all other data centers. Below: a one-pillar baseline, a drop-one-pillar ablation and a spatially blocked check. They are descriptive: they never change the weights or the scores.
ROC-AUC 0.662 (95 % bootstrap interval 0.627–0.7), 167 positives among 6,687 data centers (0.5 = chance, 1 = perfect).
Baseline: Aqueduct baseline water stress + depletion only (scarcity pillar) scores 0.678. The single pillar does as well as or better than the composite on this test. We publish that rather than tuning weights to win it: the test rewards locations where many facilities cluster near a reported event (press coverage follows clusters), while the index is meant to describe several independent pressures. Read the index as a structured profile, not a better predictor of press-reported conflict.
Drop one pillar at a time
We recompute the pressure score without one pillar (the other weights grow to fill its share) and run the same test. A negative Δ means the AUC falls without the pillar, so the pillar helps on this test. A positive Δ means the score does better on this test without it. The full score has ROC-AUC 0.662.
| Pillar dropped | Weight in pressure | ROC-AUC without it | 95 % interval | Δ vs full | 95 % interval of Δ | Pillar alone |
|---|---|---|---|---|---|---|
| Scarcity | 28 % | 0.634 | 0.601–0.672 | −0.027 | −0.036 to −0.019 | 0.677 |
| Future (2050) | 17 % | 0.653 | 0.618–0.690 | −0.009 | −0.014 to −0.005 | 0.685 |
| Seasonality & peaks | 17 % | 0.656 | 0.621–0.695 | −0.006 | −0.009 to −0.003 | 0.678 |
| Drought & groundwater | 11 % | 0.659 | 0.624–0.699 | −0.003 | −0.005 to 0.000 | 0.647 |
| Cooling demand | 17 % | 0.675 | 0.637–0.716 | +0.013 | +0.006 to +0.019 | 0.387 |
| Local competition | 11 % | 0.669 | 0.634–0.706 | +0.007 | +0.003 to +0.012 | 0.482 |
Dropping Scarcity, Future (2050) and Seasonality & peaks lowers the AUC, so these pillars carry the signal on this test. Dropping Drought & groundwater makes no clear difference (the interval of Δ includes 0). Without Cooling demand and Local competition the score does slightly better on this test. We don't re-tune the weights from this table. Few water conflicts are documented, so the whole test is fragile (see the spatial check below). Cooling demand and local competition describe how exposed a water-cooled facility is, which a press-reported dispute doesn't measure directly, so they stay in the index even when they don't help this test.
Each pillar of the validated score is dropped in turn; the remaining weights are renormalised (as for a site missing that pillar) and the rounded composite is scored against the same labels. Paired bootstrap, 500 resamples of sites, seed 7: 95 % intervals of the ablated AUC and of Δ = ablated − full. Alone = AUC of that pillar on its own.
Spatial check
Data centers and press-reported conflicts both cluster in a few places. A single worldwide AUC can therefore partly reflect which regions have reported conflicts, rather than how well the score ranks sites. We grouped the 6,687 data centers into 5° × 5° grid cells: 336 cells, of which 17 contain both a site near a reported conflict and a site that isn't. Those cells hold 166 of the 167 positives.
- Within-cell ROC-AUC 0.613 (worldwide: 0.662). Each site near a conflict is compared only with sites in the same cell. It is lower inside cells: part of the worldwide figure comes from telling regions apart. Only 1 % of the comparisons fall inside a single cell, so this figure rests on the few hubs where most reported conflicts sit.
- Resampling whole cells instead of single sites, the 95 % interval widens to 0.490–0.804 (from 0.627–0.700). It includes 0.5: once spatial clustering is allowed for, this test can't rule out chance.
- Other cell sizes: 2.5° cells (620): within-cell 0.633, cell-resampled interval 0.487–0.812; 10° cells (159): within-cell 0.621, cell-resampled interval 0.497–0.814. One lower bound sits within 0.02 of 0.5, so the conclusion is borderline rather than independent of the cell size.
- Cell by cell (10 cells with at least 3 sites of each kind): median AUC 0.518, middle half 0.124–0.675, above 0.5 in 5 of 10. The result varies a lot from one area to another.
| World region | Data centers | Positives | ROC-AUC inside the region | ROC-AUC with the region left out |
|---|---|---|---|---|
| Asia Pacific | 1,178 | 16 | 0.590 | 0.684 |
| Europe | 2,424 | 59 | 0.642 | 0.700 |
| North America | 2,365 | 44 | 0.611 | 0.685 |
| South America | 459 | 48 | 0.924 | 0.567 |
Leaving out South America lowers the worldwide AUC to 0.567, so much of the result comes from there. Regions with fewer than 10 positives or 10 negatives are not shown.
Sites are grouped in 5° × 5° latitude/longitude cells. The within-cell AUC compares a positive only with negatives in the same cell (Σ U / Σ pairs, pooled over cells), so differences between regions cannot count. The cell-by-cell distribution covers cells with ≥ 3 positives and ≥ 3 negatives. The block-bootstrap interval resamples whole cells (500 draws, seed 7). Regions: the atlas's world regions (continents) with ≥ 10 positives and ≥ 10 negatives, scored inside the region and with it held out. The weights are fixed, not fitted, so nothing is trained on the held-out blocks.
Why these weights
Expert judgement, fixed before validation: present scarcity carries the most weight (25 %) because it is measured, not projected; the 2050 projection, the worst-month signal and cooling demand (15 % each) capture what makes a site fragile for a water-cooled facility; drought/groundwater, local competition and the social signal (10 % each) complete it. Changing weights is a new version.
Limits
- Few water conflicts are documented yet, the validation is indicative and will tighten as the atlas grows.
- Aqueduct scores describe the sub-basin, not the site: a city on imported or desalinated water (e.g. Singapore) can score low on scarcity.
- Groundwater decline exists only where Aqueduct models aquifers.
- No mitigation pillar yet (air/dry cooling, disclosed WUE): planned with operator disclosures.
Derived indicator of the water pressure a data center faces at its location. It does not measure any facility's actual water use, which depends on its cooling technology.
Reproduce & cite
Computed by the atlas's indicator pipeline (version 0.2, run 2026-10-12) from the inputs listed above; every input's source, licence and coverage is in the indicator register. Weights are fixed per version; a change of weights or pillars bumps the version. What changed between versions is listed in the release notes.
Datacenter Atlas (2026). Water Stress Index for data centers (WSI-DC), v0.2. https://www.datacenter-atlas.com/water · CC BY 4.0