Water Stewardship

For two decades water was a footnote in data-center design. It was cheap, abundant near the cheap-power and cheap-fiber sites operators already wanted, and invisible to the public. The AI build-out ended that. EESI cites upper-bound estimates of up to about 5 million gallons a day for a large campus, the municipal draw of a town of 10,000–50,000 people. Virginia DEQ’s 2026 planning assumption puts a typical 100 MW-IT campus at about 300,000 gallons a day. At either scale, the load lands in exactly the arid, power-rich basins (Texas, the US Southwest, the Gulf, inland Spain) that the power-bound era pushed operators toward. Water has hardened from a siting footnote into a top-three permitting pillar, alongside power and zoning, and into the most visible front in the community fight over whether a campus gets built at all.

Stewardship is the operating, accounting and social-license discipline an operator owns for the life of the asset. The upstream siting gate (is there water, of what seniority, in what basin, under what regulator?) is run in Chapter 3.7. Cooling-plant and cooling-tower engineering lives in Chapter 5.8. Choosing wrong costs you in PUE, in permit risk, in capex, and in social license, which now binds as hard as the interconnection queue. Underneath it all sits the energy–water nexus: the cooling design that saves water spends energy, and the one that saves energy spends water. Stewardship means choosing which to spend, site by site, and accounting for it honestly.

Heat rejection is where energy and water trade against each other, and the trade depends on ambient conditions and the loop temperature you pick. Evaporative (open-loop) cooling, meaning cooling towers and adiabatic assist, uses the latent heat of vaporization. Evaporation carries heat away at the water’s operating temperature, which lets a tower reject heat at a wet-bulb temperature far below the dry-bulb air. That lower wet-bulb sink can cut compressor duty when your chiller and weather bins support it. An open tower consumes most of the water it withdraws, because that water leaves as vapor rather than as discharge. The share is (C−1)/C with negligible drift, or about 83% at six cycles of concentration. The site’s exact fraction follows its blowdown, drift and cycles of concentration. Dry, non-evaporative heat rejection sends heat to outdoor air without consuming process water. The cost is a higher approach temperature, more compressor hours in hot weather, and an annual electrical penalty set by your dry cooler, chiller and weather bins. That penalty concentrates in the summer peak that also stresses the grid.

This choice defines the water-stewardship strategy, and it is irreversible in the same sense the cooling-modality decision is: you plumb and permit a campus for towers or you do not. Choose evaporative and you get the lower wet-bulb sink plus a permanent water-consumption line that a regulator, a journalist and a county board will scrutinize for thirty years. Choose dry, non-evaporative heat rejection and you remove the process-water draw that mobilizes opposition, and you unlock arid, power-rich sites. The price is a standing energy tax on every hot-weather hour for the life of the building, with the water behind your megawatts still on your books. Either way you buy a penalty; stewardship is picking the right one for your basin. One exception is conditional: a warm-water loop rejecting to dry coolers in a temperate climate gets low PUE and near-zero site WUE together. Model the hot-weather power draw and the upstream footprint before treating an arid site as settled.

Water Usage Effectiveness (WUE) is liters of water per kWh of IT energy, and the boundary decides whether that means total input or net consumption. ISO/IEC 30134-9 defines it. Chapter 15.1 sets it against the EU’s total-water-input reporting boundary and the rest of the metric stack. LBNL’s December 2024 report modeled US annual-average site WUE at ~0.36 L of on-site water consumption per IT kWh in 2023, across all cooling types. Designs with dry, non-evaporative heat rejection trend to ~0 WUE-site. Microsoft’s FY2025 cooling-and-humidification WUE was ~0.27 L/kWh for its fully owned and controlled data centers operating for the full year. Separately, its December 9, 2024 zero-evaporation design announcement estimated avoiding more than 125 million liters per data center annually. Microsoft’s first Fairwater facility in Mount Pleasant, Wisconsin, was fully operational on June 23, 2026, with over 90% of its capacity on closed-loop liquid cooling. Its remaining air-cooled servers use water only on the hottest days. The headline number hides three drivers an operator controls.

Driver one: cooling modality. The evaporative-versus-non-evaporative choice above is the dominant lever, and nothing else moves WUE as far. Driver two: cycles of concentration. In an evaporative plant, the same makeup water can be recirculated several times before dissolved-solids buildup forces a blowdown. More cycles cut makeup and blowdown but concentrate salts. Scaling and corrosion depend on chemistry, while Legionella control also depends on temperature, residence time and the water-management program. That is a direct water-versus-chemistry trade, managed in Chapter 5.8. Driver three: water source. WUE counts liters on whatever input or consumption boundary you report, regardless of quality. The stewardship question is not just how much but which water, and source strategy is the most defensible lever you have.

Source is a decision in its own right. Potable water is the easiest source to engineer and the hardest to defend politically. Loudoun County’s data centers drew roughly 900 million gallons of water in 2023. Potable draw is what most reliably rallies community opposition and ratepayer-cost arguments. Reclaimed / non-potable water (treated municipal effluent, industrial grey water) costs more in treatment, conveyance and sometimes a dedicated purple-pipe agreement. But it is increasingly the condition of the permit itself: water-stressed jurisdictions now mandate reclaimed sourcing, and a reclaimed-water commitment is often what converts a contested rezoning into an approved one. Loudoun County reports that Loudoun Water’s reclaimed system supplied roughly 40 customers with more than 750 million gallons in 2025. Harvested / on-site sources (rainwater capture, condensate recovery, treated process water) are marginal in volume but valuable in the disclosure narrative. The downstream cost is asymmetric: potable is cheap to build and expensive to defend, while reclaimed is expensive to build and cheap to defend.

Increasingly the operator builds the reclaimed supply itself. In Spotsylvania County, Virginia, Amazon says it is funding a $480 million upgrade of the Massaponax wastewater plant and more than 20 miles of new pipeline. The county will own the system and can sell reclaimed water to other businesses (Amazon, September 2026). Amazon also says it cools with outside air about 96% of the year there, so the contract must guarantee hot-day peak delivery, not just annual volume. Check what diverting effluent does to downstream river flow before counting it as a stewardship win.

The table is the same kind of cascade the workload archetype drives in Chapter 1.1: the leftmost column is the input you control, and everything to the right is a consequence you inherit for the life of the asset. The shift that matters in 2026 is in the third row. Direct-to-chip liquid cooling, which dense NVL72 designs require (Chapter 5.4), is often confused with water consumption because it is 'water cooling', yet it consumes almost none. The technology-cooling loop is a sealed, recirculating circuit, filled once and topped up rarely. What consumes water is the heat-rejection stage at the back of that loop, and a liquid plant with dry final heat rejection consumes essentially none. So the dense-liquid future is, if anything, easier to make water-neutral than the air-cooled past: the heat is already in a contained loop, and you choose how to reject it.

An evaporative plant is not a closed system. Water flows in as makeup and out through evaporation plus blowdown. As pure water evaporates, dissolved minerals concentrate in the recirculating loop. At some threshold, the cycles of concentration, the loop must dump a fraction to drain (the blowdown) and replace it with fresh makeup. Blowdown is a regulated discharge. It carries elevated total dissolved solids, conditioning chemicals, biocides and corrosion inhibitors. In the US it falls under Clean Water Act NPDES (Section 402) permitting if discharged to surface water, or under a local pretreatment or sewer-use agreement if sent to a POTW. Underestimating blowdown chemistry is how a campus that cleared its withdrawal permit gets blindsided on the discharge side.

The other side of cycling water harder is biology. A cooling tower is a warm, aerated, nutrient-bearing aerosol generator, textbook habitat for Legionella pneumophila, the pathogen behind Legionnaires’ disease. An outbreak traced to a data-center tower is a reputational and legal event of a different order than a permit dispute. ASHRAE Standard 188 supplies the building-water risk framework. When adopted by law or contract, it requires a written Water Management Program built on a risk assessment of every device where water contacts air, with defined control limits (temperature, biocide residual, conductivity), routine monitoring, validation, and documented corrective action. ASHRAE Guideline 12 supplies the supporting risk practice. In many jurisdictions an ASHRAE-188-conformant program is the de-facto standard of care, and its absence is what plaintiffs and regulators point to after an incident.

'Water-positive' has become the headline sustainability commitment of the hyperscaler era. Microsoft reports that in FY2025, for the first time, its projects replenished more water globally (over 14.2 million m³) than it withdrew. It says global volume is not enough and that its next phase is increasingly local. That is a milestone toward its 2030 water-positive commitment, which requires sustaining the balance. Google has committed to replenish 120% of the freshwater it consumes by 2030. In 2025 its projects replenished about 7.7 billion gallons, roughly 78% of that year’s freshwater consumption, and it projects more than 19 billion gallons a year by 2030 once the projects are fully implemented (designs as of March 2026). In June 2026 Google added a pledge for each site. By 2030 the site should replenish more water than it consumes, and where a source is at high risk, Google will choose air cooling or recycled water.

Meta committed in August 2021 to be water positive by 2030. Its fleet cools mainly with direct evaporative cooling on outside air, and its December 2024 restoration approach targets 200% of consumption in high-stress basins and 100% in medium-stress basins. Restoration projects returned more than 1.6 billion gallons in 2024. These are real programs, and they are also the part of the water story a regulator or a skeptical community is most likely to pierce, because the accounting hides three distinctions that decide whether the claim means anything.

Distinction one: on-site versus total footprint. The water you withdraw and consume at the campus, the direct or on-site water, is the visible number. But the electricity that powers the campus is itself produced with water, since thermoelectric and hydroelectric generation evaporate enormous volumes. So the total footprint includes a large indirect electricity-supply and value-chain water term behind every kWh. The water-footprint convention borrowed from greenhouse-gas accounting calls on-site water scope 1, the water consumed generating the purchased electricity scope 2, and supply-chain water such as chip manufacturing scope 3 (Li et al., Making AI Less Thirsty). The indirect terms dominate at sector scale. The IEA estimates that data centers worldwide consumed about 560 billion liters of water in 2023, most of it behind the electricity, and projects up to about 1,200 billion liters a year by 2030 (Energy and AI, April 2025). A site that runs zero process water on-site can still carry a substantial water footprint through its grid mix. Honest disclosure reports the on-site figure alongside the indirect terms. A water-positive claim that leaves out the energy-embodied term is incomplete by construction.

Distinction two: replenishment versus reduction. 'Water-positive' is usually a volumetric claim. The operator funds watershed projects (wetland restoration, irrigation efficiency, leak repair, aquifer recharge) that return more water to the basin than the campus consumed. That is legitimate and valuable, but it is offsetting, not reduction, and offsets are only as good as their additionality, permanence and, critically, their locality. Replenishing a watershed in a wet basin does nothing for a stressed basin where the campus actually draws. The credible programs target same-basin replenishment. The weak ones average across geographies in a way that lets a thirsty site hide behind a wet one.

Distinction three: basin-level risk and social license. Water is the most local of all the sustainability variables. Carbon is global and fungible, and a ton avoided anywhere counts. Water is not: the only basin that matters for a community’s drought, its wells and its ratepayers is their basin. That is why water dominates the social-license fight even when the absolute volumes are modest against agriculture or municipal use. It is also why source and total-footprint disclosure, same-basin replenishment and a reclaimed-water commitment are worth more to a project’s license than any global metric. The disclosure frameworks that formalize this, CSRD/ESRS, ISSB and the ISO 30134 WUE standard, are treated in Chapter 15.7.

Trace the hot-day boundary. At six cycles, blowdown = 300/(6 − 1) = 60 m³, and tower makeup = 300 + 60 = 360 m³. Total input = 360 reclaimed + 15 potable = 375 m³/day. Discharge = 60 + 15 = 75 m³/day, so net consumption = 375 − 75 = 300 m³/day. Input-WUE = 375/240 ≈ 1.6 L/kWh, and consumption intensity = 300/240 ≈ 1.3 L/kWh. Normal operation fits the assumed reclaimed entitlement with 15 m³/day spare. Chemistry, source and discharge approvals still have to be in hand before operation.

Flip. At three cycles, blowdown = 300/(3 − 1) = 150 m³ and makeup = 450 m³, so the source is short by 75 m³/day. With the original source limit, the crossover is cycles = 375/(375 − 300) = 5. Below five cycles, the source cannot carry the full wet duty. In drought at six cycles, maximum evaporation = 300 × (6 − 1)/6 = 250 m³/day, about 17% below the requested duty. Reduce wet heat rejection, run a dry backup sized for the shortfall, or curtail load; replenishment credits cannot supply the missing water. Chapter 5.8 converts the permitted evaporation into heat-rejection duty.

Do not average the WUEs. Annual total input = 300 × 90 + 65 × 375 = 51,375 m³, and annual IT energy = 300 × 120 + 65 × 240 = 51,600 MWh. Their ratio is about 1.0 L/kWh. The day-weighted mean, [300 × (90/120) + 65 × (375/240)]/365, is only 0.89 L/kWh because the hot days carry more IT energy. The balance follows the DOE cooling-tower mass balance, and Chapter 15.1 sets the metric boundaries. Choose the source for its drought entitlement and chemistry, then carry its energy penalty into the annual bins.

A water-stewardship program that survives a contested rezoning, a CSRD audit and a drought year is built from a small number of decisions made early and documented honestly:

• Run the basin-risk screen before contracting power. Water-stress mapping, water-rights seniority and basin security under drought are siting gates. In April 2026 more than 60% of the Lower 48 was in drought (D1–D4), the highest for that time of year in the Drought Monitor’s record (AP, April 18, 2026). A refused water-service path or rezoning, as in Tucson 2025 and Pinal County 2026, can strand a campus however good the power deal was. → Chapter 3.7.
• Make the evaporate-versus-design-out call explicitly, per basin. Screen dry, non-evaporative heat rejection against summer electrical headroom in stressed basins. Keep evaporative cooling only where drought supply, treatment and discharge pass and the energy saving relieves the binding constraint.
• Commit to the most defensible source you can engineer. Use reclaimed/non-potable water where the drought delivery, treatment, conveyance and fallback case passes. Treat a potable-only design in a stressed basin as a permitting liability to eliminate.
• Own the ASHRAE-188 Water Management Program and the blowdown/discharge chain as one system. Optimize cycles, chemistry, Legionella control and NPDES/ZLD strategy jointly, under one owner.
• Disclose source-split, on-site WUE, and total/embodied footprint. Disclosure is becoming law. Under California’s AB 2619 (signed 21 September 2026), a data center must estimate its water use and name the source before its first business license is issued, then report actual use and cooling type at every renewal. From 1 January 2027, Virginia water suppliers must report deliveries of potable and reclaimed water to qualifying data centers separately (Chapter 3.7). If you claim water-positive, back it with same-basin, additional, permanent replenishment whose volumetric benefit a third party can verify.