Water stress is becoming the hidden constraint on AI infrastructure growth. As artificial intelligence computing demands skyrocket, data centers are hitting a critical bottleneck: not electricity alone, but the water needed to cool the systems that process AI workloads. By 2050, 31% of global GDP will be exposed to high water stress, up from 24% in 2010, according to research from Wood Mackenzie . This shift is forcing the energy and technology sectors to rethink cooling strategies at a moment when AI is pushing computing density to unprecedented levels.

Why Is Water Becoming a Critical Constraint for AI Data Centers?

The problem is straightforward but urgent: traditional power generation and modern AI infrastructure both depend heavily on water for cooling, yet water availability is becoming increasingly volatile. Thermoelectric, nuclear, and hydroelectric plants produced 80% of global power in 2025, and all require continuous water access for cooling . Meanwhile, the explosion of AI computing is pushing data centers toward liquid cooling systems that can handle 250 kilowatts per rack, roughly 10 times what air cooling can manage . This parallel demand spike arrives precisely as water stress intensifies across South Asia, the Middle East, North Africa, and the western United States, where aquifers are being drawn down faster than natural recharge rates.

Recent episodes across Europe illustrate the real-world impact. High river temperatures and low flows forced nuclear output cuts and temporary reactor curtailments, exposing a systemic vulnerability in energy infrastructure that depends on water availability . As climate variability increases, this pattern will likely repeat in other regions, creating operational and economic risks for both power generation and AI deployment.

How Are Data Centers Adapting to Water Constraints?

The technology sector is responding with a shift toward more efficient cooling methods. Here are the primary approaches data centers are adopting:

• Single-Phase Direct-to-Chip Liquid Cooling: Hyperscalers are standardizing systems that circulate warm water (30-45 degrees Celsius, compared to 15-25 degrees for air systems) through cold plates attached to GPUs and CPUs, handling 60-150 kilowatts per rack while improving energy efficiency .
• Two-Phase Refrigerant Systems: Advanced cooling using refrigerants can exceed 250 kilowatts per rack, enabling the highest-density AI training clusters to operate within water constraints .
• Hybrid and Dry Cooling for Power Plants: Traditional thermal power plants are shifting toward hybrid cooling systems and dry cooling, which eliminates water use entirely but carries a 7-percentage-point efficiency penalty and adds $160 per kilowatt in capital costs .

The efficiency gains from liquid cooling are substantial. Enterprise facilities historically split power 60/40 between IT equipment and cooling; hyperscalers have pushed that ratio to 90/10, meaning nearly all energy now goes to compute rather than facility overhead . Power usage effectiveness is expected to drop to 1.2 by 2028, a significant improvement . Companies like Vertiv, CoolIT, and Asetek are producing manifold and coolant distribution systems aligned to NVIDIA's H100, GB200, and GB300 platforms .

What's the Real Water Cost of AI Infrastructure Versus Power Generation?

Here's a critical insight: while data center cooling is becoming more water-efficient, the power plants that supply electricity to those data centers remain far more water-intensive. Thermal power generation uses 10 to 20 times more water than data center on-site cooling . Traditional once-through cooling systems withdraw 132.5 cubic meters per megawatt-hour but consume only 0.9 cubic meters. Wet recirculating towers, now the industry standard, reduce withdrawals to 4.6 cubic meters per megawatt-hour but triple consumption to 3.1 cubic meters through evaporation .

"AI clusters generate heat loads that air simply cannot handle at scale. Liquid cooling isn't optional anymore, it's the foundation for next-generation compute," stated Jom Madan, Principal Analyst for Scenarios and Technologies at Wood Mackenzie.

Jom Madan, Principal Analyst, Scenarios and Technologies at Wood Mackenzie

Madan added an important observation about where the real exposure lies: "The water question hasn't gone away; it's moved from the data hall to the power plant and that's where the real exposure sits. Thermal power generation remains 10 to 20 times more water-intensive than data centre on-site cooling. As water stress intensifies, the case for wind, solar, and dry cooling becomes operational, not just environmental" .

Madan

What Does This Mean for the AI Chip Market and Infrastructure Investment?

The broader AI chip market is expanding rapidly, with the global market valued at $102.89 billion in 2025 and projected to reach $1,354.35 billion by 2035, growing at a compound annual growth rate of 29.4% . GPUs dominated the market with 46% share in 2025, as they are highly efficient at handling AI-related tasks . However, this explosive growth in chip demand directly translates to increased power consumption and cooling requirements, making water management a strategic business issue, not just an environmental one.

The AI infrastructure challenge is attracting significant venture capital attention. Energy tech startup Emerald AI recently raised $25 million to help data centers balance power use with grid capacity amid rising AI demand . The funding round was led by Energy Impact Partners and included participation from NVIDIA's venture arm NVentures, Salesforce Ventures, Samsung Ventures, and others, signaling broad recognition that energy and water management are critical to scaling AI infrastructure .

Emerald AI's platform, called Emerald AI Conductor, allows data centers to function as flexible grid resources rather than static energy users by adjusting power consumption based on grid conditions and distributing computing workloads across networks of data centers . This approach helps ease strain during peak demand periods while maintaining strict computing standards. The company's advisory board includes NVIDIA, Salesforce Ventures, National Grid, Eaton, GE Vernova, and Siemens, reflecting the collaborative effort required to tackle energy and water challenges at scale .

What Policy and Infrastructure Changes Are Needed?

The technology exists to manage water constraints more effectively, but deployment is lagging behind market demand. According to Wood Mackenzie's analysis, the missing piece is a policy framework to accelerate deployment at the speed the market requires . India, Mexico, Egypt, and Turkey account for over half of the global GDP exposed to high water stress, making this a geographically distributed challenge that requires coordinated responses .

New power plant builds are increasingly favoring wind and solar, which require far less water to operate, over traditional thermal generation . This shift aligns with the operational and economic case for water-efficient infrastructure. However, existing thermal and nuclear plants will continue to dominate the power mix for decades, meaning hybrid cooling systems and dry cooling technologies will remain essential investments.

The convergence of AI growth, water scarcity, and energy infrastructure constraints is reshaping investment priorities across the technology and energy sectors. Companies deploying AI infrastructure must now factor water availability and cooling efficiency into site selection and operational planning. For policymakers, the challenge is creating regulatory frameworks that incentivize water-efficient cooling technologies and renewable energy sources fast enough to keep pace with AI's explosive growth.