Water Usage and Cooling Efficiency in AI Compute

This article surveys the direct and indirect water-consumption mechanisms, the regulatory framing, and the operational practices that the foundational program adopts.

Direct water consumption

The dominant direct water-consumption mechanism is evaporative cooling. The most efficient air-cooling systems use cooling towers in which water is evaporated to dissipate heat to the atmosphere; the evaporated water is “consumed” in the sense that it leaves the data center as water vapor and is no longer available for downstream use. A typical evaporatively-cooled data center consumes 1 to 3 liters of water per kilowatt-hour of IT-equipment energy.

For an AI training run that consumes 100 megawatt-hours of IT-equipment energy, the direct water consumption is in the range of 100,000 to 300,000 liters — comparable to the annual residential water use of a small number of households. For a high-traffic AI inference service that consumes thousands of megawatt-hours per year, the direct water consumption can be in the millions of liters per year.

Public research has quantified the water footprint of specific frontier-model training runs. The Li et al. analysis estimated that training GPT-3 in Microsoft’s U.S. data centers consumed approximately 700,000 liters of clean fresh water — and that a similarly-sized training in Asian data centers would have consumed three times as much because of less-efficient cooling.¹

The alternative cooling architectures have different water profiles. Air cooling without evaporation has near-zero direct water consumption but lower energy efficiency. Direct-to-chip liquid cooling has low direct water consumption (the coolant is a closed loop) and high energy efficiency. Immersion cooling is similar to direct-to-chip from a water-consumption perspective. Adiabatic cooling uses water evaporation only when ambient temperatures exceed a threshold, dramatically reducing water consumption in most operating conditions.

Indirect water consumption

The upstream electricity-generation system also consumes water. Thermoelectric power plants (coal, natural gas, nuclear) use large quantities of cooling water — typically 1 to 5 liters per kilowatt-hour generated. Hydroelectric generation has its own evaporation losses. Wind and solar generation have minimal water footprints. The grid-mix water intensity therefore varies dramatically across regions: a fossil-heavy grid produces high indirect water consumption per kilowatt-hour delivered, while a wind-and-solar-heavy grid produces near-zero indirect consumption.

The combined water footprint — direct facility water plus indirect upstream water — is the figure that the most rigorous water-accounting frameworks report. For a coal-grid evaporatively-cooled facility, the total water footprint can be 5-10 liters per kilowatt-hour; for a wind-grid liquid-cooled facility, the total can be under 0.5 liters per kilowatt-hour.

Locality matters

Unlike greenhouse-gas emissions — which mix into the global atmosphere — water consumption is a local impact. A data center drawing 100 million liters per year from an aquifer in a water-stressed region produces a different (and more consequential) impact than a data center drawing the same volume from an aquifer in a water-abundant region. Several jurisdictions have begun to introduce data-center water-permit limits, drought-response curtailment requirements, and disclosure requirements that focus on local water-stress context.

The McKinsey State of AI surveys have begun to surface water as a procurement-decision factor for enterprise AI buyers, particularly for buyers operating in water-stressed regions or with public sustainability commitments that include water targets.²

Regulatory framing

The European Union Corporate Sustainability Reporting Directive (CSRD) and the European Sustainability Reporting Standards (ESRS) include water-and-marine-resources (ESRS E3) as a material disclosure category. Organizations within scope of the CSRD must disclose water consumption in operations, water stress in operating regions, and water-management strategies — and AI-related water consumption is increasingly recognized as a material subcategory.³

The EU AI Act Article 95 voluntary code of conduct on sustainability is expected to encourage providers of general-purpose AI models to disclose water consumption in their training and serving facilities, alongside the energy and emissions disclosures.⁴

The Greenhouse Gas Protocol does not directly cover water but the broader water-accounting frameworks (the Water Footprint Network’s standard, the Alliance for Water Stewardship standard) provide the methodology that the integrated reporting can use.

Maturity Indicators

The COMPEL D19 maturity rubric does not separate water consumption from energy and carbon at the per-level indicators, but the rubric’s broader framing of “monitoring, measuring, and minimizing the environmental footprint of AI systems including energy consumption, carbon emissions, and resource usage” makes water an in-scope category at every level above Foundational.⁵ An organization at Level 3 (Defined) on water is reporting per-system water consumption alongside per-system energy and carbon. An organization at Level 4 (Advanced) is integrating water into the same dashboards as energy and carbon and is including water-stressed-region procurement criteria in its facility-selection practices.

The Stanford Foundation Model Transparency Index (FMTI) compute-layer scoring is beginning to include water disclosure as a measured indicator, which is creating procurement-decision visibility for water in the same way it has for energy.⁶

Practical Application

A foundational practitioner who is integrating water into the AI sustainability program should produce four artifacts.

Artifact 1: the per-facility water-intensity table. A table that, for each facility hosting AI workloads, records the direct water consumption per kilowatt-hour of IT-equipment energy and the indirect water consumption derived from the local grid mix. The table is the input to all per-workload water calculations.

Artifact 2: the water-stress overlay. A geographic overlay that, for each facility, records the local water-stress level (typically using the World Resources Institute’s Aqueduct water-risk atlas or equivalent). Facilities in high-water-stress regions are flagged for additional governance attention.

Artifact 3: the cooling-architecture decision criteria. The criteria the organization applies when commissioning new AI infrastructure — typically requiring liquid or immersion cooling for new high-density deployments and requiring air-cooled or adiabatic alternatives in water-stressed regions.

Artifact 4: the water-disclosure narrative. A written disclosure (for the ESG report and for customer questionnaires) that explains the AI program’s water consumption, the share in water-stressed regions, the cooling-architecture choices, and the trajectory toward lower water intensity.

The Green Software Foundation has begun to publish principles that explicitly include water alongside energy and carbon, providing the framing that the integrated water-and-carbon accounting can use.⁷ The International Energy Agency Electricity 2024 report’s data-center growth projections imply substantial water-consumption growth that the program lead will need to manage proactively.⁸

Summary

Water consumption is the second-largest environmental impact category for AI workloads. Direct consumption is dominated by evaporative cooling at 1-3 liters per kilowatt-hour for typical air-cooled facilities; indirect consumption is dominated by thermoelectric power generation at 1-5 liters per kilowatt-hour for fossil-heavy grids. Locality matters because water consumption is a local impact, not a global one. Regulatory framing under the CSRD ESRS E3 standard and the EU AI Act Article 95 voluntary code is increasingly requiring water disclosure for AI workloads. The COMPEL D19 maturity rubric implicitly includes water in its “resource usage” framing, with Level 3 requiring per-system water tracking and Level 4 requiring integration into ESG reports and procurement decisions. The next article, M1.9Hardware Efficiency: TPUs, NPUs, and Custom Silicon for AI, develops the hardware-architecture lever that determines the IT-equipment energy and water consumption that all the facility-level practices then multiply against.

© FlowRidge.io — COMPEL AI Transformation Methodology. All rights reserved.