You have seen the headlines. “AI is draining our reservoirs.” “ChatGPT drinks a bottle of water per prompt.” “Data centers will leave your taps dry.” The narrative is everywhere, and it sounds alarming. But when you actually follow the numbers, a very different picture emerges.

The claim that AI and data centers consume a dangerous amount of water is, to put it directly, not supported by the evidence.

The Numbers Behind the Noise

Let us start with the scale of the problem as presented. A widely cited 2023 study from the University of Colorado Riverside suggested that training GPT-3 consumed roughly 700,000 liters of water. That number made headlines. It sounds enormous. But context changes everything.

In 2023, all U.S. data centers combined — the vast majority of which run internet infrastructure, not AI — used between 200 and 250 million gallons of freshwater per day. The United States consumes approximately 132 billion gallons of freshwater per day. That means all data centers account for roughly 0.2% of national freshwater consumption. AI-specific workloads represent a fraction of that: approximately 0.008% of total U.S. freshwater use, or about 10.6 million gallons per day.

To make that concrete: all AI workloads in all American data centers use about eight times as much water as the local water utility in a town of 16,000 people provides to its residents. That is not nothing, but it is a far cry from a crisis.

Most of That Water Is Not “Consumed”

The word “consumed” does most of the heavy lifting in these headlines. Water use falls into two categories:

• Consumptive use: Water removed from a local system, primarily through evaporation.
• Non-consumptive use: Water drawn, used for cooling, and returned to its source essentially unchanged.

For data centers, the breakdown is approximately:

Category	Share of AI Water Use	What Happens
Non-consumptive, indirect (power plants)	~90%	Returned to source
Consumptive, indirect (power plant evaporation)	~7%	Lost to atmosphere
Consumptive, direct (data center onsite)	~3%	Lost onsite

The onsite water that data centers actually consume is treated to potable standards — it is drinking water. The irony of the “AI is drinking our water” narrative is that the water data centers use is the same water that comes out of your tap, and the amount is trivial compared to agricultural, industrial, and residential use.

The Personal Scale Puts It in Perspective

One AI prompt uses between 0.3 mL and 2 mL of water (onsite plus offsite combined). The average American’s total consumptive water footprint — mostly from food and manufacturing — is about 422 gallons per day. That means in a single day, you consume enough water to power roughly 800,000 chatbot prompts.

Consider the water embedded in everyday objects:

• A pair of leather shoes: equivalent to 4,000,000 AI prompts
• A smartphone: equivalent to 6,400,000 AI prompts
• A pair of jeans: equivalent to 5,400,000 AI prompts
• A cotton t-shirt: equivalent to 1,300,000 AI prompts

If you want to save a lifetime’s worth of chatbot prompts, skipping a single pair of jeans does more than quitting AI entirely.

What About Local Impact?

The most compelling version of the water scarcity argument is local: a data center moves into a water-stressed region and strains the supply. This has happened, but it is rare and often misrepresented.

There is essentially one documented case — Newton County, Georgia — where data center expansion contributed to rising water costs. Even there, it was one factor among several, including inflation, housing growth, and other industries.

In most documented cases, data centers improve local water infrastructure:

• The Dalles, Oregon: Google funds water system upgrades as part of its development agreement.
• Council Bluffs, Iowa: Google pays for expanded water treatment capacity.
• Quincy, Washington: Microsoft co-funded a $31 million water reuse utility, increasing regional resilience.
• Goodyear, Arizona: Microsoft invested $40-42 million to expand wastewater capacity.
• Umatilla/Hermiston, Oregon: AWS recycles approximately 96% of its cooling water and provides it to local farmers at no charge.

Some widely reported incidents were not about operational water use at all. The New York Times headline “Their Water Taps Ran Dry When Meta Built Next Door” described a data center that was not yet operational and did not draw groundwater. The issue was construction sediment, not cooling systems.

The Cooling Systems That Change the Equation

The water narrative also ignores how rapidly cooling technology is improving. Modern data centers are deploying systems that minimize or eliminate water consumption entirely.

Scaleway: Adiabatic Cooling at Scale

Scaleway’s DC5 data center in France uses an adiabatic cooling system that replaces traditional mechanical air conditioning with free cooling enhanced by water evaporation. Outside air is humidified with pure water to drop its temperature by nearly 10 degrees Celsius before being injected into cold aisles. The system is managed by 2,200 sensors and a real-time algorithm that adapts every 17 milliseconds.

The result: a Water Usage Effectiveness (WUE) below 0.15, compared to the industry average of 1.8. That is more than a 10x improvement. A 6 MWh ice storage unit provides backup cooling and peak load shifting, and the entire system operates with zero carbon emissions from cooling.

CSCS: Lake Water and Waste Heat Recovery

The Swiss National Supercomputing Centre (CSCS) in Lugano uses Lake Lugano as a natural cooling source, drawing water from 45 meters depth where it sits at approximately 6 degrees Celsius. Three pumps move up to 760 liters per second over 2.8 kilometers and up 30 meters in elevation.

The system uses a dual-circuit design: the first circuit cools high-performance supercomputers (up to 14 MW), while the second circuit uses the warmed water at 16-17 degrees Celsius for lower-density systems and office buildings. The heated water exiting at 15-20 degrees Celsius is reused to heat the CSCS office building in winter and supplies hot water to the city of Lugano. Micro-turbines at the pumping station generate electricity from the returning water’s potential energy, covering over 30% of the pumping station’s own energy needs.

Infomaniak: Zero Water Consumption, 100% Heat Reuse

Infomaniak’s D4 data center in Geneva goes further: it uses zero water for cooling. The system is a chiller-free, closed-loop solution that captures 100% of the energy consumed by servers and repurposes it for heating.

Two Trane XStream RTWF heat pumps capture low-temperature waste heat (40-45 degrees Celsius) from servers, inverters, and ventilation. They upgrade it to 67 degrees Celsius in summer for domestic hot water and up to 85 degrees Celsius in winter for district heating. The cold released during compression cools the servers to approximately 28 degrees Celsius. At full capacity, the system recovers 1.7 MW continuously, enough to heat 6,000 homes in winter or provide 20,000 showers daily in summer.

The data center achieves a Power Usage Effectiveness of 1.09 and an Energy Reuse Factor of 0.95. It avoids 3,600 tons of CO2 equivalent annually from natural gas displacement.

These are not prototypes. These are operational facilities running today.

The Economic Argument

Even setting aside the engineering, the economic case against the water narrative is strong. Data centers generate approximately 50 times more tax revenue per gallon of water than golf courses in Maricopa County, Arizona. Golf courses in that county use 29 billion gallons of water per year. All data centers in the county use 0.12% of the region’s total water supply.

If Arizona replaced all its golf courses with data centers using the same water, it could gain an estimated $42 billion per year in tax revenue — nearly double its 2023 state tax revenue. In water-scarce regions, data centers are among the most economically productive uses of water available.

Why the Narrative Persists

If the numbers are this clear, why does the water scare persist? Several reasons:

1. Large numbers without context. “700,000 liters” sounds scary. “0.008% of national freshwater use” does not. Headlines favor the scary number.

2. Conflation of electricity and water. Most data center water use is indirect, through power generation. When articles cite water figures that include power plant cooling, they are attributing the power grid’s water use to AI specifically.

3. Genuine concern about AI’s environmental impact. People who care about sustainability are looking for leverage points. Water is tangible and emotionally resonant. But directing concern at data center water use specifically, rather than at the far larger water footprints of agriculture, textiles, and manufacturing, misallocates that concern.

4. Local politics. In communities where data centers are being built, water is sometimes a proxy for broader concerns about growth, change, and corporate presence. That is a legitimate political conversation, but it should be grounded in actual water data.

The Bottom Line

AI and data centers use water. So does everything else. The claim that AI poses a unique or disproportionate threat to water resources does not survive contact with the actual data.

At the national level, AI accounts for roughly 0.008% of U.S. freshwater consumption. At the personal level, your daily water footprint is equivalent to hundreds of thousands of AI prompts. At the local level, data centers more often improve water infrastructure than strain it. And at the engineering level, the most advanced cooling systems in the world — from Scaleway’s adiabatic design to Infomaniak’s zero-water closed loop — are proving that high-density computing and water efficiency are not in conflict.

The next time you see a headline about AI drinking your water, remember: the glass is mostly full, and AI barely took a sip.

---

Sources:

• Andy Masley, “The AI Water Issue Is Fake,” blog.andymasley.com, October 2025
• IEEE Spectrum, “The Real Story on AI Water Usage at Data Centers,” spectrum.ieee.org
• Scaleway DC5 adiabatic cooling system documentation
• CSCS (Swiss National Supercomputing Centre) lake water cooling infrastructure
• Infomaniak D4 data center closed-loop heat recovery system