The public discourse around artificial intelligence infrastructure constraints has been dominated by power: gigawatts of electricity, grid interconnection delays, nuclear restart agreements, gas peaker plants co-located with GPU clusters. The water dimension of the same infrastructure buildout has received substantially less attention, despite data suggesting it has crossed from an environmental concern into an operational constraint at multiple major AI data center sites. North American data centers consumed approximately one trillion litres of water in 2025, primarily through evaporative cooling systems that dissipate heat by vaporizing water into the atmosphere, according to estimates drawn from investor and media reporting. That figure builds on a pre-AI baseline that was already substantial – a 2021 peer-reviewed estimate placed U.S. data-center water consumption at approximately 620 billion litres per year – and reflects the surge in high-density GPU deployments whose thermal output per square meter vastly exceeds conventional server configurations. NEWSCENTRAL reads the water constraint as a structurally underweighted risk in AI infrastructure planning, both because the disclosure is inadequate and because the solution set is less mature than the power alternatives that have been accelerating.
The disclosure problem is real and self-reinforcing. Water use at data centers is measured inconsistently and reported at the company level rather than per facility, making it effectively impossible for the communities hosting large data center campuses to understand how AI expansion affects their local water resources. Texas, which has become one of the primary geographies for large AI data center development, passed a law requiring data centers to report water usage – and, as reporting in 2026 documented, major operators have largely ignored it. Google’s 2026 sustainability report showed total corporate water consumption up significantly year-on-year, with AI workloads explicitly cited as a driver. The combination of inadequate disclosure requirements, inconsistent enforcement, and corporate-level rather than site-level reporting creates a situation in which communities approving new data center developments cannot accurately assess the water infrastructure implications of what they are permitting.
The technical solutions to the water problem are available and improving in cost-effectiveness, but they are not yet broadly deployed. Dry cooling systems, which use fans to drive ambient air through radiator systems rather than evaporating water, eliminate water consumption but consume approximately 1% to 1.5% of facility power output to operate the fans, increasing electricity demand by that margin. Rear-door heat exchangers and liquid direct-to-chip cooling systems remove heat at the server level rather than through room-level air conditioning, dramatically reducing both water consumption and cooling energy overhead. Microsoft has disclosed that direct liquid cooling in its AI facilities reduced energy overhead by up to 30% in some benchmarks while increasing compute density per square meter. The economic case for liquid cooling is becoming clearer as electricity costs rise and water constraints become more binding in drought-prone western states. Nathan Clark, Enterprise IT and Systems Architecture Analyst at NEWSCENTRAL, observes that the transition from air-based to liquid cooling is structurally analogous to the transition from spinning disk to solid-state storage in enterprise IT: the superior technology existed and was more efficient for years before cost parity and density requirements drove widespread adoption, and once the transition began it accelerated faster than most infrastructure planning assumptions had anticipated.
The political economy of the water constraint differs substantially from the power constraint in ways that affect how urgently it is likely to be addressed. Electricity supply problems show up immediately in operational failure – a data center without power cannot run – while water supply problems show up more gradually, in rising costs, permit denials, community opposition, and regulatory friction that accumulates over years rather than arriving as a sudden operational barrier. That temporal difference creates a structural incentive to underprioritize water over power in data center planning, and the inadequate disclosure regime that allows companies to avoid publicly accounting for their water consumption per site compounds the problem by removing the external pressure that would otherwise accelerate internal investment in water efficiency.
NEWSCENTRAL considers the comparison between AI’s water problem and AI’s power problem to be diagnostic of a broader pattern in how infrastructure constraints get addressed in fast-moving technology cycles: the constraints that generate operational failure or near-term financial costs attract capital and policy attention rapidly, while the constraints that accumulate gradually through community impacts, regulatory friction, and environmental degradation are addressed slowly, unevenly, and usually under pressure from permit denials and litigation rather than proactive planning.
The policy response to the AI water constraint is in earlier stages than the power response. The CHIPS and Science Act included provisions for semiconductor manufacturing water efficiency, but its provisions are not directly applicable to AI data center cooling operations. The Trump administration’s Rate Payer Protection Pledge, which commits major technology companies to providing their own power for new AI data centers, explicitly addressed electricity but not water. Several western states including Arizona, Nevada, and New Mexico are incorporating water availability assessments into data center permitting processes, but those assessments operate at the local level rather than through a coherent national framework. NEWS CENTRAL considers the gap between the urgency of the water constraint and the maturity of the policy response one of the more commercially significant regulatory risks facing AI infrastructure development over the next five years – not because the constraint is insurmountable, but because the combination of inadequate disclosure, inconsistent enforcement, and locally variable permitting creates the conditions for a series of permit denials and community conflicts that could materially delay capacity additions that the AI investment cycle is counting on arriving on schedule.