The AI build-out changed what a data center risks. Rack densities have climbed from a historical 10–15 kW to 40–100+ kW for GPU clusters, and every one of those kilowatts arrives as heat that has to leave the building — historically, by evaporating water. At the same time, sites chosen for cheap power, land and tax treatment cluster in exactly the basins where water is contested: the US Sun Belt, southern Europe, northern India, parts of East Asia. The result is a portfolio of facilities whose core input — cooling — depends on a resource their hosts may not be able to supply through a drought year. This guide sets out the assessment method operators and their ESG teams are converging on: screen the hazards geospatially, stress the thermal design honestly, measure water stress rather than water volume, and extend the analysis upstream to the fabs the whole industry depends on.
The hazard set
| Hazard | Driver | Impact pathway | Financial expression |
|---|---|---|---|
| Extreme heat | Rising ambient temperatures; urban heat islands | Evaporative cooling efficiency falls with wet-bulb temperature; mechanical chillers carry the load | PUE deteriorates; energy costs spike; thermal throttling in the tail |
| Water scarcity & drought | Precipitation shifts; aquifer depletion | Municipal curtailment; industrial water prices rise | Forced reduction in cooling withdrawal; dry-cooling retrofits; tanker logistics in extremis |
| Flood & inundation | Extreme rainfall; riverine and coastal surge | Substations flood; fuel deliveries for backup generators are cut | Long outages; asset damage; insurance repricing (see our stranded-assets guide) |
| Grid stress & wildfire | Peak demand colliding with heat; transmission fire risk | Load shedding; public-safety shutoffs | Generator hours accumulate — Scope 1 diesel, fuel-supply exposure, permit pressure |
The assessment method, in three passes
Pass one — geospatial exposure mapping. Plot every facility against downscaled climate projections (CMIP6 under SSP2-4.5 and SSP5-8.5 are the working pair) at 2030, 2040 and 2050 horizons: temperature and wet-bulb trajectories, flood designations, basin-level water stress. The output is a risk taxonomy — low / medium / high / critical — per asset, not per portfolio. Portfolio averages conceal the only thing that matters, which is which specific buildings sit in which specific basins.
Pass two — thermal and electrical sensitivity. Stress the cooling design against the projected wet-bulb tail (the 28–32°C wet-bulb band is where evaporative systems begin failing to hold ASHRAE environmental envelopes unaided) and calculate the crossover: at what ambient condition does the facility need full mechanical support, what does that do to PUE and power draw, and does the utility contract — and the grid behind it — carry that load on the same afternoon every other customer peaks?
Pass three — redundancy and vulnerability scoring. Grade what the site can survive: N+1 / N+2 depth, on-site fuel autonomy (48–72 hours is the common design band), dual utility feeds, flood barriers and elevation, and — routinely underweighted — the water redundancy: second source, storage volume, reclaimed-water optionality.
Water: measure stress, not just volume
A litre consumed in a wet basin and a litre consumed in a stressed basin are different facts, and volumetric reporting treats them as identical. The working toolkit corrects for that:
- WUE (Water Usage Effectiveness) — annual water use ÷ IT equipment energy (L/kWh). The efficiency baseline. Evaporative fleets run well above 0.2 L/kWh; the design frontier — dry cooling, closed-loop and direct-to-chip liquid cooling — targets below 0.2 and approaches zero on site. Report WUE alongside PUE or not at all: the two trade against each other, and optimizing one in secret is how facilities end up efficient on paper and fragile in fact.
- WRI Aqueduct (and the WWF Water Risk Filter) — basin context: baseline water stress (High at 40–80%, Extremely High above 80%), seasonal variability, groundwater decline, regulatory and reputational exposure. A facility in an Extremely High basin needs a mitigation plan regardless of its WUE.
- ISO 14046 water footprinting — life-cycle assessment of direct and indirect water (including the water embedded in purchased electricity), scarcity-weighted with characterization factors such as AWARE. This is the step that catches the shell game: a “zero-water” site on a coal-heavy grid has moved its water, not eliminated it.
- Source Water Vulnerability Assessment — the local audit: the municipal supplier’s own sustainability, recharge rates, drought allocation hierarchies (who gets cut first, by law), and competing agricultural and residential demand growth.
Disclosure anchors
ESRS E3 (Water and Marine Resources) requires in-scope companies to report water consumption — specifically in water-stressed areas — with watershed context, alongside ESRS E1 for the physical-risk scenario analysis itself. IFRS S2 requires the financial-effects version: what physical risk does to position and cash flows, and what adaptation capital is allocated against it. The TNFD’s LEAP approach (Locate, Evaluate, Assess, Prepare) gives the nature-dependency framing site selection teams increasingly use. And beneath the frameworks sits the layer that binds first in practice: municipal permitting. Local ordinances are progressively restricting potable-water extraction for industrial cooling and pushing facilities toward reclaimed and non-potable feeds — the permit, not the framework, is where water strategy becomes compulsory. (The framework map is in our compliance frameworks overview.)
Upstream: the fab problem
A data center’s water risk does not end at its fence line; its supply chain concentrates in the same hazard. Advanced semiconductor fabrication is ultra-pure-water intensive, and fab capacity concentrates in drought-exposed geographies — Taiwan (whose 2021 drought forced water trucking to fabs) and Arizona among them. A single-point failure there is a server-availability event everywhere. The same interdependency runs through power: the embedded water footprint of regional generation — hydro reliability in drought, thermal-plant cooling limits in heatwaves — is the data center’s availability risk wearing another name. A complete assessment maps the top upstream nodes (fabs, cold-plate and PDU manufacturing, grid regions) through the same hazard screen as the facilities themselves.
The execution roadmap
Phase 1 (months 1–3) — audit. Screen 100% of owned and leased locations through Aqueduct and the downscaled projections; classify every site into the four-band taxonomy; commission Source Water Vulnerability Assessments for the high and critical bands.
Phase 2 (months 4–12) — cooling transition. In stressed basins, mandate the low-water architectures for new high-density deployments — dry cooling, direct-to-chip liquid, rear-door heat exchangers — and retrofit evaporative systems with adiabatic assist to cut peak-season draw. Design decisions made in this phase are 15-year commitments; the basin projections, not this year’s rainfall, are the design basis.
Phase 3 (months 12–18) — circular sourcing. Exit potable municipal water for industrial cooling: reclaimed municipal wastewater, industrial greywater, on-site rainwater harvesting. This is also the phase where allocation politics get managed — a facility that does not compete with residents for drinking water holds a materially stronger position in a drought allocation fight.
Phase 4 (ongoing) — replenishment and transparency. Watershed restoration in operating basins — wetland creation, aquifer recharge, invasive removal — against volumetric replenishment targets. The pledge benchmark is set by the largest operators: Microsoft’s commitment to be water positive by 2030, and Google’s target of replenishing 120% of its freshwater consumption. Pair the projects with public, near-real-time usage reporting; in stressed basins, secrecy about consumption reads as guilt about consumption.
Frequently asked questions
What’s the difference between WUE and PUE — and why report both?
PUE measures energy overhead (total facility energy ÷ IT energy); WUE measures water overhead (annual water ÷ IT energy). Cooling choices trade one against the other: evaporative cooling improves PUE by spending water; dry cooling saves water by spending energy. Reporting only one invites optimizing the reported number at the expense of the hidden one.
Does direct-to-chip liquid cooling eliminate water risk?
It largely eliminates on-site evaporative demand — the loop is closed. It does not eliminate the water embedded in the electricity the facility buys (generation cooling and hydro dependence), nor the upstream fab water in its servers. ISO 14046 footprinting exists precisely to keep that distinction honest.
What water-stress level should trigger a mitigation plan?
On WRI Aqueduct’s baseline water stress scale, High (40–80%) warrants a documented mitigation plan; Extremely High (above 80%) warrants one before further capacity is added. The trigger is the basin classification, not the facility’s own consumption — a small consumer in a critical basin still faces curtailment politics and allocation hierarchies.
Why do hyperscalers publish replenishment pledges — is that operations or PR?
Both, and the operational part is real: replenishment projects (wetlands, recharge) raise the basin’s effective supply for every user including the operator, and the pledges — Microsoft water-positive by 2030, Google at 120% of consumption — are volumetric commitments that can be audited against reported consumption. The PR value collapses quickly if the underlying consumption data isn’t published with them.
Sources: WRI Aqueduct 4.0 (baseline water stress bands); The Green Grid WUE metric definition; ISO 14046 water footprint standard and AWARE characterization factors; ASHRAE thermal environmental envelopes; CMIP6 scenario framework (SSP2-4.5, SSP5-8.5); ESRS E3 and E1; IFRS S2; TNFD LEAP approach; Microsoft (water positive by 2030) and Google (120% replenishment) public commitments. Site-level conclusions require site-level hydrological data; basin classifications are screening instruments, not engineering determinations.