OpenAI published its first comprehensive environmental impact report this month, covering energy consumption, carbon emissions, and water usage across its data centre operations for the 2025 calendar year.
The headline water figure: 6.1 billion litres - a 280% increase from the company's estimated 2024 consumption, driven by the expansion of training infrastructure for GPT-5 and the scaling of inference capacity to serve over 300 million weekly active users.
Water is used in data centres primarily for cooling. Servers generate heat; evaporative cooling systems use water to dissipate it. The process is efficient but consumptive - the water evaporates and is not returned to the local supply.
What has drawn scrutiny is where that water is being consumed. Approximately 2 billion litres - a third of the total - was drawn from regions classified as "high" or "extremely high" water stress by the World Resources Institute. This includes data centres in West Des Moines, Iowa, and The Dalles, Oregon - both communities that have experienced drought conditions and water rationing in recent years.
In West Des Moines, residents near the Microsoft-operated data centre campus that processes OpenAI workloads reported a 12% increase in municipal water rates in 2025. The city council attributed the increase to infrastructure upgrades needed to meet growing demand, though it did not specifically cite the data centre.
"The water doesn't disappear from the planet. It evaporates into the atmosphere," said an OpenAI spokesperson in response to questions. "Our facilities operate within permitted water allocations in every jurisdiction."
This is technically true. Data centre operators apply for and receive water use permits from local authorities. But environmental advocates argue that the permitting process does not adequately account for cumulative impact.
"When you permit one data centre, the water impact is manageable," said Dr. Kimberly Pham, a water policy researcher at the Pacific Institute. "When you permit ten in the same watershed over three years, which is what's happened in central Iowa, the cumulative draw changes the hydrology. The permits were issued individually. No one assessed the aggregate."
Google and Meta have published similar figures. Google's 2025 environmental report disclosed 5.2 billion litres of water consumption. Meta reported 3.8 billion litres. Microsoft, which operates data centres for both its own Azure platform and OpenAI, reported 7.3 billion litres across all operations.
The industry's total water footprint is growing faster than its efficiency improvements. Data centre water usage efficiency (WUE) has improved by approximately 8% per server, but the total number of servers has increased by over 40%.
Several data centre operators, including Google and Microsoft, are investing in waterless cooling systems (air cooling, liquid immersion cooling) for new facilities. But retrofitting existing data centres is expensive and slow, and the fastest-growing AI workloads are running on existing infrastructure.
What we know for certain
OpenAI's data centres consumed 6.1 billion litres of water in 2025, a 280% increase from 2024. Approximately a third came from drought-affected regions. West Des Moines municipal water rates increased 12%. Google, Meta, and Microsoft disclosed comparable figures.
What we are inferring
The permitting process for data centre water use is not accounting for cumulative impact in regions with multiple facilities. Industry-wide water consumption is growing faster than efficiency gains. The shift to waterless cooling will take years to materially change the trajectory.
What we couldn't verify
Whether the West Des Moines water rate increase was directly caused by data centre demand or by other infrastructure factors. The exact breakdown of OpenAI vs. Microsoft workloads at shared facilities. Whether any community has experienced water shortages directly attributable to data centre operations.