The Cloud Is Not Weightless

Photo by Geoffrey Moffett on Unsplash

As temperatures in central Iowa climbed past thirty-five degrees Celsius, Microsoft’s cluster of data centres west of Des Moines pumped 11.5 million gallons of water from the watershed of the Raccoon and Des Moines rivers. The purpose of this extraordinary draw was not agriculture, not municipal supply, not firefighting. It was cooling. Somewhere inside those warehouse-sized buildings, thousands of GPU servers were training GPT-4, the large language model that would soon be marketed as a revolution in human productivity. The West Des Moines Water Works, alarmed by a single corporate client consuming six per cent of the district’s water during the hottest month of the year, issued what amounted to an ultimatum: future data centre projects would be approved only if Microsoft could demonstrate technology to “significantly reduce peak water usage.” The residents of West Des Moines were, in effect, being asked to share their drinking water with a machine learning how to write poetry. Nobody had asked them whether they found that trade acceptable.

The metaphors we use for digital technology are almost uniformly ethereal. The cloud. The stream. Wireless. These words perform an act of dematerialisation so thorough that most users of AI have no idea their queries have a physical substrate, let alone a hydrological one. But every computation generates heat, every hot server requires cooling, and the dominant cooling method for data centres is evaporative: water is pumped through systems that absorb thermal energy, convert it to vapour, and vent it into the atmosphere. The word “consume” is precise here and worth pausing over. Unlike water used in, say, a household washing machine, which can be treated and returned to the supply, roughly eighty per cent of the water drawn into data centre cooling towers evaporates. It does not come back. It enters the atmosphere as water vapour and, depending on local conditions, may not precipitate anywhere near where it was extracted.

This vanishing act has quantifiable dimensions. A 2024 report from the Lawrence Berkeley National Laboratory estimated that U.S. data centres consumed seventeen billion gallons of water directly through cooling in 2023, with projections suggesting that figure could double or quadruple by 2028. But the direct cooling number captures only part of the story. The same report estimated an additional 211 billion gallons consumed indirectly, through the water used by the power plants that generate the electricity data centres require. When you account for both the water evaporating in the cooling tower and the water turning to steam at the coal or gas plant feeding it electricity, a single large data centre’s total water footprint becomes staggering. The Environmental and Energy Study Institute reports that the largest facilities consume up to five million gallons per day, equivalent to the domestic water supply for a town of ten to fifty thousand people.

The argument of this essay is that the water footprint of artificial intelligence represents not a minor externality but a structural feature of the AI economy, one that accelerates climate stress, deepens environmental injustice, and exposes the fundamental dishonesty of an industry that markets its products as “clean” while externalising their most elemental physical costs onto communities least equipped to absorb them. The cloud, it turns out, is not weightless. It is wet, and it is getting wetter precisely as the planet gets drier.

The peer-reviewed research that first brought AI’s water consumption into public view emerged from the University of California, Riverside, and the University of Texas at Arlington. In a study published in Communications of the ACM, Pengfei Li, Shaolei Ren, and their collaborators calculated that training GPT-3, with its 175 billion parameters, consumed approximately 700,000 litres of freshwater on-site in Microsoft’s U.S. data centres, with total water consumption (including the electricity supply chain) reaching 5.4 million litres. For inference, the ongoing process of actually running the model and generating responses, they estimated that every twenty to fifty medium-length exchanges required the equivalent of a standard 500-millilitre bottle of water. This figure has since been contested. Critics note that it assumed longer responses than typical users generate, and that newer, more efficient models consume less per query. But even the most conservative recalculations place the per-query cost in the range of five to twenty-five millilitres: small for any individual prompt, enormous when multiplied by the billions of queries processed daily across platforms like ChatGPT, Gemini, and Claude.

The corporate sustainability reports tell a corroborating story, though the companies themselves would prefer a gentler reading. Microsoft’s global water consumption spiked thirty-four per cent from 2021 to 2022, reaching nearly 1.7 billion gallons. Google’s rose twenty per cent over the same period, with its data centres consuming approximately 5.6 billion gallons in 2022 and 6.1 billion gallons in 2023. Shaolei Ren, the UC Riverside researcher whose team produced the foundational estimates, told the Associated Press plainly: “It’s fair to say the majority of the growth is due to AI.” A December 2025 study by Alex de Vries published in the journal Joule estimated that the water footprint of AI systems alone could reach 312 to 765 billion litres in 2025, a range whose upper bound approaches the global annual consumption of bottled water.

These numbers deserve a moment of critical scrutiny. One difficulty is that no major technology company separately reports its AI-related water consumption from its non-AI workloads. Google’s sustainability reports do not distinguish between the water used to serve a YouTube video and the water used to generate a Gemini response. Microsoft bundles Azure cloud services, Xbox infrastructure, and OpenAI partnership workloads into a single figure. This opacity is not accidental. It insulates the companies from accountability for the specific environmental cost of their most aggressively marketed product category. When an industry cannot or will not disaggregate its own impacts, independent researchers are forced to estimate, and estimates carry uncertainty that corporations are only too happy to exploit when the numbers prove inconvenient.

The relationship between AI water consumption and climate change is not merely additive. It is recursive. Data centres consume water partly because they consume electricity, and much of that electricity is generated by thermoelectric power plants that themselves consume water. As global temperatures rise, data centres require more cooling. As they require more cooling, they consume more water and more electricity. As they consume more electricity, the power plants supplying them consume more water and emit more carbon. As carbon emissions drive temperatures higher, the cycle tightens. This is a positive feedback loop in the thermodynamic sense and a catastrophic one in every other.

The geographic dimension makes the feedback loop vicious in ways that aggregate statistics obscure. A study by the Houston Advanced Research Center and the University of Houston projected that data centres in Texas alone will consume 49 billion gallons of water in 2025, potentially rising to 399 billion gallons by 2030. Texas, a state whose water infrastructure has already buckled under consecutive droughts. In Arizona, where data centre construction is accelerating, acquifers are declining and the Colorado River system is under existential stress. An MSCI analysis of roughly fourteen thousand data centre assets worldwide found that one in four existing facilities may face significantly more water-scarcity days by 2050. The industry is, with remarkable consistency, building its most water-intensive infrastructure in the places least able to sustain it.

In Newton County, Georgia, a Meta data centre opened in 2018 that consumes 500,000 gallons of water per day, fully ten per cent of the entire county’s water supply. The county continues to field requests for new permits, some of which would use up to six million gallons daily, more than doubling the county’s total current consumption. In Northern Virginia, the self-proclaimed “data centre capital of the world,” all data centres collectively consumed close to two billion gallons in 2023, a sixty-three per cent increase from 2019. In Aragón, Spain, Amazon’s largest European data centre requires 500 million litres of drinking water annually for cooling. When the company requested a forty-eight per cent increase in its water permits in December 2024, citing the need for more cooling capacity as climate change raises temperatures, the protest movement Tu Nube Seca Mi Río (“Your Cloud Is Drying My River”) called for a moratorium on new data centres. Months later, Aragón petitioned the European Union for drought aid.

The cooling tower, visible and auditable, is only the most obvious site of water consumption. Semiconductor manufacturing, the process that produces the chips powering every AI server, is itself extraordinarily water-intensive. Fabricating microprocessors requires ultrapure water for cleaning, etching, and rinsing at each stage of production. Approximately 1,500 gallons of piped water are needed to produce 1,000 gallons of the ultrapure variety, and a typical chip fabrication plant consumes nearly ten million gallons per day. Every GPU in every data centre arrives with an embedded water debt that no sustainability report currently accounts for. This scope-3 water footprint, the water consumed upstream in the manufacturing supply chain, is a blind spot in every major technology company’s environmental accounting.

There is also the question of what kind of water is being consumed. Despite the existence of treated wastewater and reclaimed water alternatives, the majority of data centres use potable water. The reasons are partly technical (reclaimed water can cause corrosion and microbial growth in sensitive equipment) and partly economic (potable water is cheaper and more reliably available through municipal systems). But the consequence is that data centres are drawing from the same supply that serves hospitals, schools, and households, and doing so at volumes that alter the calculus of municipal water planning. Loudoun County, Virginia, home to roughly two hundred data centres, now derives such a large proportion of its tax revenue from data centre operators that its board of supervisors is considering adjusting tax rates to avoid fiscal dependency on a single industry. The political economy of water has become, in these communities, inseparable from the political economy of computation.

Every major technology company has pledged to become “water positive” by 2030, a formulation that sounds reassuring until you examine what it means. Water positivity, as defined by the industry, involves replenishing more water than the company consumes, typically through investments in watershed restoration, reforestation, or water access projects in developing countries. The structural problem is one of fungibility. Water replenished in a reforestation project in Brazil does not replace water evaporated from a cooling tower in Iowa. Hydrological systems are local. A litre restored to a watershed in one hemisphere does not rehydrate the acquifer being drained in another. The accounting may balance on a corporate spreadsheet. It does not balance in the physical world.

Google, which previously claimed “100% renewable” energy through annual matching of renewable energy credits, has since shifted to reporting actual hourly carbon-free energy, achieving sixty-six per cent in 2024, an acknowledgement that annual matching obscured rather than resolved the gap between aspiration and reality. No comparable reckoning has occurred for water. The Li et al. study identified a particularly perverse tension: the hours and locations most efficient for reducing carbon emissions (midday, when solar energy peaks) are precisely the hours and locations least efficient for water consumption (midday, when temperatures are highest and evaporative cooling demands are greatest). Optimising for carbon and optimising for water pull in opposite directions. An industry that claims to be solving for sustainability has, in practice, been solving for one metric at the expense of another, and disclosing only the metric that flatters.

The distributional dimension of this problem deserves unflinching attention. Data centres are disproportionately sited in communities with lower land costs, weaker regulatory frameworks, and less political leverage. The tax revenues they generate can be significant; Loudoun County expects nearly $900 million in data centre property taxes in fiscal year 2025, approaching the county’s entire operating budget. But the employment effects are minimal, as data centres require remarkably few permanent workers, and the water costs are borne by the entire community. This is a familiar pattern in extractive industries: concentrated benefit, diffuse cost. The difference is that the extraction here is invisible. No mine shaft, no smokestack, no pipeline. Just a nondescript warehouse humming in a field, quietly converting a town’s water supply into the sensation of artificial intelligence.

The global picture is sharper still. An investigation by SourceMaterial and The Guardian found that thirty-eight active data centres owned by three major firms operate in parts of the world already facing water scarcity, with another twenty-four under development in similar regions. In Santiago, Chile, Google’s data centre in Cerrillos has faced court challenges over its consumption of potable water during severe drought. The pattern echoes the extractive logic that has characterised the relationship between the Global North and the Global South for centuries: resources flow outward, serving consumption demands elsewhere, while the environmental costs accumulate locally. The cloud may be global. The drought is always local.

None of this constitutes an argument against artificial intelligence as such. The technology has genuine applications in climate modelling, water system optimisation, agricultural efficiency, and drought prediction that could, if properly governed, mitigate the very crises it currently accelerates. The irony is precise and worth naming: the most water-intensive information technology in human history could be deployed to solve water scarcity, if the economic structures governing its development prioritised that outcome over quarterly returns. But they do not. And the question that should trouble us is not whether AI uses water (it does, in volumes that defy the industry’s own sustainability rhetoric) but why the communities whose water is being consumed had no seat at the table when the decision was made. When the cloud drinks your river, who gave it permission, and who profits while you go thirsty?