Water usage effectiveness is the single number sitting underneath this week’s claim that AI data centers are less thirsty now. On 7 September 2026 an AFP report carried statements from Nvidia, Microsoft, Amazon Web Services and Meta arguing that the cooling water problem is being engineered away. The evidence they offered was, almost entirely, a ratio: litres of water consumed for every kilowatt-hour of computing delivered.
That ratio is real and it is genuinely improving. AWS now reports 0.12 litres per kilowatt-hour. Microsoft reports 0.27. Meta reports 0.19. Nvidia says its DSX reference design can take a facility to near zero. Those are not marketing inventions; they are published figures with methodologies behind them.
But a ratio is not a volume. Water usage effectiveness measures how hard each litre works, not how many litres are drawn. Both Microsoft and AWS improved the ratio between 2022 and 2025 while their total water use went up, because the denominator — compute — grew faster than the efficiency gain. Rystad Energy still expects global data center cooling demand to rise from 222 billion litres in 2025 to as much as 644 billion by 2030.
This article takes the claim seriously and then tests it. It covers what the metric measures, what each operator has actually published, where the closed-loop engineering genuinely works, what it costs in electricity, and what the number leaves out. If you buy compute, run infrastructure, or answer sustainability questions in a tender, the Data Center Operations implications are at the end.
Table of contents
- What Water Usage Effectiveness Actually Measures
- The Water Usage Effectiveness Numbers the Tech Giants Just Published
- Why Water Usage Effectiveness Improves While Total Water Rises
- Nvidia’s 45C Closed Loop and the Water Usage Effectiveness Ceiling
- The Water Usage Effectiveness Trade: Litres Saved, Kilowatt-Hours Spent
- Water Usage Effectiveness Ignores the Water You Cannot See
- Rystad’s 2030 Projection and the Water Usage Effectiveness Gap
- Why Water Usage Effectiveness Reporting Cannot Be Compared Yet
- What Water Usage Effectiveness Means for AI Buyers
- Water Usage Effectiveness Questions to Put to Your Cloud Provider
- Water Usage Effectiveness FAQ
- References
What Water Usage Effectiveness Actually Measures
Water usage effectiveness, usually shortened to WUE, is a facility-level efficiency metric. It was defined by The Green Grid as a companion to power usage effectiveness, and it answers one narrow question: for every unit of energy the IT equipment consumes, how much water does the site consume to keep that equipment cool?
The formula is deliberately simple. Annual site water consumption in litres, divided by annual IT equipment energy consumption in kilowatt-hours. A facility that consumes 1.2 million litres and delivers 10 million kilowatt-hours of IT load reports a water usage effectiveness of 0.12 L/kWh.
The Formula Behind the Metric
Simplicity is the metric’s strength and its weakness. Because water usage effectiveness has one numerator and one denominator, it can be gamed at either end without anyone lying. Shrinking the numerator by moving cooling load onto chillers lowers the ratio. Growing the denominator by packing more GPUs into the same hall lowers it too, even if the site draws more water than it did last year.
Consumption Versus Withdrawal
The distinction that causes most confusion is consumption against withdrawal. Withdrawal is the volume a site takes from a source. Consumption is the volume that does not go back — mostly what evaporates from a cooling tower. Amazon disclosed 2.5 billion gallons withdrawn across its global data center estate in 2025, a figure much larger than its consumption number, and the two are routinely quoted as if they were the same thing.
Why the Denominator Matters
A water usage effectiveness figure is only comparable when the denominator is defined identically. Some operators use IT load. Some use total facility energy, which is larger and therefore flatters the ratio. Because power usage effectiveness sits at 1.17 across Microsoft’s global fleet in FY25, switching the denominator from IT load to facility load divides the same water volume by a number 17% larger, cutting the reported figure by around 15% without a single litre changing hands.
The Water Usage Effectiveness Numbers the Tech Giants Just Published
Here is what the operators actually report, compiled by Rystad Energy alongside its own demand modelling. The spread is wide enough that the phrase “industry average” carries very little meaning.
| Operator | Reported WUE (L/kWh) | Period | Note |
|---|---|---|---|
| Amazon Web Services | 0.12 | 2025 | 52% better than its 2021 baseline |
| Meta | 0.19 | 2024 | Closed-loop designs, no net water loss claimed |
| Microsoft | 0.27 | FY2025 | Improved from 0.30 in FY2024 |
| Digital Realty | 0.59 | 2025 | Colocation, mixed tenant load profiles |
| Industry average cited by AWS | 0.84 | Comparison basis | Used to support a “7x more efficient” claim |
| Equinix | 0.91 | 2025 | Large legacy colocation footprint |
Reading the League Table
Scaled against the highest reported figure, the gap between the best and worst water usage effectiveness result in this table is close to eightfold. That is not a rounding difference between comparable facilities; it is the difference between a hyperscale campus built around dry coolers and a twenty-year-old colocation hall built around evaporative towers.
The Averages Nobody Agrees On
AWS frames its 0.12 figure as roughly seven times better than an industry average of 0.84 L/kWh. The arithmetic holds — 0.84 divided by 0.12 is exactly seven — but the average itself is the contested part. There is no regulator collecting these figures, no mandatory disclosure boundary, and no audit requirement, so a water usage effectiveness average is assembled from whoever chose to publish.
Why Water Usage Effectiveness Improves While Total Water Rises
This is the part of the story that gets lost in the headline. AFP reported that Microsoft improved its water efficiency by 25% between 2022 and 2025, and AWS by 37% over the same window — and that both companies used more water in total across that period.
Efficiency Is a Rate, Not a Budget
Nothing about those two facts is contradictory. Water usage effectiveness is litres per kilowatt-hour. If the ratio falls by a third while the compute fleet more than doubles, absolute consumption still climbs. An operator can honestly report its best-ever efficiency figure in the same year it draws its largest-ever volume, and both statements will survive an audit.
Microsoft and AWS Both Grew
Microsoft’s own disclosure shows the ratio moving from 0.30 L/kWh in FY2024 to 0.27 in FY2025, a 10% year-on-year improvement in the metric. Amazon reports a 2% decrease in water withdrawn at owned and operated sites between 2024 and 2025 against a 2.5 billion gallon total. Those are real gains at the margin. They are not a reversal of the trend line, and the operators have not claimed they are.
The Growth Term Dominates
The uncomfortable arithmetic is that efficiency has to improve faster than capacity grows just to hold volumes flat. Rystad’s central projection assumes exactly the opposite: efficiency improves, capacity grows faster, and the total nearly triples. Improving water usage effectiveness is necessary. On its own, it is not sufficient.
Nvidia's 45C Closed Loop and the Water Usage Effectiveness Ceiling
The strongest engineering claim in the story comes from Nvidia. Its DSX reference design for AI factories circulates a sealed mixture of roughly 75% water and 25% propylene glycol directly across the chips, filled once and run closed for the life of the facility.
What the DSX Reference Design Changes
The trick is temperature. Coolant enters the rack at 45C rather than the roughly 32C typical of closed-loop systems in 2024, and leaves at around 55C. Because the loop is already hot, the heat can often be rejected to outdoor air through dry coolers rather than evaporative cooling towers. Ali Heydari, Nvidia’s director of data center cooling and infrastructure, describes the result as zero water consumption with chillers running “outside of maybe 1% of the year”.
The Numbers Behind the Claim
Nvidia puts conventional cooling-tower systems at roughly 2.6 million gallons of water per megawatt per year, and says the new design cuts that toward zero. It also estimates that a 50-megawatt hyperscale facility can save more than $4 million a year in combined cooling energy and water costs, on the industry rule of thumb that raising chiller temperatures by 1C cuts cooling energy by about 4%.
Where the Loop Actually Ends
The important caveat is Nvidia’s own. Zero water consumption applies to dry-cooler designs “in the right geographic locations”. The company contrasts the Scottish Highlands with Phoenix, Arizona for a reason. A sealed loop still has to dump its heat somewhere, and in a hot climate that means either mechanical chillers burning more electricity or evaporative assistance burning water. The loop is closed; the building is not.
| Cooling approach | Water impact | Electricity impact | Climate dependency |
|---|---|---|---|
| Evaporative cooling tower | Highest; around 2.6m gallons per MW-year | Lowest | Works almost anywhere |
| Closed loop at ~32C inlet | Low, but often needs evaporative assist | Moderate | Moderate |
| Closed loop at 45C inlet with dry coolers | Near zero in suitable climates | Low to moderate | High; fails in hot regions |
| Mechanical chillers | Near zero | Highest | None |
| Free air cooling | None while conditions allow | Lowest | Very high; seasonal |
The Water Usage Effectiveness Trade: Litres Saved, Kilowatt-Hours Spent
Andy Masley, an independent researcher who works on AI and data centers, put the physics plainly to AFP: there is a fairly direct trade-off between how much water is used and how much energy is used. Heat has to leave the building, and you pay for that in one currency or the other.
The Dry-Cooling Penalty
Rystad quantifies the exchange rate. Switching to dry cooling saves roughly 2.15 litres of water per kilowatt-hour of IT load, and costs an additional 0.30 to 0.74 kilowatt-hours of electricity to do it. Whether that is a good trade depends entirely on local water scarcity and on how the local grid generates its power, which is why a single global water usage effectiveness target would be a bad idea.
Climate Decides the Price
In a cool, humid region the penalty is small and the water saving is close to total. In Phoenix or Jakarta the same design either fails or falls back on chillers. AWS itself illustrates the spread: its regional water usage effectiveness ranges from 0.02 L/kWh in Stockholm to 2.85 L/kWh in Jakarta — a 142-fold difference inside one company operating one playbook.
Cheap Water Weakens the Incentive
Shaolei Ren, an engineering professor at UC Riverside, gave AFP the commercial reason progress has been uneven: because water is generally much cheaper than electricity, companies have less financial incentive to cut water use. On his reading, the recent acceleration is driven at least as much by public backlash and planning objections as by cost. That matters, because reputational pressure is reversible in a way that a price signal is not.
Water Usage Effectiveness Ignores the Water You Cannot See
The metric’s boundary stops at the fence line. Everything upstream of the meter — the water evaporated by the power stations feeding the site, and the ultrapure water consumed fabricating the chips inside it — sits outside the ratio entirely.
The Electricity Footprint
IEEE Spectrum’s analysis of a single GPT-3 text response puts total water at 16.9 millilitres: 2.2 millilitres for on-site cooling and 14.7 millilitres for the electricity that produced the answer. That is 87% of the footprint outside the building. Rystad reaches a similar conclusion at national scale, noting that in the United States indirect water consumption from electricity generation can exceed direct consumption by more than twofold.
The Fabrication Footprint
Semiconductor fabrication is the other omission. Producing the accelerators that fill an AI hall consumes very large volumes of ultrapure water, and none of it appears in any operator’s water usage effectiveness disclosure. A site can drive its ratio to zero and still sit at the end of a supply chain with a substantial water bill attached.
What a Complete Number Would Include
A defensible figure would report on-site consumption, the water embedded in purchased electricity, and the fabrication footprint of the hardware, split by watershed. Nobody publishes that. Until somebody does, water usage effectiveness is a useful operational metric being asked to carry an environmental argument it was never designed to support.
Rystad's 2030 Projection and the Water Usage Effectiveness Gap
Rystad Energy’s Tech Thirst analysis is the source of the volume figures in the AFP story, and it is the most useful counterweight to the efficiency claims because it models both at once.
Three Scenarios, One Baseline
Data centers consumed 222 billion litres directly for cooling in 2025. Rystad’s risked central case takes that to just under 644 billion litres by 2030. A more water-efficient pathway lands at 543 billion, and an aggressive mitigation case at 388 billion. Even the best case is a 75% increase on today.
| Scenario | 2030 volume | Multiple of 2025 | What it assumes |
|---|---|---|---|
| 2025 actual | 222bn litres | 1.0x | Baseline; 59bn US gallons |
| 2030 best case | 388bn litres | 1.75x | Aggressive adoption of low-water cooling |
| 2030 moderate case | 543bn litres | 2.45x | More water-efficient pathways, partial rollout |
| 2030 risked central case | 644bn litres | 2.90x | Current trajectory, no step change |
Water-Stressed Regions Carry the Risk
The distribution matters more than the total. Rystad expects water-stressed regions to account for 34% of global direct data center water consumption by 2030, and calculates that adopting the least water-intensive technologies in those places could cut their consumption by 45%. That is where a water usage effectiveness improvement buys the most, and it is not where capacity is cheapest to build.
Older Sites Are the Hard Part
Minh Khoi Le, Rystad’s global head of data center and hydrogen research, notes that older facilities are expensive to retrofit. New builds can be designed around a 45C loop from day one. A hall commissioned in 2016 around evaporative towers usually cannot be, which is why fleet-average water usage effectiveness will improve more slowly than any single flagship campus suggests.
Why Water Usage Effectiveness Reporting Cannot Be Compared Yet
If you are trying to hold a provider to account, the reporting problem is the one that will frustrate you first. Every figure quoted in this article is self-published, and no two are assembled the same way.
No Common Boundary
Operators differ on whether they report withdrawal or consumption, whether potable and recycled water are separated, whether leased colocation space is in scope, and whether the denominator is IT load or facility load. Google, for instance, publishes total withdrawal volumes without a fleet-wide water usage effectiveness figure at all. Comparing two numbers built on different boundaries produces a ranking that means nothing.
No Common Cadence
Reporting periods do not line up either. Microsoft reports on a fiscal year ending 30 June. Most others report on the calendar year. Meta’s most recent published figure is a year older than AWS’s. A league table assembled across those cadences is comparing different states of a fast-moving buildout.
No Independent Audit
There is no regulator collecting water usage effectiveness, no assurance requirement, and no penalty for a generous boundary choice. AFP noted that SpaceX, which now owns xAI, has never published an ESG report at all and received MSCI’s lowest rating in June. Absence of disclosure is currently costless, which tells you how much weight the disclosed numbers can bear.
What Water Usage Effectiveness Means for AI Buyers
For anyone procuring AI capacity rather than building it, this is not an abstract sustainability debate. It shows up in tender questionnaires, in ESG reporting obligations that flow down from your own customers, and increasingly in planning risk attached to the regions you deploy into.
Region Choice Is the Biggest Lever
The single most effective decision available to a buyer is where the workload runs. AWS’s own spread from 0.02 to 2.85 L/kWh across regions dwarfs any difference between vendors. Moving a training workload from a water-stressed region to a cool one changes its water footprint by more than switching provider ever will, and it is usually a configuration change rather than a migration.
Efficiency Cuts Both Ways
The same shift that improves water usage effectiveness raises electricity demand, so a water-optimised region choice can worsen a carbon figure if the local grid is fossil-heavy. Treat the two as one decision. Sound Cost Optimization practice already forces this trade-off into the open, because dry cooling shows up on the power bill.
Ask for Volumes, Not Ratios
Request absolute consumption by region and by watershed alongside the ratio. A provider that can answer that question is measuring properly; one that can only quote a fleet-wide water usage effectiveness figure is not. This belongs in your AI Strategy documentation rather than in a sustainability appendix nobody reads.
Water Usage Effectiveness Questions to Put to Your Cloud Provider
These five questions separate operators who measure from operators who market. They work in a tender, a quarterly business review, or a due diligence pack, and none of them requires a specialist to interpret the answer.
The Five Questions
First, is your published water usage effectiveness based on consumption or withdrawal, and is the denominator IT load or facility load? Second, what is the figure for the specific region my workload runs in, not the fleet average? Third, what proportion of that region’s supply is potable? Fourth, is the site in a water-stressed watershed as defined by an external index? Fifth, what does the cooling design fall back to during a heatwave?
What a Good Answer Looks Like
A strong answer gives you a regional number, states the boundary explicitly, and names the fallback mode. A weak answer quotes the fleet average and changes the subject to a replenishment or “water positive” pledge — which is a separate offsetting commitment, not a statement about the water this site draws from this catchment.
Where It Belongs in Governance
Put the answers in the same register you use for cybersecurity and continuity obligations, and re-ask annually, since both the design and the region mix change. Good Vendor Management makes this routine rather than a scramble when a customer questionnaire lands. Teams already tracking cloud adoption and data analytics commitments will find it fits the existing cadence.
Water Usage Effectiveness FAQ
Is a lower water usage effectiveness figure always better?
Not necessarily. A lower ratio usually means less water per unit of compute, but it may have been achieved by shifting the cooling load onto electricity. Read it next to power usage effectiveness and the local grid mix, or you are optimising one number by quietly worsening another.
Are AI data centers really less thirsty now?
Per unit of compute, yes — the published water usage effectiveness figures support that. In absolute terms, no. Microsoft and AWS both improved efficiency between 2022 and 2025 while their total water use rose, and Rystad expects global cooling demand to nearly triple by 2030 on current trajectory.
Does Nvidia’s closed-loop design really use zero water?
Inside the loop, effectively yes: it is filled once and sealed for the life of the facility. Whether the building uses zero water depends on how the heat leaves it. Nvidia’s own claim is scoped to dry-cooler designs in suitable climates, and it expects chillers to run a small number of days a year elsewhere.
Why do companies not just publish total litres?
Some do, and the numbers are large — Amazon’s 2.5 billion gallons withdrawn in 2025, for example. A ratio is more flattering during rapid growth because it improves even as volumes climb, and there is no rule requiring either figure, so most disclosures lead with the ratio.
What is a good water usage effectiveness target for an enterprise facility?
There is no universal answer, because the right target depends on climate and water stress. The published range in this article runs from 0.12 to 0.91 L/kWh. A more useful goal is a design that does not draw potable water from a stressed watershed at all, whatever the ratio says.
References
Tech Thirst: Data center water consumption could triple by 2030 without efficiency gains
AI data centers are less thirsty now, tech giants say
Hotter Than a Hot Tub: The 45C Breakthrough to Cool AI’s Biggest Machines
Amazon’s data centers are 7x more water-efficient than the industry average
Amazon Sustainability: Water Stewardship
Measuring energy and water efficiency for Microsoft datacenters
Inside Microsoft’s two-decade push to cut water intensity while scaling for growth
How Much Water Do AI Data Centers Really Consume?