The next constraint on AI infrastructure - water?

AI data centres are already facing growing scrutiny over their use of both power and water. As AI workloads expand and rack densities rise, attention is turning to how water can be managed more cost-effectively through reuse, recovery and improved cooling system design - helping facilities to significantly reduce their dependence on public mains supplies or private aquifers. 

With compute demand continuing to accelerate and periodic droughts placing additional pressure on regional water resources, infrastructure planners and operators are now confronted with a set of practical operational predicaments that can no longer be ignored. This isn't just a sustainability issue anymore. It’s becoming a core part of infrastructure planning, influencing where facilities can be built, how they are cooled, and how reliably they can scale.

A report delivered by the Water Research Centre (WRC), published in February 2026 and funded by the Strategic Panel’s Market Improvement Fund to better understand water use in data centres, offers advice. It recommends steps to take to safeguard water resources while supporting sector growth. Included is an optional reporting framework to incorporate water use and water efficiency, registering data centres on a Critical National Infrastructure register, and reducing barriers to using alternatives to drinking water for cooling.

Reducing barriers can mean opening the door to circular water strategies, including the treatment and reuse of alternative sources such as contaminated groundwater or treated wastewater, where site conditions and local regulations allow.  

In parallel to the demand for water by data centres, power strategy is maturing as demand for electricity and power grids grows, and sustainability targets and operating costs come under pressure. Renewable procurement, grid resilience and long-term capacity planning are now established board-level conversations, where Power Usage Effectiveness (PUE) is routinely tracked and optimised.

However, while water underpins almost every cooling architecture in modern data centres and is emerging as a serious constraint on infrastructure expansion and cooling strategy, the question remains: is there a clear strategy for water? 

There is a strategy for energy. Water is playing catch-up

The immense computational needs of AI training and inference require specialised infrastructure that supports rapid data transfer via massive power delivery. High-density AI computing, using advanced hardware that runs at significantly higher power and thermal densities than traditional data centre equipment, creates extraordinary heat loads. Rack power densities can exceed 100 kW - up to 1MW now - compared with roughly 5-15 kW per rack in conventional air-cooled environments.

High-end performance has been made possible through a strategy for energy use. Playing catch-up is the strategy for cooling to prevent equipment failure.  Conventional air-cooling is no longer sufficient.  

What does a water strategy look like? By necessity, many facilities are adopting more advanced liquid-based cooling systems, such as direct-to-chip and immersion cooling, to handle the additional heat while maintaining reliable operation. These systems - evaporative, adiabatic or liquid-based - depend heavily on water. 

Liquid cooling - a strategic variable that alleviates risk 

The rapid adoption of liquid cooling in AI deployments has accelerated the shift towards water becoming a strategic variable and, with it, changes to the infrastructure risk profile. With liquid cooling improving thermal efficiency, higher rack densities are enabled. Simultaneously, the cooling increases sensitivity to coolant quality, filtration performance and hydraulic stability. In environments like these, water-based cooling systems sit closer to the heart of infrastructure risk.

Recent hyperscale projects in the United States illustrate the shift. Working via a specialist design-build engineering contractor, Envirogen supported next-generation AI infrastructure for a global digital platform headquartered in Silicon Valley. The requirement was not simply water treatment; it also involved protecting high-density, liquid-cooled compute environments operating at a significant scale.

Training and full-scale computing

Training a large AI model such as GPT-4 is a modest operation in some regards, despite consuming up to some 700,000 litres (153,979 gallons) of water over its training lifecycle. That’s a one-time demand during an intensive compute period [1]. 

Modest because, in contrast, a full-scale, hyperscale data centre can use up to 1.5 million litres (330,000 gallons) of water every day for cooling and humidification [2], reflecting the scale of ongoing infrastructure-level scale resource demand.

Across EMEA, demand for, and growth in, the number of AI data centres remains strong. Based on internal analysis of industry data, UK capacity is currently around 2.6GW and is expected to increase significantly by 2030, potentially reaching 6.3GW if investment is fully realised, or 4.7GW under more conservative scenarios. At a European level, total installed capacity is currently estimated at 21–22GW and is expected to rise steadily over the same period, potentially reaching 30–36GW as investment programmes are delivered.

Frankfurt is overtaking London as the top data centre hub in Europe, leading in capacity among the FLAP-D (Frankfurt, London, Amsterdam, Paris, Dublin) markets due to strong hyperscale demand. A number of the FLAP-D facilities are in water-stressed catchments, highlighting the importance of, and need for, water strategies now and going forward.

Tier two locations

Some operators are beginning to explore regional or “tier two” locations as part of their long-term infrastructure strategy. While major metropolitan hubs continue to dominate deployment, site selection is increasingly influenced by local constraints around land availability, power capacity and water resources.

In some cases, regional locations offer greater flexibility to design new infrastructure from the outset, including the integration of water reuse and recovery systems. Together, these factors mean that data centres can be designed for expansion rather than constraint, resulting in more appropriately-sized infrastructures to cater for demand – but still with water recovery and re-use elements in place. 

Regardless of where a new data centre will be built, it is increasingly common for companies to work with design-build engineering contractors to ensure that the bigger picture of future capacity is considered. In urban areas, that will mean avoiding higher costs and less stress on the water supply for the general population and businesses.

As a result, Water Usage Effectiveness (WUE) has improved across Europe. The European Data Centre Association says that average WUE fell from 1.85 in 2015 to 1.40 in 2020, with leading facilities achieving even lower ratios [3]. The Climate Neutral Data Centre Pact has meanwhile reinforced commitments to responsible water use [4]. Change is happening, with Data centres now designed to achieve 1.2 on the WUE scale.

Metrics alone don’t tell the whole story

Metrics may have improved, but they do not remove structural risk and, on their own, do not tell the whole story. In regions facing discharge constraints, abstraction limits or drought pressure, water availability can directly influence planning approvals and long-term expansion capability. Consequently, water is shifting to a strategic variable not only because of the rapid adoption of liquid cooling, but also from operational utility.

Secondary closed-loop data hall filtration

In the hyperscale AI infrastructure project referenced earlier, involving a global digital platform working with a specialist design-build contractor, filtration for the data hall plays a critical role in maintaining coolant quality, protecting downstream components and ensuring stable hydraulic performance. In high-density liquid-cooling environments, it becomes a fundamental part of overall system reliability.

In such projects, custom basket strainer systems are engineered directly into containerised, modular or skid-mounted Cooling Distribution Units (CDUs) to safeguard closed-loop cooling circuits. The role of the strainers is to protect downstream equipment, limit particulate-related risk and support stable hydraulic performance within the cooling loop. Wetted components of the CDUs were selected for compatibility with Dowfrost LC25, a widely used glycol-based coolant in advanced liquid cooling systems, helping to maintain material integrity and reliable long-term performance within the cooling loop.

Leading chip and rack manufacturers require that the secondary cooling circuit be equipped with a filtration system rated at 25 microns or finer. This specification is critical for ensuring optimal cooling performance, system reliability and long-term operational stability.

The design objectives of the project were practical but critical. Downstream heat exchangers required protection from particulate contamination such as corrosion debris, pipe scale, and installation residue, which can restrict flow paths, reduce heat transfer efficiency and increase wear on system components.

The pressure drop needed to be minimised in order to reduce pump burden, while the filtration surface area had to be maximised within tight spatial constraints. Each decision directly influenced reliability and long-term operating performance.

In parallel, high-flow filtration housings were integrated to support condensate capture and reuse, because continuous cooling cycles generate substantial condensate volumes. The overall result is effective filtration, enabling safe recirculation of water, reduced discharge and improved overall water efficiency.

These filtration solutions were not merely corrective measures introduced after commissioning. Instead, they were integrated during the design and construction phases, emphasising the importance of proven end-to-end system design from the outset rather than bolt-on fixes added later.

Designing for reuse, recovery and long-term water resilience.

Where water is becoming a constraint, resilience must be embedded at the infrastructure stage.

RO systems are central to controlling mineral scaling in adiabatic and evaporative cooling systems. Modern high-recovery RO configurations, particularly when combined with advanced membrane stages, can attain water recovery rates of 90 percent, sometimes higher, depending on feedwater quality and system design, significantly reducing discharge volumes.

Direct nanofiltration (dNF), using advanced hollow fibre membrane technology, enables treatment of rainwater, surface water and certain wastewater streams to standards suitable for reuse within cooling and utility systems. Notably, by diversifying supply sources, operators reduce their reliance on potable water abstraction and strengthen their long-term water resilience. 

Engineered water resilience

Some AI data centres are evaluating Zero Liquid Discharge (ZLD) strategies to maximise recovery and minimise environmental discharge. While ZLD is not appropriate for every site, when combined with advanced membrane technologies, it can materially reduce dependency on external water. Together, these technologies represent a move from reactive water treatment to engineered water resilience.

Diversification, combined with reuse and recovery, can help address concerns from regulators, local authorities, community groups and the wider public, while strengthening planning and water resilience.

The pillars of engineered water resilience

For operators looking to bridge the gap between energy strategy and water management, three areas really matter:

i. True circularity: Resilience starts with diversifying supply. Using Direct Nanofiltration (dNF) for rainwater harvesting or treating adiabatic blowdown for reuse, for example, helps reduce reliance on potable mains. This can extend across multiple on-site water streams, including process and utility water, to improve overall water efficiency and break dependency on external supply. It’s not just about sustainability; it’s about making sure the site can keep running when water resources become constrained locally

ii. Protection of high-value assets: As liquid cooling becomes more common, water quality is no longer optional. CDUs and GPU cold plates need protection against the increased sensitivity to coolant quality that accompanies high-density systems. As noted earlier, fine filtration (typically 25 microns or finer) is needed to remove the particulates and corrosion debris that can otherwise lead to performance issues, reduced heat transfer efficiency, or total system failure

iii. Operational efficiency and speed (prefabricated, modular delivery): To scale properly, water systems need to be treated as a fundamental part of the infrastructure, not something added later. That means using prefabricated, modular systems to speed up delivery and having proper monitoring in place to track performance and spot issues early. It also requires a whole-site approach to water management, where reuse, treatment, and distribution are considered together to optimise overall efficiency and performance.

Reduced use of chemicals 

Where higher purity water is a necessity - in steam boilers supporting on-site or associated energy infrastructure, for example - integrated RO with electrodeionisation (EDI) systems, such as MultiPro RO-EDI, provide stable purification with reduced dependence on chemical-intensive treatment approaches. Condensate polishing further lowers abstraction requirements and improves overall water balance. It does this by removing corrosion elements (e.g. iron oxides) and dissolved solids from condensed steam, thus protecting boilers and turbines from scaling and corrosion.

Infrastructure partnership, not equipment on its own

Infrastructure resilience depends not only on what is installed, but on the team that supports it over time. The same applies to water resilience, which does not depend on equipment alone. It requires early collaboration - a partnership - between operators, consultants, including cost consultants, Mechanical, Electrical and Plumbing (MEP) designers, project managers and specialist water providers.

Water systems must integrate seamlessly in hyperscale environments with cooling distribution systems, prefabricated plantrooms, automation platforms and maintenance regimes. Also essential are excellent delivery capability, engineering depth and experience of manufacturing coordination.

This model is already being seen on major AI-ready developments in Northern Europe. On one current project, Envirogen is working alongside a specialist data centre design-build contractor to deliver an integrated water package within a wider MEP scope. The system has been designed to provide high-recovery treatment for Adiabatic Air Cooling (ADAC) and Air Handling Unit (AHU) applications, helping to reduce overall site water consumption. 

To accelerate delivery and reduce onsite risk, storage tanks, booster sets and RO skids are being supplied as prefabricated offsite-manufactured modules, with the scope extending through mechanical installation, Factory Acceptance Testing (FAT), commissioning and final handover.

Long-term operational support is just as important, whether delivered in-house or by an outsourced partner team. Water systems need monitoring, optimisation, consumables management, including reliable supply of the consumables, membrane replacement cycles, spare parts and service expertise. 

For organisations scaling AI infrastructure, partner selection therefore becomes a strategic decision.

When growth meets constraint

With AI infrastructure investment continuing to accelerate, McKinsey estimates that up to 7 trillion dollars (£5.2 trillion) in capital investment may be required globally by 2030 to support AI-ready data centres [5].

Although power strategy is being solved at both national and corporate levels, data centre infrastructure does not scale on electricity alone. In water-constrained regions, operators who cannot demonstrate credible reuse, recovery and lifecycle modelling may encounter slower approvals, tighter permitting conditions or rising lifecycle costs – and public opposition. While none of these appears dramatic in isolation, with the possible exception of public opposition, together they can shape the pace of expansion.

Data centres that treat water as a secondary consideration may find that growth is limited not by demand or capital, but by resource availability. Water is rapidly becoming one of the factors that will determine how far, and how fast, AI data centre capacity can grow in line with demand.

In conclusion…

As the headline of this article suggests, the next constraint on AI data centre infrastructure could very well be water. However, the risk is significantly reduced when non-potable sources of water are treated and reused in situ, and when water resilience is integrated into infrastructure design from the outset alongside energy strategy, allowing facilities to scale more predictably and robustly.

In contrast, data centres that treat water as a secondary consideration might discover that growth is limited not by demand or capital, but by resource availability. Water is therefore rapidly becoming the next determinant of how far, and how fast, AI data centre capacity can grow in line with demand without being held back unnecessarily.  

Water resilience drives greater data centre resilience, through effective cooling systems enabling more uptime and the ability to maximise output, with the caveat that the data centre is being reliably cooled at a higher output.

References

[1] Strubell, E., Sharma, A., & McCallum, A. (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv. https://arxiv.org/abs/2304.03271

[2] Juniper Networks. (2025). Executive Circle: 5 Key Data Centre Trends for 2025.

https://www.juniper.net/content/dam/www/assets/ebooks/us/en/2025/5-key-data-center-trends-for-2025.pdf

[3] European Data Centre Association (EUDCA). (2020). WUE Performance Report: Water Usage Effectiveness in European Data Centres.

https://www.eudca.org/resources

[4] Climate Neutral Data Centre Pact. (2024 Revision). White Paper on the Responsible Use of Water by Data Centres in Europe.

https://www.climateneutraldatacentre.net/wp-content/uploads/2025/01/White-paper-on-DC-responsible-use-of-water-2024-Revision.pdf

[5] McKinsey & Company. (April 2025). The cost of compute: A $7 trillion race to scale data centers. McKinsey Quarterly.

https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers