Scope and source compliance
This note targets three linked questions: why AI data centers consume water, which permits govern that consumption and discharge, and where the investment exposure sits. Source discipline used here:
- Tier 1 technical/IP/performance claims: peer-reviewed journals with DOI-verified identifiers, granted USPTO patents, EPA regulator documents.
- Tier 2 adoption/deployment context: company sustainability reports, company 10-K filings, named regulator proceedings.
- Prohibited: SEO content farms, anonymous benchmarks, AI summaries, unretrievable secondary quotations.
A key distinction applied throughout: a granted patent is evidence of recorded claims, not proof of deployed performance or commercial viability. A peer-reviewed paper survives independent review; a preprint does not. No arXiv preprints are used here because journal versions are available.
Technical system: why AI data centers use water
Data center heat rejection is the core physical driver. Servers and AI accelerators convert electricity into heat; if the heat is not removed, silicon throttles or fails. Cooling designs determine whether water is used, consumed, recycled, or avoided.
Common cooling architectures:
- Air-cooled / dry-cooled systems use fans and outside air; they consume little or no water on site but can use more electricity, which creates indirect water demand at the generating plant.
- Evaporative / cooling-tower systems use water evaporation to reject heat efficiently; they consume water on site and create blowdown wastewater.
- Direct-to-chip liquid cooling circulates a working fluid through cold plates attached to processors; it reduces evaporative losses but may require water for heat rejection from the secondary loop.
- Immersion cooling submerges hardware in dielectric fluid; it is less water-intensive on site but adds fluid handling, leasing, and disposal concerns.
- Body-of-water heat rejection uses a lake, river, or ocean as a heat sink. Microsoft’s granted patent US 9,521,784 B2, titled “Underwater data center,” granted 2016-12-13, describes a submersible data center using ambient water as a heat sink (US9521784B2, grant date 2016-12-13, accessed 2026-04-26). The patent is a granted patent, not an application, so it can support a claim that the technical approach was recorded and assigned; it does not establish deployment, commercial viability, or water-permitting outcomes.
Peer-reviewed work confirms that cooling choices are only part of the water story. Siddik et al. 2021, Environmental footprint of data centers in the United States, Environmental Research Letters, DOI 10.1088/1748-9326/abfba1 (published 2021, accessed 2026-04-26), finds that the water footprint of U.S. data centers extends well beyond the data-center building: electricity generation, especially thermoelectric generation, embeds water consumption upstream. That finding matters because an “air-cooled” data center can still impose substantial water demand through the electricity it consumes.
Mytton, Data centre water consumption, npj Clean Water 3, Article 11, DOI 10.1038/s41545-020-0058-3 (published 2020, accessed 2026-04-26), reviews data center water-use drivers and concludes that water use is highly site-dependent and frequently under-reported when only direct on-site water is counted.
The AI-specific tailwind is power density. A traditional cloud rack may run 5–15 kW per rack; AI training clusters can push 30–100+ kW per rack. Higher power density per square foot raises heat-rejection density, which pushes operators toward liquid cooling and larger cooling-water/wastewater systems. This is a source-bound statement to the extent that Nvidia’s 10-K reports that its data center segment revenue is driven by demand for accelerators used in AI and accelerated computing; the rack-density escalation is the engineering consequence of accelerator deployment (NVIDIA 10-K via SEC, accessed 2026-04-26).
1.1 Water withdrawal vs. consumption
Permitting and company disclosures often blur two terms:
- Water withdrawal is water removed from a source; a power plant or cooling tower may withdraw large volumes and return most of it at higher temperature.
- Water consumption is water evaporated or incorporated into product and not returned to the same watershed.
For data center cooling, evaporation is the main consumption pathway. Blowdown is a wastewater discharge and is regulated separately. This distinction is operational, not rhetorical: a facility can have high withdrawal and low consumption, or low withdrawal and high consumption. EPA’s National Pollutant Discharge Elimination System (NPDES) program governs point-source discharges, including cooling water blowdown, when the discharge reaches waters of the United States (EPA NPDES, accessed 2026-04-26).
Quantitative worked calculation
A useful back-of-envelope calculation for a hypothetical 100 MW AI data center campus:
Formula
\[
W = P_{IT} \times 1000 \frac{kW}{MW} \times U \times H \times WUE
\]
where:
- \(W\) = annual water consumption, liters/year
- \(P_{IT}\) = IT load, MW
- \(U\) = annual utilization fraction
- \(H\) = hours per year
- \(WUE\) = water usage effectiveness, liters of water consumed per kWh of IT energy
Inputs and calculation
- \(P_{IT} = 100\ MW\)
- \(U = 0.80\)
- \(H = 8,760\ h/yr\)
- \(WUE = 1.8\ L/kWh\)
\[
Energy = 100\ MW \times 1000 \times 0.80 \times 8,760 = 700,800,000\ kWh/yr
\]
\[
W = 700,800,000 \times 1.8 = 1,261,440,000\ L/yr
\]
Converting to U.S. gallons:
\[
1,261,440,000\ L \div 3.78541 = 333,200,000\ gallons/yr
\]
Sensitivity
| Assumption | WUE | Annual water |
|---|---|---|
| Dry cooling / efficient climate | 0.5 L/kWh | 350.4 million liters ≈ 92.6 million gallons |
| Midpoint | 1.8 L/kWh | 1.261 billion liters ≈ 333 million gallons |
| Evaporative, hot climate | 2.5 L/kWh | 1.752 billion liters ≈ 463 million gallons |
Parameter table
| Parameter | Value | Source | Tier | measured/derived/assumed/inferred | cross-verified |
|---|---|---|---|---|---|
| IT campus load | 100 MW | Modeling assumption | n/a | assumed | n/a |
| Utilization | 0.80 | Modeling assumption | n/a | assumed | n/a |
| Hours/year | 8,760 h | Physical constant | n/a | measured | n/a |
| WUE midpoint | 1.8 L/kWh | Published WUE ranges in Siddik et al. 2021, DOI 10.1088/1748-9326/abfba1; Mytton 2020, DOI 10.1038/s41545-020-0058-3 | Tier 1 | derived | yes |
| WUE low | 0.5 L/kWh | Peer-reviewed range [Siddik et al. 2021; Mytton 2020] | Tier 1 | inferred | yes |
| WUE high | 2.5 L/kWh | Peer-reviewed range [Siddik et al. 2021; Mytton 2020] | Tier 1 | inferred | yes |
| Conversion | 1 U.S. gallon = 3.78541 L | NIST/standard conversion | Tier 1 | measured | yes |
Source-bound caveat: The 1.8 L/kWh midpoint is a modeling choice inside the published peer-reviewed range, not a fleet-wide audited measurement. The result illustrates scale; it is not a forecast for any specific facility. [SINGLE-SOURCE for midpoint] because the midpoint is a derived assumption, not a directly measured average from two independent institutions.
The permitting stack
There is no single federal “AI data center water permit.” Instead, the relevant permissions stack across federal, state, and local authorities.
3.1 Clean Water Act NPDES
Cooling tower blowdown, boiler blowdown, and construction dewatering discharges are point-source discharges. If they reach waters of the U.S., they require an NPDES permit or equivalent state-issued permit. The permit imposes effluent limits, monitoring, and reporting. EPA’s NPDES program page is the primary regulatory source (EPA NPDES, accessed 2026-04-26).
3.2 Clean Water Act Section 404
Data center campuses often grade land, fill wetlands, or alter streams. Discharge of dredged or fill material into waters of the U.S. requires a Section 404 permit from the Army Corps of Engineers, with EPA oversight and water-quality certification by states/tribes under Section 401. The permit is a schedule risk because wetland delineations, mitigation, and public interest reviews take time (EPA CWA 404, accessed 2026-04-26).
3.3 Cooling water intake structures
For facilities that withdraw surface water for cooling, Clean Water Act Section 316(b) regulates cooling water intake structures to reduce fish impingement and entrainment. A data center that withdraws water directly from a river or lake for cooling should evaluate whether state 316(b)-equivalent requirements or federal permitting thresholds apply. The EPA cooling-water-intakes page is the primary regulator source (EPA Cooling Water Intakes, accessed 2026-04-26). Separately, the power plants that supply electricity to data centers are canonical 316(b) subjects; this is the main channel by which data center electricity growth creates water-intake permitting tension even when the data center itself does not withdraw surface water.
3.4 State water rights and public water supply
In Western states, surface and groundwater withdrawals are governed by prior appropriation or correlative rights, with permits required from state agencies. A data center cannot simply buy land and start withdrawing; it must identify a vested water right, lease or purchase water rights, or contract with a municipal utility. USGS water-use data and state water-rights records are the underlying factual base for assessing regional water stress (USGS Water Use in the United States, accessed 2026-04-26). In regulated municipal systems, a new hyperscale campus may trigger system-development charges, water-system adequacy studies, and rate-base investment by the serving utility.
3.5 Climate and electricity-grid intersection
NERC’s 2024 Long-Term Reliability Assessment identifies large loads, including data centers, as a driver of resource-adequacy stress and notes that load additions may occur faster than generation and transmission additions (NERC 2024 LTRA, published 2024-12-17, accessed 2026-04-26). This matters for water because thermoelectric generation remains a major water user, so data center electricity demand indirectly drives power-plant water withdrawal and discharge permits under the CWA.
Company disclosures vs. 10-K risk factors: a contradiction
Google’s 2024 Environmental Report discloses that its global data center fleet consumed 6.1 billion gallons of water in 2023 (Google 2024 Environmental Report, published 2024, accessed 2026-04-26). That is a self-reported sustainability disclosure, not an audited SEC financial metric.
Microsoft markets a “water positive by 2030” commitment and reports water replenishment projects in its Environmental Sustainability Report (Microsoft 2024 Environmental Sustainability Report, accessed 2026-04-26).
But the same companies’ SEC filings use more cautionary language. Microsoft’s Form 10-K for fiscal year ended June 30, 2024 lists the availability and quality of utility services, including water, among operational risks that could disrupt operations and increase costs (Microsoft 10-K via SEC, accessed 2026-04-26). Alphabet’s Form 10-K similarly discusses the operational dependence of data centers on utilities and environmental compliance (Alphabet 10-K via SEC, accessed 2026-04-26). Amazon’s 10-K includes environmental regulatory compliance risk for AWS data centers (Amazon 10-K via SEC, accessed 2026-04-26). Meta’s 10-K includes utility and infrastructure dependence for its data centers (Meta 10-K via SEC, accessed 2026-04-26).
Contradiction: Marketing and sustainability documents emphasize restoration and “water positive” goals; 10-K risk sections treat water as a scarcity/liability risk. These are not necessarily inconsistent—one is a future pledge, the other a current risk—but the disclosure asymmetry matters. Sustainability reports are selected metrics with company-chosen boundaries; SEC risk factors are mandatory and cautionary. Investors should not convert a “water positive” pledge into an assumption that water risk is neutralized.
Unit economics / business model
The named technology is not a single product but a stack: water rights, treatment, cooling equipment, discharge compliance, and permitting labor. Monetization differs by layer.
- Hyperscalers and AI cloud providers monetize AI compute via cloud subscriptions or API consumption. Water is an operational expense, a schedule risk, and a community/permit risk, not a revenue line. They capture value if AI demand scales faster than water-related project delays. Company risk disclosures support this cost/risk framing (Microsoft 10-K, Alphabet 10-K).
- Water and wastewater utilities monetize via regulated volumetric charges, fixed service charges, connection charges, and authorized capital recovery in rate cases. A new data center can add large load to a water system, but the utility’s return depends on the regulator-approved rate base and authorized ROE. Exact authorized ROE varies by jurisdiction and filing; needs primary source validation for any single citation.
- Cooling/thermal equipment suppliers monetize equipment sales plus service contracts. Vertiv’s 10-K describes thermal management and cooling solutions for data centers, including liquid cooling, as part of its data center offering (Vertiv 10-K via SEC, accessed 2026-04-26). Trane Technologies supplies commercial HVAC and data center cooling; its filings discuss data center cooling demand (Trane 10-K via SEC, accessed 2026-04-26).
- Water equipment/treatment suppliers monetize pumps, membranes, valves, analyzers, and water-reuse systems. Xylem’s 10-K discusses water infrastructure and treatment demand (Xylem 10-K via SEC, accessed 2026-04-26).
- Engineering and permitting consultants monetize study hours, permit application packages, biological assessments, wetland mitigation design, and water-rights due diligence. Their economics are embedded in developer soft costs; no public standard margin is cited here because a verifiable public figure was not retrieved.
Where value is captured: In a water-constrained site, the party that controls a usable water right, a discharge permit, or the equipment that reduces consumptive use captures scarcity value. The data center operator captures AI compute economics only if permitting timelines hold.
Catalyst map
Catalyst researched first:
- subject: NERC 2024 Long-Term Reliability Assessment flags data-center large-load growth and resource-adequacy stress, tightening the electricity-water-permitting link
- event_type: policy_decision
- event_date: 2024-12-17
- source_url: nerc.com
Additional distinct dated events for future second-pass work:
- subject: Google 2024 Environmental Report releases 2023 data center water consumption figure
- event_type: financial_disclosure
- event_date: 2024-07-30
- source_url: sustainability.google
- subject: FERC technical conference on large loads co-located with large generators, relevant to data center siting and grid-water-permit sequencing
- event_type: policy_decision
- event_date: 2024-11-01
- source_url: ferc.gov
- subject: EPA finalizes National Primary Drinking Water Regulation for PFAS, affecting water treatment and industrial pretreatment obligations
- event_type: regulatory_decision
- event_date: 2024-04-10
- source_url: epa.gov
Disconfirming Evidence
The thesis that water infrastructure and environmental permitting are a growing bottleneck for AI data center buildout would be weakened or disproven if:
- Dry/closed-loop cooling reaches mass deployment at data center scale. If adiabatic dry coolers or liquid immersion systems eliminate consumptive on-site water use and do not raise electricity-driven indirect water use, water scarcity loses relevance for siting.
- Permission timelines compress. If state and federal agencies adopt categorical exclusions, pre-approved water permits, or blanket general permits for data centers, the permitting bottleneck disappears.
- Utility water supply expands faster than expected. If municipalities, desalination providers, or water reuse projects add abundant supply without material rate increases, water scarcity no longer constrains campus deployment.
- Company-reported water intensity falls faster than AI load growth. If hyperscalers disclose declining liters-per-kWh or liters-per-inference at a rate that offsets compute growth, the aggregate water constraint weakens.
- Site selection shifts to water-abundant regions with fast permitting. If AI data center investment decisively migrates to areas with surplus water rights and streamlined wetlands/stormwater approvals, water stress ceases to be a portfolio-level material risk.