Water for data centres: How low can we flow?

AI and the continuing demand for cloud computing is rapidly reshaping the scale and delivery of data centres. Facilities are getting larger, power densities are rising and cooling loads are pushing well beyond what earlier generations of infrastructure were designed to handle. As this expansion accelerates and timelines for delivery compress - water has become a point of tension, particularly in places where supplies are already constrained.

Data centre water demand is often presented as fixed, inevitable and incompatible with community needs. In practice, this framing misses something important. Water use is not always an inherent feature of data centres; it is largely a consequence of design choices - many of which mirror decisions other industries have already navigated.
  
Power stations, large industrial complexes and more recently, hydrogen projects have been navigating the same terrain of complex design choices and trade-offs for decades . Experience suggests that early estimates tend to overstate eventual water demand and that real-world constraints - such as schedule, approvals, waste management and ultimately cost - almost always force complex design choices and trade-offs that drive systems to evolve and optimise.  And learnings from other industries matter, because data centres are now passing through a similar phase.

Part of the challenge stems from how cooling is considered. Cooling is often treated as a single function, when it’s really three separate ones: removing heat from equipment, moving that heat through various cooling circuits and then rejecting it to the surrounding environment.  
        
What drives water demand is how heat is rejected to the environment - the cooling source. This is where design decisions have the biggest consequences for water consumption and waste. Evaporative systems like cooling towers use water directly and can require significant make-up water if not carefully optimised through onsite reuse or feed water quality management.

Air-based systems largely avoid continuous water use, but in warmer climates often rely on adiabatic cooling - pre-cooling with water sprays - during hot periods, creating sharp, intermittent demand peaks. In all cases, these approaches bring trade-offs in energy use, footprint and cost. They are also typically used in conjunction with chillers, where it is heat from the condenser unit that is rejected to the environment.

Closed-loop cooling circuits are used to transfer heat from the equipment to the cooling source.  These circuits are typically filled with a mix of treated water and additives to prevent freezing, corrosion, scaling and biological growth.  Very small volumes of water are required for commissioning and often for ongoing make up. 
There is no universal answer. Cooling strategies are shaped by site selection, local climate, noise constraints, available land and how water can be supplied and discharged.

The power sector shows how these dynamics play out over time. Early power stations made extensive use of once through cooling using water extracted directly from rivers, lakes or the ocean. That approach fell away as environmental impacts from warm water discharges became harder to justify. Evaporative cooling became dominant, reducing withdrawals significantly but creating concentrated wastewater streams that also needed to be managed.

As water availability tightened further, power stations began to adopt dry and hybrid systems, recovery of water from waste streams and invested in increasingly complex treatment and disposal solutions (e.g. zero liquid discharge processes). These shifts were not driven by ideology. They were practical responses to regulation, cost and risk.

Hydrogen projects have followed a similar trajectory - compressed into only a few years. Initial water estimates were low, in some cases because cooling was largely overlooked. These were followed by conservative projections that suggested very high demands. Only once water quality, waste handling, ambient temperatures and energy penalties were properly accounted for did designs begin to settle into more balanced, site specific solutions.

Data centres are now reaching that same point. The early numbers attract attention, but they rarely survive contact with delivery reality once considered integrated system design is undertaken.

Across sectors, the same levers consistently shape water outcomes.

Siting is often decisive. Access to non potable supply, recycled water, discharge pathways and local climate can matter more than the choice between one cooling technology and another. Increasingly, sophisticated geospatial analysis is used to weigh these factors upfront.

Cooling system selection remains important, but it is never made in isolation. Facilities with stringent latency or noise constraints, often those near population centres face tighter limits on the cooling options available.

Water quality also matters more than is sometimes acknowledged. For example, underestimating the implications of saline or brackish waste streams can quickly undermine both cost and operability.

Closed-loop cooling circuit operating temperature is another method enabled by liquid immersion or direct-to-chip liquid cooling. In the hydrogen sector, small increases in allowable temperature unlocked substantial reductions in cooling demands and water demand. Data centres operate within narrower thermal bands, but even modest shifts can deliver meaningful gains.

Perhaps the largest opportunities lie in integrated design. When power, cooling, building systems, IT hardware, water supply and waste management are considered together, trade offs become visible that are missed in siloed approaches.

A recurring mistake in fast growing industries is reliance on single point demand estimates. In practice, technology improves, operating strategies change and extreme conditions tend to affect peaks rather than averages.

More robust approaches acknowledge that uncertainty and plan around it. Probabilistic modelling, whole of system analysis and adaptive pathways planning allow decisions to be tested against a range of futures rather than a single forecast. This matters for data centres in particular, where capacity is often modular and demand grows in stages. Probabilistic modelling also allows water utilities and government to plan for the impact of “phantom demands” which has resulted in unrealistically elevated demand for water in many locations due to multiple applications for water supply from each data centre.

The type of data centre (AI versus cloud data storage) and the type of developer (e.g. Hyperscalers like Meta, Google and Microsoft, colocation providers and smaller operators) will affect the future demand of water.  Hyperscalers and AI data centre developers can afford more expensive higher electricity demand low water demand solutions such as “dry cooling” that take advantage of direct-to-chip and liquid immersion cooling.  Whereas colocation providers and smaller operators often face capital constraints and retrofit limitations.  In terms of industry trajectory and the potential impact on water consumption, typically low water demand AI data centres will start to dominate data centre capacity from 2030.

Water stress exposure for data centres is projected to increase over time, not stabilise – due to climate change increasing ambient temperatures and driving up cooling needs - water supply and scarcity become more significant. The water-based constraints that are becoming increasingly prominent will only increase in the future.  This may drive the industry towards the use of recycled water and perhaps, as a key opportunity, assist water utilities create new recycled water infrastructure that can increase the water security of the entire community by accessing these new water sources.

The industry is building the next decade's infrastructure today – which makes integrated planning, site selection and modelling of future scenarios even more crucial for long term success.

Rather than locking in one solution, these approaches preserve flexibility, allowing systems to adapt as constraints, technologies and expectations evolve.

The water story for data centres will split. AI facilities - typically high-density, hyperscaler-backed - can adopt near-zero water use technologies like liquid immersion, though at higher cost. Conventional data centres, run by colocation providers and smaller operators facing capital constraints, will be pushed towards more water-intensive cooling. Forecasts that treat all data centres as one category misread where the real pressure lies - a distortion compounded by phantom demand: developers lodging water supply applications across multiple sites for a single facility, meaning reported demand can overstate reality by up to an order of magnitude.

There is a counterpoint. Data centres are large, creditworthy consumers of non-potable water - ideal anchor customers for recycled water schemes. Their guaranteed offtake can underwrite wastewater treatment infrastructure, unlocking purified recycled water not just for the facility but for the wider community.

The aim should not be zero water at any cost nor acceptance of inflated early estimates as inevitable. The more useful goal is fit for purpose water use: cooling systems that meet performance requirements while respecting local water availability and community expectations.

If there is one clear lesson from power, hydrogen and other industries, it is that water demand is rarely fixed. Once constraints become real, systems adapt, often quickly.

The choices being made now about location, cooling strategy and system integration will shape whether data centres place long term pressure on urban water systems or whether they are designed to coexist with them.