The public debate around artificial intelligence’s water footprint has become fixated on a handful of headline-grabbing figures: 700,000 liters of freshwater “lost” while training GPT-3, a potential global water demand of 4.2–6.6 billion cubic meters by 2027, or the nearly 950 TWh of electricity that data centers are expected to consume by 2030 — roughly as much as the whole of Japan consumes today.

The figures are real and come from reputable sources. The problem is not that they are false, but that they are often presented without context: without considering the cooling technology actually being used, without comparison with other industries and, most importantly, without distinguishing between data centers built ten years ago and those being deployed in 2026.

A question of units before anything else

The first layer of dramatization simply comes down to the unit of measurement chosen for reporting.

Virtually every other form of water consumption — by a household, a farm or a factory — is conventionally expressed in cubic meters (m³). A typical household consumes somewhere around 100–150 m³ of water per year, a figure that almost no one perceives as alarming.

Express the same amount in liters — 100,000–150,000 liters — and the impression changes dramatically, even though the actual quantity of water is identical.

This is precisely the mechanism used, intentionally or otherwise, in most articles about AI and water: the figures are almost always reported in liters rather than cubic meters, making them appear three orders of magnitude larger than they actually are when compared with other household or industrial water consumption figures reported conventionally.

The “700,000 liters” used to train GPT-3, for example, translates into 700 m³ — roughly the annual water consumption of 5–7 typical households. It is a real figure, but one with a completely different visual impact from “700,000.”

The choice of unit does not invalidate the underlying data, but it does explain a significant part of the perception of a “crisis” that has taken hold in the public debate — and it is precisely why the aggregated figures deserve to be examined consistently, alongside the actual technologies behind them.

Where the figures come from — and what they leave out

The 700,000 liters associated with training GPT-3 originate from the field’s landmark study, “Making AI Less Thirsty”, conducted by researchers from the University of California, Riverside, and the University of Texas at Arlington

In a subsequent paper, the same research team estimated that global AI-related demand could reach 4.2–6.6 billion cubic meters of water by 2027 — including both direct server cooling, or “on-site” consumption, and the water used by thermoelectric power plants to generate the electricity consumed by those data centers, or indirect “off-site” consumption.

One detail that is usually missing from media coverage is that the underlying figures largely come from data centers built in arid regions of the United States — Arizona, Texas and Nevada — where water stress is already a structural issue and where conventional evaporative cooling has historically been used extensively. These systems rely on open cooling towers that continuously lose water through evaporation during each cooling cycle.

Extrapolating these figures to every data center, regardless of its cooling technology or the region in which it operates, is one of the main sources of confusion in the public debate.

Where the industry is heading: closed-loop, direct-to-chip and immersion cooling

The essential technical distinction is between open-loop and closed-loop systems.

A conventional evaporative cooling tower operating in an open-loop configuration continuously loses water: water evaporates in order to absorb heat, and those losses must constantly be replenished from the public water network or other sources. This is why the average WUE of a “generic” data center has been reported at approximately 1.8–1.9 liters/kWh by The Green Grid and confirmed by the Environmental and Energy Study Institute.

The industry is rapidly moving toward a different model:

• Closed-loop cooling — the cooling fluid circulates within a sealed circuit and is continuously recirculated, without evaporation through an open cooling tower. The system is filled once during commissioning, after which the water essentially remains within the system.
• Direct-to-chip (D2C) — cold plates installed directly on the motherboard remove heat from CPUs and GPUs, eliminating the need to cool the entire volume of air inside the data hall. This is becoming the standard solution for the high densities required by AI clusters — racks in the 40–130+ kW range — where conventional air cooling is reaching its physical limits.
• Immersion cooling — servers are submerged in a dielectric liquid, almost entirely eliminating the need to cool the air within the server room. It is one of the most aggressive approaches in terms of efficiency. Although large-scale adoption is still at an early stage, deployment is increasing in high-density GPU clusters.

The difference is an order of magnitude, not merely a matter of rhetoric.

Microsoft, one of the relatively few operators publishing audited data, reported an average WUE of 0.27 L/kWh in 2025, compared with 2.3 L/kWh in the early 2000s — an improvement of almost 90% over two decades, achieved largely by moving away from industrial chiller-based cooling toward direct evaporative systems and, more recently, toward designs that use no evaporation at all (Microsoft, June 2026).

The company has announced that its new generation of data centers will use zero-water-evaporation closed-loop cooling systems, filled once during construction and recirculated throughout the facility’s operating life (Microsoft Cloud Blog).

The trend is clearly visible across the entire industry. The high-density racks required by AI clusters — 40 kW and increasingly more than 100–150 kW — are effectively forcing the transition to liquid cooling, particularly closed-loop and direct-to-chip configurations, because conventional air cooling is no longer physically capable of removing such heat densities efficiently.

Operators that have already made this transition report reductions in water consumption of up to 80% compared with conventional evaporative systems, while the industry’s focus is shifting from “how much water do we use?” to “how much water can we recirculate instead of losing?”

Alongside direct-to-chip systems, immersion cooling — in which servers are submerged in a dielectric fluid — represents another emerging technology capable of reducing water requirements even further by almost entirely eliminating the need to cool the air inside the server room.

What WUE do Microsoft, Amazon and Google have today — and what are they doing to reduce it further?

The three major hyperscalers report their figures differently, but the direction is the same: WUE is declining, even as absolute water consumption increases with the construction of new facilities.

Microsoft achieved 0.27 L/kWh in fiscal year 2025, down from 0.30 L/kWh the previous year, and officially announced that it had become “water positive” — replenishing more water than it consumed — five years ahead of its 2030 target (Microsoft, June 2026; Data Center Dynamics).

The company says that approximately 90% of its owned fleet in 2025 operated using cooling systems with low or zero water consumption, with water required less than 5% of the time in cooler regions such as Dublin and Amsterdam and up to 40% of the time in hotter climates such as Phoenix.

At its Phoenix facilities, operational adjustments alone — recalibrated temperature set points, real-time weather analysis and periodic audits — delivered a 23% year-on-year improvement in WUE in 2025.

Microsoft also reports the use of recycled or non-potable water at key locations: 74% of consumption in Quincy, Washington; 99% in Singapore; and 79% in San Antonio.

Amazon Web Services published an absolute global water consumption figure for the first time in June 2026 — 2.5 billion gallons, or approximately 9.5 million m³, in 2025 — alongside a WUE of 0.12 L/kWh, the best figure officially reported by a major hyperscaler. AWS compares this with an industry average of 0.84 L/kWh, making its operations approximately seven times more water-efficient (Data Center Dynamics; aggregated analysis by GPUSmith).

AWS also reported a 2% reduction in water withdrawals at sites it directly owns and operates compared with 2024.

Google is the most widely discussed case because it is the only one of the three whose absolute water consumption has risen visibly alongside its AI expansion: 10.9 billion gallons, or approximately 41 million m³, in 2025 — up 34% from 2024 and more than double the 2021 level, according to its Environmental Report 2026 (Axios; Data Centre Magazine).

The company offsets part of this consumption through water-replenishment projects — 7.7 billion gallons replenished in 2025 through 165 projects across 97 watersheds, equivalent to 78% of its consumption, with a target of reaching 120% by 2030.

It is important to put this into perspective: the increase in Google’s absolute water consumption primarily reflects the expansion of installed capacity — electricity consumption at Google data centers increased by 27% in 2024 alone — rather than a deterioration in efficiency per unit of energy consumed. Nevertheless, it is one of the reasons Google remains a primary target of criticism regarding transparency in its reporting methodology.

The pattern across all three companies is consistent: WUE per kWh continues to decline as new facilities adopt closed-loop and direct-to-chip cooling; absolute water consumption can nevertheless increase simply because the number of data centers and the amount of installed power capacity are growing considerably faster than per-unit efficiency can compensate for.

This is a technical distinction that most media articles systematically omit, and it accounts for much of the public confusion surrounding the issue.

Context: who actually consumes large amounts of water?

Compared with other economic activities, even a modern data center using conventional evaporative cooling is far from being one of the largest water consumers — and a facility operating with closed-loop cooling uses even less:

• Semiconductor fabs consume between 2 million and 10 million gallons of ultrapure water per day, per facility — meaning that a chip already carries a significant water footprint before it ever reaches a server (World Economic Forum).
• The brewing industry: producing a single liter of beer can require up to 25 liters of water across the entire production chain.
• A natural-grass football pitch consumes an average of approximately 100,000 liters of water per day for irrigation alone.
• Golf courses in the United States consume approximately 2.08 billion gallons of water per day for irrigation, compared with approximately 449 million gallons per day used by all U.S. data centers combined for cooling, according to estimates by the Florida Water & Pollution Control Operators Association.
• Agriculture remains by far the world’s largest consumer of freshwater, accounting for approximately 72% of total freshwater withdrawals according to the latest FAO AQUASTAT 2025, followed by industry at approximately 16% and the municipal sector at approximately 12%.

At an aggregate level, all U.S. data centers combined account for less than 1% of the country’s total water consumption, according to estimates from Lawrence Berkeley National Laboratory — including both indirect consumption through electricity generation and direct consumption through cooling.

Even the more pessimistic projections from the International Energy Agency, which foresee global data center electricity consumption rising from 485 TWh in 2025 to approximately 950 TWh in 2030 — close to 3% of global electricity consumption — do not fundamentally change the order of magnitude.

Even under an accelerated AI-driven growth scenario, the sector remains several orders of magnitude below other industries in terms of pressure on freshwater resources.

What the European industry says about the issue

Independent confirmation of this perspective comes from within the European data center industry itself.

In a position paper specifically dedicated to the subject — “Seeing Through the Mist: Data Centres and Water Usage” — the European Data Centre Association (EUDCA), based on its own industry data, reaches the same structural conclusion.

The report introduces a technical distinction that most public comparisons overlook: the difference between water usage — the total volume of water withdrawn from a source for use — and water consumption — the portion of that water that is not returned to the source because it has been lost through evaporation or contamination.

A closed-loop system may have water usage comparable to an evaporative system, while its actual water consumption can be close to zero because the water is treated and returned to the circuit or source.

This is precisely the distinction that makes many of the “headline” figures presented in the media misleading when water usage and water consumption are treated as synonymous.

In terms of comparative context, EUDCA cites figures similar to those discussed above, but expressed consistently in cubic meters: an average U.S. golf course consumes approximately 431,537 m³ of water per year, compared with 416,395 m³ for an average 1–5 MW data center — equivalent to the annual water consumption of around 1,000 households.

Across Europe as a whole, the European Environment Agency figures cited by EUDCA show that water withdrawals are distributed as follows: 36% for power-plant cooling, 29% for agriculture, 19% for public water supply — including households and tourism — and 14% for manufacturing. Data centers and other “emerging users,” such as hydrogen production, remain within the remaining 2% not covered by these major categories.

From a technical perspective, the report confirms the pattern described above: the average WUE across European colocation facilities was 0.31 L/kWh in 2025, below the 0.4 L/kWh target established under the Climate Neutral Data Centre Pact, while hyperscalers generally report even better performance due to standardized and higher operating temperatures.

Final considerations

The figures circulating in the public debate about the water supposedly “lost” by AI are not fabricated, but they are almost always decontextualized: they originate from studies modeling older infrastructure, located in arid regions and relying on open evaporative cooling.

Although data centers can be significant water consumers at the local level in certain locations, the sector cannot be classified as a water-intensive industry in the same sense as agriculture, power generation or manufacturing.

At the same time, the rapid pace of development across the data center industry requires careful monitoring and transparent reporting of the resources being used, with a clearly stated objective of moving toward near-zero consumption of potable water and significantly greater water reuse.