Amazon for the first time releases data center water usage data: 2.5 billion gallons of water in 2025

Amazon first disclosed data center water usage data: by 2025, global water consumption will reach 2.5 billion gallons, with a water usage efficiency of 0.12 liters per kilowatt-hour, about 7 times more efficient than the industry average of 0.84 liters, claiming to be better than peers.
(Background: Why AI hasn't caused large-scale unemployment for software engineers? Latest research: Humans are irreplaceable in judgment and accountability.)
(Additional context: The Information: Google plans to commission Samsung to produce the 10th generation AI chip "Icefish," diversifying TSMC supply risks.)

Table of Contents

Toggle

• The accounting gap behind impressive numbers
• The true limits of cooling technology
• The difference between water and electricity

Amazon recently disclosed its data center water usage data for the first time. The figures are indeed impressive. Amazon states that by 2025, global data center water consumption will reach 2.5 billion gallons (about 9.5B liters), with a water efficiency of 0.12 liters per kilowatt-hour, a 2% decrease from the previous year, despite continued expansion of operational scale.

The report includes comparison charts showing that Google, Microsoft, and Meta all have higher water use per kilowatt-hour than Amazon, and claims that Amazon's efficiency is 7 times better than the industry average. Simply put, Amazon is saying: "We are the most water-efficient in this industry."

The accounting gap behind impressive numbers

0.12 liters vs. industry average of 0.84 liters, the difference is indeed significant. But this set of numbers relies on a key assumption, hidden in the report's notes: Amazon's statistics do not include the indirect water consumption at power plants supplying electricity, nor the water used during the construction of new data centers.

This exclusion is not trivial. Cooling water consumption at thermal power plants is often the largest single component of a data center's entire lifecycle water footprint, estimated to account for a much higher proportion of total water use than the direct water used at the facility. Every drop of water saved by the data center itself may be offset by the evaporation of dozens of times that amount at power plant cooling towers located hundreds of kilometers away. In other words, the 2.5 billion gallons on Amazon's books are only the direct costs; the indirect costs are unaccounted for—and that ledger could be much thicker.

On the other hand, 0.12 is an "average": it averages out peak electricity demand periods and data centers located in arid regions, diluting the actual water burden during the hottest weeks or in the most water-scarce facilities. A beautiful annual average does not reflect the real burden during the hottest weeks or in the most water-stressed plants.

Amazon also cites data from Google, which is worth noting. The unfavorable comparison figure for Google in the report seems to mainly focus on water use at the Gemini AI data center, whereas Amazon reports on all data center operations. The two denominators are fundamentally different—one is specific to AI workloads, the other to overall business. Comparing them directly is methodologically flawed, which significantly reduces the persuasiveness of this comparison chart.

The true limits of cooling technology

Amazon explains how it can achieve 0.12 liters: about 90% of the time, it uses air cooling that consumes no water; only during the hottest days and peak hours does it activate evaporative water cooling, continuously increasing the servers' heat tolerance.

This logic is sound, and the technological approach aligns with industry trends. Microsoft is integrating closed-loop, zero-evaporation cooling systems into multiple data centers, each saving over 125 million liters of water annually. Cooling technology improvements are real, but can they keep pace with the rapid growth in computing power demand?

This highlights a fundamental structural contradiction facing the industry: each company demonstrates efficiency improvement curves, with water use per kilowatt-hour decreasing, but total electricity and water consumption continue to rise. The denominator of efficiency shrinks, but the numerator—scale—grows even faster. The end result is that total consumption does not decrease but increases. Some studies warn that, if current trends continue, water consumption for AI data centers by 2030 could be equivalent to the annual water needs of 1.3 billion people.

The difference between water and electricity

Finally, I want to add a concept: although 2.5 billion gallons may still be within an acceptable range on a global scale, water is not like electricity that can be scheduled across grids. It is drawn from local water systems and specific communities. A data center built in a water-scarce region, even with impressive WUE figures, may still compete with local residents and agriculture for the same underground water sources or river water rights.

This is why local governments in places like Seattle prefer to pause or restrict data center development: global averages cannot address local water accounting issues.