Ireland, where data centers reportedly represent 21% of total metered electricity use, has already paused new approvals in parts of Dublin until 2028
Decision Focus
A report published by the UN University Institute for Water, Environment and Health (UNU-INWEH) in June 2026 quantifies AI infrastructure’s combined electricity, water, and land demands through 2030. The operational signal for data center energy heads is not the carbon headline. It is the finding that low-carbon energy choices can reduce emissions while simultaneously increasing water and land pressure—and that regulators are already acting on those secondary constraints in markets where AI capacity is most concentrated.
90-Second Brief
As the week closes, the UNU-INWEH report projects that AI-powering data centers could consume 945 TWh of electricity globally by 2030, roughly double the 2025 baseline the report cites for all data centers combined. Alongside electricity, the report projects 9.3 trillion liters of water use and over 14,500 square kilometers of land by the same date. Ireland, where data centers reportedly represent 21% of total metered electricity use, has already paused new approvals in parts of Dublin until 2028. The governance argument at the center of the report lands directly where energy procurement decisions intersect with water and land planning.
What Is Really Happening?
The report’s core argument challenges a premise that has shaped most hyperscaler sustainability strategy: that renewable energy procurement resolves AI’s environmental liability. The authors note that some low-carbon generation sources can reduce carbon emissions while increasing water consumption or land footprint—a trade-off the current reporting framework, focused on Scope 2 emissions and 24/7 carbon-free energy matching, does not capture.
The structural driver is inference workload. The report attributes 80 to 90 percent of AI energy consumption to inference rather than training, meaning the load is continuous and grows proportionally with user adoption rather than with model development cycles. That characteristic makes demand growth harder to manage through efficiency gains alone. The report describes a rebound dynamic: as AI becomes cheaper and faster per query, aggregate usage rises, offsetting per-unit improvements. A single widely adopted AI application is cited as consuming hundreds of gigawatt-hours annually from inference alone—a scale that, multiplied across the application landscape, compounds faster than infrastructure planners have historically modeled.
Geographic concentration adds another layer. The report notes that more than 90% of global AI data center capacity sits in just two countries, while only 32 countries host AI-specialized infrastructure at all. That concentration means regulatory and resource constraints in a small number of markets can directly constrain global capacity planning cycles in ways that distributed build-outs historically did not.
Why It Matters for Global Heads of Data Center Energy
The practical exposure arrives at three intersecting points.
First, site selection criteria are widening. If regulators in key markets begin conditioning data center approvals on water and land impact assessments alongside grid load studies, the due diligence framework for new sites changes materially. Dublin’s pause is one data point. The report also notes that data center expansion in Mexico and Uruguay has coincided with severe drought conditions, suggesting the pattern may extend to emerging capacity markets that operators have been counting on for geographic diversification.
Second, PPA and renewable energy strategy carry a new dimension. The report’s finding that low-carbon choices can increase water or land impact adds a variable to energy procurement due diligence that standard REC and 24/7 CFE frameworks do not surface. A solar configuration with evaporative cooling that achieves carbon accounting targets may create water liability in a drought-stressed region. That liability may not yet translate to a financial exposure in most jurisdictions, but markets already showing grid stress are credible candidates for early regulatory movement on multi-resource disclosure.
Third, capital planning models should be stress-tested against the inference trajectory. The report’s scenario that AI’s share of data center energy use could roughly double from its current level to 40% by 2030 is a projection, not a guarantee. Operators who have built capital plans around moderate AI load growth should verify that their demand assumptions account for inference adoption rates rather than only training infrastructure additions.
Forward View
Three fronts carry the most operational weight if the report’s framing gains traction with regulators and institutional investors.
Water disclosure requirements are the most proximate candidate for near-term regulatory action. The EU’s data center sustainability reporting requirements and emerging disclosure frameworks in the United States have focused primarily on carbon. If the UNU analysis influences policy dialogue, water use intensity metrics could enter formal disclosure requirements within this planning cycle, particularly in European jurisdictions where grid constraints have already produced approval moratoria.
Market access risk in concentrated geographies is the second front. Dublin, Singapore, and Northern Virginia have each experienced grid or permitting constraint events in recent years. If multi-resource assessment becomes a condition of permitting rather than a voluntary reporting choice, approval windows in those markets narrow further and lead times extend.
The third front is competitive positioning on resource efficiency. Operators who can demonstrate low water and land intensity alongside low carbon per megawatt-hour will have a structurally stronger position with regulators, host communities, and institutional investors as environmental screening criteria broaden. That positions water and land efficiency as a future procurement and infrastructure design parameter, not only a sustainability communications decision.
What Is Still Uncertain
The source summarizing this research is a secondary report. The underlying methodology behind several projections—particularly the electricity and water figures—should be verified against the UNU-INWEH primary publication before being used in internal forecasting or regulatory submissions. How the report allocates water and land footprints across specific cooling technologies, energy mixes, and geographies is not detailed in the available summary, which limits the precision with which operators can map these findings to their own portfolio.
It is also not confirmed whether any regulator is actively developing water or land use thresholds as formal conditions of data center approval, or whether Dublin’s pause represents a transferable regulatory template. The report recommends that governments include AI infrastructure in water and energy planning, but a policy recommendation and a binding regulatory instrument are different stages of a governance process. The timeline and jurisdiction for any such shift remain open.
One Question for Your Team
Which of our planned or active sites in the next three years are located in water-stressed markets, and does our current PPA or cooling technology choice in those markets increase or reduce water consumption relative to the regional grid baseline?
Sources
- Carboncredits — AI’s Environmental Cost: Data Centers Now Rival Entire Nations in Energy, Water, and Land Use (Link)
