By Stuart Kerr, Technology Correspondent – LiveAIWire
Published: August 2025 | Updated: August 2025
Contact: [email protected] | @LiveAIWire
Meta description: AI data centres consume vast amounts of energy and water. The overlooked environmental costs of training models may shape the sustainability debate as much as carbon emissions.
The Other AI Footprint
The story of AI’s environmental toll is often framed in terms of carbon. Headlines focus on how training massive models burns through electricity, fuels fossil plants, and inflates emissions. Yet another resource is quietly strained in the process: water. Cooling the sprawling server farms that power AI requires millions of litres of freshwater, drawing from rivers, reservoirs, and sometimes already stressed local supplies.
Research from the Environmental and Energy Study Institute (EESI) suggests U.S. data centres already consume around 163.7 billion gallons annually, with large facilities using up to five million gallons per day (EESI Report). With AI workloads accelerating, that figure is only expected to climb.
A Thirst for Intelligence
Cooling is the most visible driver of water demand, but it is not the only one. A report by the IEEE ITCC outlines how AI’s water footprint extends across the lifecycle: generating electricity for servers, fabricating chips, and managing hardware all require water (IEEE Blog). In regions already prone to drought, these withdrawals can strain ecosystems and communities.
The United Nations Environment Programme (UNEP) has warned that unchecked growth of AI data centres risks worsening global water stress, alongside related challenges like e-waste and mineral extraction (UNEP Report). This is not just an engineering issue but a matter of environmental justice.
Measuring the Hidden Flows
Numbers underscore the urgency. A UK government report found that Microsoft’s data centre water use jumped by 34% in 2022 to 6.4 million cubic metres, while Google reported 19.5 million cubic metres the same year. Globally, AI-driven demand could reach 4.2 to 6.6 billion cubic metres annually by 2027—roughly half the UK’s total yearly consumption (UK Government Report PDF).
Academic work echoes these findings. Researchers at the University of California calculated that training GPT-3 consumed an estimated 700,000 litres of freshwater, evaporated in cooling processes. Their paper warns that future growth in AI could intensify global water withdrawals dramatically (ArXiv PDF).
Beyond Carbon Metrics
Carbon accounting is slowly becoming part of corporate disclosures, with AI firms under pressure to show greener energy mixes. But water remains a blind spot. Few companies release comprehensive breakdowns of their withdrawals, and even fewer disclose how consumption is distributed across facilities. Without transparency, communities near data centres struggle to measure the true costs.
This blind spot mirrors earlier debates about AI’s compute scale. Just as Google’s nuclear bet on AI infrastructure highlighted hidden energy dependencies, water scarcity reveals another layer of risk. And like the integration of Gemini into Google Workspace, the expansion of AI into daily life means its environmental costs ripple ever further.
Communities on the Frontline
In many cases, the burden of water-intensive data centres falls hardest on nearby communities. Residents in Arizona, the Netherlands, and Ireland have voiced concerns over new server farms drawing water from already stressed local supplies. For small towns, the arrival of a hyperscale facility can mean difficult trade-offs: jobs and tax revenue on one side, pressure on reservoirs and rivers on the other. Without proper consultation and safeguards, the social contract between tech giants and communities risks fraying.
The same dynamic has been seen in energy debates. Just as wind farms and solar plants have raised questions about land use, water-hungry AI operations demand public engagement and fair distribution of costs and benefits. Civil society groups are increasingly calling for binding environmental impact assessments before construction, not after.
Policy Gaps and Future Pathways
Governments are beginning to pay attention, but regulation lags behind. Some regions have introduced disclosure rules, but these remain patchy and inconsistent. A global standard for water reporting in AI and cloud infrastructure does not yet exist. As a result, companies can trumpet carbon neutrality while quietly expanding water withdrawals that go largely unmeasured.
Policymakers face a balancing act. AI is an economic growth engine and a national security priority, but it is also a voracious consumer of resources. The challenge is to ensure that innovation does not undermine sustainability. Options include mandating annual water disclosures, incentivising investment in dry cooling technologies, and steering new facilities toward less water-stressed geographies.
Toward Sustainable Intelligence
The path forward requires acknowledging water alongside carbon. Cooling technologies can be improved, siting decisions can prioritise less water-stressed regions, and AI companies can commit to disclosing withdrawals as rigorously as emissions. Communities deserve to know the trade-offs when a data centre arrives in their backyard.
Ultimately, the question is not whether AI will consume resources—it will. The real test is whether it can do so in a way that is sustainable, equitable, and transparent. Without that, the thirst for intelligence may come at a cost no algorithm can optimise away.
About the Author
Stuart Kerr is a technology correspondent at LiveAIWire, covering artificial intelligence, climate, and society. His reporting investigates the hidden costs of emerging technologies and their implications for policy, environment, and everyday life. More at About LiveAIWire.