The Liquid Cooling Revolution: How Data Centres Are Rewriting the Rules of Water, Energy, and AI Infrastructure

By Catie Owen | Regional Content & Insights Lead
Published: October 6, 2026 | 7-minute read

Executive Overview

As the artificial intelligence (AI) boom accelerates, the digital world is increasingly crashing into the hard limits of the physical world. While the public discourse often centers on the computational marvels of large language models and advanced machine learning algorithms, the rapid expansion of AI is fundamentally a story of physical infrastructure: chips, power grids, millions of gallons of water, and, above all, the relentless physics of heat removal.

With rack densities routinely climbing past 100 kW—and pushing toward 130 kW in hyper-specialized deployments—traditional air-cooling infrastructure is approaching a hard technological ceiling. The central challenge facing the global data centre industry has fundamentally shifted. Operators are no longer merely asking how to provision more power to a facility; they are urgently trying to figure out how to cool those racks efficiently without imposing unsustainable burdens on local water supplies and municipal utility grids.

To unpack this shifting paradigm, Capacity sat down with Matt McLaughlin, Corporate Account Manager for Ecolab’s data centre business. McLaughlin offers a front-row seat to the engineering and environmental pressures reshaping the industry. According to McLaughlin, modern data centre operators view water and energy infrastructure not just as operational inputs, but as the core determinants of a facility’s "license to operate."

As this report explores, navigating the future of AI infrastructure requires abandoning legacy, linear cooling models in favor of a holistic, circular approach that views heat, water, and power as an interconnected ecosystem.

Detailed Chronology: The Evolution from Air to Liquid

To understand why the industry is undergoing a massive retrofit wave, it is helpful to examine how data centre cooling has evolved alongside compute workloads over the past two decades.

• The Era of Low-Density Air Cooling (Late 1990s – Early 2010s): For years, standard enterprise servers operated at modest rack densities ranging from 3 kW to 5 kW. Standard Computer Room Air Conditioner (CRAC) and Computer Room Air Handler (CRAH) units easily managed these thermal loads by pushing large volumes of chilled air across raised floors. Water usage was largely confined to evaporative chillers rejecting heat at the facility perimeter.
• The Cloud & Hyper-Scale Expansion (2010s – Early 2020s): As virtualization and hyperscale cloud computing took off, rack densities crept up to 10 kW and 20 kW. Air-containment aisles and hot/cold aisle configurations became standard practice. However, water usage effectiveness (WUE) began to draw intense regulatory and community scrutiny, particularly in arid regions like Northern Virginia, Dublin, and Phoenix.
• The Generative AI Inflection Point (2023 – Present): The mainstream explosion of generative AI necessitated clusters of specialized GPUs (such as NVIDIA’s H100 and Blackwell architectures) that generate unprecedented thermal loads. Racks routinely exceed 100 kW. Air cooling, which relies on fans and chillers moving massive volumes of air, has become physically incapable of handling these concentrated hotspots efficiently.
• The Liquid-Cooling Mainstream (2026 and Beyond): The industry has reached an inflection point where direct-to-chip liquid cooling and advanced hybrid facility loops are transitioning from experimental alternatives to baseline design requirements for any new high-density build or enterprise retrofit.

Supporting Context & Metrics: The Physics of Thermal Management

The transition to liquid cooling is not driven by marketing hype, but by fundamental thermodynamics.

Starting with the Physics

As McLaughlin emphasizes, solving the modern thermal puzzle begins with understanding basic physics:

“Liquid—specifically water—moves heat far more effectively than air. Water can carry roughly 3,500 times more heat than the same volume of air and transfer heat about 23 times faster.”

To visualize this, consider a high-performance combustion engine. While an external fan can cool the surface of the engine block, circulating liquid coolant directly through the core removes heat from the exact point of generation. The same principle applies to modern high-density compute clusters.

Air cooling forces fans and chillers to churn through monumental volumes of air across blazing-fast silicon components, devouring massive amounts of parasitic power in the process. By bringing liquid closer to the heat source—either via secondary water loops at the rack level or specialized dielectric coolants circulating directly over the chip—operators capture thermal energy much closer to the source at higher temperatures, using a fraction of the energy.

The Hidden Water Footprint: On-Site vs. Off-Site

A critical insight highlighted by Ecolab’s data centre division is that the energy story is inherently a water story.

Much of the global electrical grid relies heavily on water resources somewhere along the generation chain—ranging from the cooling towers of thermal power plants to the massive reservoirs feeding hydroelectric dams. Consequently, every kilowatt-hour saved by an optimized cooling system also reduces the upstream water consumed to generate that electricity.

[ On-Site Cooling Efficiency ] 
      │
      ├──> Reduces direct water consumption inside the facility fence
      │
      └──> Lowers power demand 
                │
                └──> Reduces "embedded water" consumed off-site at power generation plants

McLaughlin notes that this upstream water consumption is rarely captured on standard facility balance sheets:

“It’s real, and it means efficient cooling improves the water math in two places at once. On-site, water is used purposefully and kept in circulation. Off-site, lower power demand reduces the water embedded in the grid that feeds the facility.”

Official Insights & Strategic Frameworks

Addressing high-density thermal loads requires abandoning siloed metrics and adopting comprehensive, multi-stage cooling chains.

Breaking Down the Cooling Chain

Framing cooling as a binary choice between competing technologies is a common pitfall. In practice, modern high-density data centres operate via a sophisticated cooling chain:

1. Chip to Rack: Thermal energy is captured directly at the processor via direct-to-chip liquid cooling or immersion techniques, tackling the most extreme thermal loads.
2. Rack to Facility: Intermediate coolant loops transfer heat away from the server racks toward centralized facility distribution units (CDUs).
3. Facility to Atmosphere: Heat is ultimately rejected externally through evaporative cooling towers, dry coolers, adiabatic systems, or hybrid units, depending entirely on local climate, electrical availability, and water scarcity.

The Danger of Metric Isolation

One of the most dangerous traps for modern operators is optimizing metrics in isolation. Water Usage Effectiveness (WUE) and Power Usage Effectiveness (PUE) are frequently tracked as independent scorecards. However, they are deeply and inversely linked.

An engineering approach designed to minimize on-site water consumption might inadvertently demand significantly more electrical energy. That additional energy draw, in turn, can trigger higher water consumption upstream at regional power plants.

“Where operators can go wrong is optimising one metric in isolation,” McLaughlin warns. “The better approach treats heat, water, and energy as one connected balance, measured inside and outside a facility based on climate, power grid, water availability, and performance requirements.”

Fluid as Critical Infrastructure

Whether an operator deploys direct-to-chip liquid cooling, hybrid evaporative loops, or closed-loop systems, efficiency rapidly collapses if the fluid itself is neglected.

Water and specialized coolants are not "set-and-forget" utilities. Left unmanaged, factors such as corrosion, microbiological growth, particulate accumulation, and coolant drift will steadily erode heat-transfer efficiency.

McLaughlin highlights real-world examples where proactive intervention prevented catastrophic failures:

• Hyperscale Site: Continuous automated monitoring detected micro-levels of corrosive particles that routine scheduled maintenance had completely missed, allowing engineers to remediate the issue before hardware damage occurred.
• Colocation Facility: Real-time chemistry monitoring flagged diluted coolant parameters during the critical commissioning phase, allowing technicians to make immediate corrections and keep the multi-million-dollar facility strictly on schedule.

Future Outlook: What This Means for Operators

As the global footprint of artificial intelligence expands over the remainder of the decade, the demand for high-density compute will only accelerate. Yet, resource waste does not have to scale in lockstep with compute capacity. While electrical generation capacity can theoretically be scaled up through new renewable infrastructure and grid investments, local water availability remains strictly bounded by geographical and ecological limits.

For engineering teams planning their next high-density build or brownfield retrofit, McLaughlin outlines four non-negotiable strategic pillars:

1. Start with the Site: Local climate, regional power sources, watershed stress, and workload profiles must dictate the cooling architecture—never the other way around.
2. Count Water and Energy Together: Model WUE and PUE side-by-side, explicitly factoring in the embedded water footprint of upstream power generation.
3. Match the Method to the Heat: Deploy liquid cooling where high-density compute demands it, and pair it with facility-level heat rejection systems tailored to local environmental conditions.
4. Treat Fluid as Infrastructure: Implement real-time monitoring of coolant chemistry, temperature, and flow from the commissioning phase onward to safeguard performance over the entire lifecycle of the facility.

Ultimately, the data centre operators who successfully navigate the AI era will be those who master the delicate equilibrium between digital ambition and physical resource stewardship—delivering maximum compute capacity with minimal environmental impact.