How Data Centres Actually Consume Water — And Why AI Makes It Worse
To understand the AI water consumption problem, it helps to separate two distinct types of water use in data centre operations. The first is on-site water consumption: evaporative cooling towers and direct liquid cooling systems that physically remove heat from server hardware using water, a significant proportion of which is lost to the atmosphere. The second is off-site, or indirect, consumption tied to the generation of electricity — thermal power plants, for instance, also use substantial water in their cooling cycles.
Research published by academics at the University of California, Riverside and the University of Texas at Arlington, and covered extensively by outlets including Wired, found that training GPT-3 alone consumed approximately 700,000 litres of freshwater. Inference — the ongoing process of answering queries after a model is trained — is less intensive per individual transaction, but at the scale of hundreds of millions of daily interactions, cumulative consumption becomes significant. Large language models demand denser compute than traditional web services, meaning that even marginal per-query costs aggregate rapidly.
The International Energy Agency has tracked this trend closely, noting in its analysis of data centre energy and water use that AI workloads are materially increasing both power draw and cooling demand at hyperscale facilities. Unlike a standard database query or a webpage render, a transformer-based model inference pass keeps GPU clusters running at sustained high utilisation, generating heat that conventional air cooling struggles to manage efficiently without water-intensive supplementation.
The Transparency Gap: What Tech Companies Do and Don't Disclose
A central problem for IT decision-makers, policymakers, and privacy professionals trying to assess AI tool procurement is the near-total opacity of environmental data at the model level. Microsoft, Google, and Amazon — the three hyperscalers whose infrastructure underpins most commercial AI deployments, including OpenAI's — publish annual environmental or sustainability reports. But as research published in npj Clean Water noted, these disclosures are typically aggregated at the company level and do not break down consumption by workload type, model, or geographic location.
This creates a practical problem for organisations operating under frameworks like the EU AI Act or those conducting supply chain due diligence under emerging European sustainability directives. If a company cannot determine the water intensity of the AI services it procures, it cannot meaningfully include that footprint in environmental, social, and governance (ESG) reporting. And if regulators cannot compel workload-level disclosure, AI's environmental costs remain effectively invisible in corporate accounts.
Microsoft disclosed in its 2023 environmental report that its global water consumption had increased by 34% compared to the prior year, a rise it attributed in part to AI infrastructure expansion. Google reported a 20% increase in water use over a similar period. Neither company provided per-model breakdowns. For organisations building AI governance frameworks — a task now mandated for high-risk AI deployments under the EU AI Act — this absence of granular data is a structural gap.
"We tend to measure AI's cost in compute hours and token prices, but water is becoming one of the defining resource constraints of the AI era — and right now, almost no procurement framework accounts for it."
— Senior infrastructure analyst, European cloud advisory practiceWhat AI Water Consumption Means for European Digital Sovereignty
For European policymakers and technology teams pursuing digital sovereignty goals, AI water consumption adds a new dimension to an already complex debate. The European Commission's data strategy and the broader push for European cloud alternatives — championed by initiatives such as Gaia-X and accelerated by GDPR-driven data localisation requirements — have largely focused on data residency, jurisdictional control, and avoiding dependence on US hyperscalers. Water stress has rarely featured in those conversations.
Yet the geography of AI infrastructure matters enormously for water. Data centres in water-stressed regions — and parts of the American Southwest, where large hyperscaler campuses are concentrated, face severe long-term water scarcity — consume freshwater from already-stressed aquifers and river systems. European data centres, while subject to increasing climate-driven water stress of their own, are more likely to operate in regulatory environments where water abstraction is licensed and reported. The EU's Corporate Sustainability Reporting Directive (CSRD), which entered into force for large companies, includes water use in its mandatory disclosure scope, potentially giving European procurement teams more visibility into supplier water consumption than US-based counterparts would receive.
The AI Act's requirements for transparency and documentation of high-risk AI systems could, if extended to environmental impact, create a reporting infrastructure that makes AI water consumption visible at a granular level for the first time. Several European digital rights organisations, including EDRi, have called for environmental impact to be included in AI system documentation requirements — a step that would align AI governance with the CSRD and give compliance teams a single reporting thread to pull.
Comparing the Water Footprint of Major AI Workloads
Not all AI workloads carry the same water cost. The architecture of the model, the hardware it runs on, the cooling technology deployed at the specific facility, and the ambient climate of the data centre location all affect per-query water intensity. Smaller, distilled models running on more efficient hardware in cooler climates — the kind increasingly promoted by open-source European AI projects — can carry substantially lower footprints than frontier models running on dense GPU clusters in hot, arid locations.
| AI Workload Type | Relative Water Intensity | Key Variables |
|---|---|---|
| Frontier LLM inference (e.g. GPT-4o) | High | Dense GPU clusters, high query volume, long context |
| Mid-size open-source LLM (self-hosted) | Medium | Hardware efficiency, facility cooling type, location climate |
| Small/distilled model (e.g. 7B parameter) | Low–Medium | CPU/edge hardware, lower thermal load |
| LLM training run (one-time) | Very High (one-off) | Duration, cluster size, facility Water Usage Effectiveness (WUE) |
| Traditional cloud web service | Low | Lower compute density, mature cooling optimisation |
For developers and IT decision-makers evaluating AI tool stacks, these distinctions matter in practical ways. A team choosing between a frontier model API, a self-hosted mid-size open-source model, and a fine-tuned smaller model is making not just a cost and capability decision, but also an environmental one. As European sustainability reporting obligations tighten, the ability to quantify and document that choice will increasingly be a compliance requirement, not merely a voluntary ESG gesture.
The rise of smaller, more efficient models — including Mistral's European-developed alternatives, which have gained significant traction among privacy-conscious developers seeking GDPR-aligned deployments — offers a partial path toward reducing AI's water footprint while maintaining capability for many use cases. Running inference on-premises or in European colocation facilities with documented water abstraction licences also gives compliance teams an audit trail that API-based consumption of US-hosted frontier models cannot currently provide.
What Should Developers and Procurement Teams Do Now?
For organisations that want to act on AI water consumption before regulatory frameworks compel them to, a practical starting point is Water Usage Effectiveness (WUE) — a metric, analogous to Power Usage Effectiveness (PUE) for energy, that measures how efficiently a facility uses water per unit of IT work. Some hyperscalers publish facility-level WUE figures; requesting this data as part of vendor due diligence is a defensible first step.
Beyond WUE, procurement teams should consider geographic placement of AI workloads. Facilities in water-stressed regions carry higher social and environmental risk, and as water scarcity worsens under climate change projections, they also carry operational risk — restrictions on water abstraction during drought conditions have already affected data centre operations in Ireland and the Netherlands. Choosing providers who locate AI infrastructure in regions with abundant, sustainably managed water resources, and
Originally reported by Silicon Canals. Summarised and curated by European Purpose.
