Amazon Says Its Datacenters Drank 2.5 Billion Gallons of Water in 2025, and Wants Credit for It

Amazon has published a figure most cloud operators would rather keep quiet: its datacenters worldwide used about 2.5 billion gallons of water in 2025. In a company blog post, the retail and cloud giant framed the disclosure as a milestone, saying it is now 75 percent of the way toward a 2022 pledge to become "water positive" by 2030, meaning its facilities would return more water to the environment than they take out.

The number is real, the framing is the part that deserves scrutiny. Amazon softened the figure by comparing it to the amount of water Americans pour onto lawns and gardens, a rhetorical move that tells you more about how the company wants the story read than about what the water means to the towns hosting its AWS facilities.

What Amazon actually disclosed

The headline metric is Water Usage Effectiveness, expressed as liters of water consumed per kilowatt-hour of compute. Amazon put its own figure at 0.12 L/kWh. It then did something operators rarely do in public: it named competitors. By Amazon's accounting, Microsoft used 0.27 L/kWh in 2025, Meta sat at 0.19 L/kWh in 2024, and Google was the heaviest drinker at 1.15 L/kWh over the same period. The Register has asked the three named companies to respond, and those figures are Amazon's characterization, not independently audited cross-company numbers.

That distinction is the whole game. When a company both sets the metric and grades its rivals on it, the disclosure becomes a marketing instrument as much as an accountability one. Water Usage Effectiveness can be gamed by what you choose to count. On-site evaporative cooling shows up in the number. The water consumed by the power plant generating the electricity, often far larger, frequently does not. A facility can post an impressive WUE while still drawing heavily on a stressed regional aquifer through its energy supply chain.

Why the transparency gap matters to the people nearby

For residents, the relevant question is not how Amazon compares to Google. It is how many gallons leave their local watershed during a drought year. Those are different numbers, and the second one is much harder to find.

The accountability problem is concrete. A 2022 investigation found Google's datacenters were consuming more than a quarter of all the water used in The Dalles, Oregon, a figure that only surfaced after a legal fight over whether the usage should be public at all. The company had argued the numbers were a trade secret. That case set the template for the current standoff: operators treat granular, location-specific water draw as confidential, while publishing tidy global averages that cannot be checked against any particular community's supply.

Global totals and efficiency ratios are not what a county water board needs. A town sitting on a single aquifer needs the local withdrawal figure, the discharge quality, and the seasonal peak. Right now most of that lives behind non-disclosure agreements baked into the economic development deals that bring datacenters to a region in the first place.

The regulatory pressure building behind the disclosure

Amazon's voluntary blog post does not exist in a vacuum. It lands as regulators on both sides of the Atlantic start treating datacenter resource use as something to be measured and reported rather than self-certified.

In Europe, the Energy Efficiency Directive now requires datacenter operators above a certain size to report energy and water consumption into a central EU database, with the European Commission working toward an efficiency rating scheme. The Register has noted that Brussels' scorecard could come with a credit warning attached, tying a facility's resource profile to how it is assessed. The direction of travel is clear: in the EU, the kind of figure Amazon just volunteered is becoming a legal obligation with standardized definitions, which removes the operator's freedom to pick the most flattering metric.

In the US the picture is more fragmented. There is no federal water disclosure mandate for datacenters, so the pressure is coming from states, utilities, and local permitting boards. Public sentiment is shifting fast. A recent Ipsos survey found most Americans do not want a datacenter built near them, citing electricity prices, the buildings themselves, and water consumption. A separate finding captured the mood bluntly: Americans would sooner accept a nuclear plant next door than a datacenter. That opposition is starting to translate into permit conditions that demand exactly the local-level transparency the industry has resisted.

The AI multiplier

Underneath the disclosure war is a trend that is not going away. Datacenter water use has climbed for years, driven by the raw number of new facilities and by AI hardware that runs hotter and demands more aggressive cooling than traditional servers. Microsoft's water consumption jumped 34 percent to 6.4 million cubic meters in 2022, a spike the company linked to generative AI workloads. Analysis by The Guardian found that many US facilities now in the pipeline are slated for regions already under drought stress, which is the worst possible place to add a new industrial water draw.

Amazon's defense is its cooling design. The company says about 90 percent of the time its facilities use "free air cooling," pulling outside air across the servers to carry heat away without touching water. Only during the hottest stretches does it switch to evaporative cooling, which does consume water. That architecture genuinely helps, and it is a reasonable thing to engineer toward. But free air cooling depends on climate, and the hottest days, when the system falls back on water, are precisely the days when local supplies are most strained and demand peaks. The annual average hides that correlation.

What actually changes

For users and the public, the practical shift is not Amazon's pledge but the slow standardization of what operators must report and to whom. A voluntary blog post lets a company choose its metric, its comparison, and its framing. A regulatory reporting regime, like the one taking shape in the EU, takes those choices away and makes the numbers comparable across operators and verifiable by someone other than the operator's own communications team.

The honest bottom line, one The Register has made before, is that datacenters cannot get to zero water use no matter how the figures are presented. Cooling a building full of high-density AI accelerators will always cost something, whether in water at the site or water at the power plant feeding it. The goal of disclosure is not to pretend otherwise. It is to give the communities carrying that cost the data to decide whether the trade is worth it, on terms they can verify rather than terms the operator hands them. Amazon's 2.5 billion gallon admission is a step, but the version that counts is the one a regulator, not a marketing department, signs off on.