Amazon has announced that its global data centre operations used just 0.12 litres of water per kilowatt-hour (L/kWh) in 2025, making them more than seven times more water-efficient than the industry average of 0.84 L/kWh.
In a statement, the company said it uses significantly less water per unit of computing power than most operators in the global data centre industry, which collectively accounts for less than 0.5 per cent of industrial water use worldwide.
"We're continuing to get even more efficient year over year. These gains are the result of years of investment in custom cooling technology, smarter systems, and a commitment to minimise water use wherever possible," Amazon said.
Data centres power a wide range of digital services, from video calls and online banking to virtual healthcare and education. However, maintaining the servers that enable these services requires effective temperature management.
"To deliver that computing reliably, we need to maintain optimal temperatures," said Joern Tinnemeyer, a data centre engineering leader at Amazon. "My team focuses on thermal management—taking the heat generated as a by-product of computing operations and removing it as effectively and efficiently as possible."
According to Amazon, around 90 per cent of the time its data centres rely on "free air cooling", which involves drawing in outside air, passing it through server halls to absorb heat, and releasing it back outdoors without using water.
"It's kind of like in your house. It's a nice summer morning, it's not that hot out, and you open the windows instead of turning on the air conditioner," Tinnemeyer explained.
The company said it has spent years increasing the temperature thresholds at which its data centres operate, while designing servers capable of functioning reliably at higher temperatures. This has reduced the need for water-based cooling systems.
Amazon now uses water to cool incoming air only when ambient temperatures exceed approximately 85°F (29.4°C), significantly reducing water consumption across most operating environments.
"This is how we innovate at Amazon," Tinnemeyer said. "We set an ambitious target that benefits our customers, iterate relentlessly, and validate with data—in this case, proving we could cut water use in half without any impact on performance."
To verify the approach, Amazon analysed thousands of hours of operational data across its data centre campuses. According to company officials, higher operating temperatures did not increase equipment failure rates.
The company also highlighted its broader sustainability goals, stating that it aims to become water positive by 2030 and is already about 75 per cent of the way towards achieving that target.
AI's growing environmental footprint
Amazon's announcement comes amid increasing concern about the environmental impact of artificial intelligence.
A recent report by the United Nations University (UNU) warned that AI-related water consumption could eventually match the basic annual domestic water needs of 1.3 billion people by the end of this decade. The report also estimates that AI infrastructure could occupy more than 14,500 square kilometres of land globally—roughly twice the size of the Jakarta metropolitan area.
The study argues that discussions around AI's environmental impact often focus heavily on greenhouse gas emissions, particularly those associated with training large language models, while overlooking other critical factors such as water use, land consumption and electronic waste.
Researchers caution that some solutions designed to reduce carbon emissions may unintentionally increase pressure on water resources and land availability, especially in regions already facing environmental stress.
Daily AI use drives most energy demand
Contrary to popular perception, the report found that everyday use of AI systems—not model training—accounts for around 80 to 90 per cent of total AI energy consumption.
One widely used AI service is estimated to process approximately 2.5 billion prompts each day, consuming hundreds of gigawatt-hours of electricity annually.
Energy requirements also vary dramatically depending on the task. Generating an AI-created image can consume more than 1,000 times the energy required for basic text classification, while AI-generated video demands even greater computational resources.
The report warns that efficiency improvements alone are unlikely to curb overall resource consumption because lower operating costs and better performance often encourage greater adoption—a phenomenon known as the rebound effect.
Uneven environmental costs
The environmental burden of AI infrastructure is not distributed equally around the world.
In some countries, data centres already account for a substantial share of national electricity consumption, while in others, rapidly expanding facilities are placing increasing pressure on local water resources, sometimes in drought-prone regions.
The report also projects that AI infrastructure could generate up to 2.5 million tonnes of electronic waste annually by 2030. Much of this waste is expected to end up in lower-income countries that often lack adequate recycling and disposal systems.
In addition, the extraction of critical minerals required for AI hardware raises concerns about environmental degradation and social inequalities in mining regions.
A widening global divide
The report highlights growing disparities in AI infrastructure ownership and access. More than 90 per cent of AI-specialised computing capacity is concentrated in just two countries—the United States and China—while over 150 countries have little or no significant domestic AI infrastructure.
Researchers warn that this imbalance could deepen both economic and environmental inequalities, with some regions bearing the environmental costs of AI development without enjoying a proportional share of its benefits.
Call for responsible AI development
Despite its findings, the UNU report does not argue against the adoption of AI. Instead, researchers are calling for the creation of a "responsible AI ecosystem" that balances technological progress with environmental sustainability.
The framework proposed by the study emphasises transparency, efficiency-by-design, equity, lifecycle responsibility, international cooperation and sustainable use to ensure that AI development remains within planetary limits.