On June 3, the United Nations University Institute for Water, Environment and Health published a report titled 'The Environmental Cost of AI Energy Consumption: Carbon, Water and Land Footprints,' warning that water and energy consumption and pollution from data centers will double within four years due to artificial intelligence (AI) usage growth. The report projects that by 2030, global data center electricity demand will reach 945 terawatt-hours, with associated water consumption equivalent to the annual basic living water needs of 1.3 billion people, and land occupation exceeding 14,500 square kilometers. The report argues that assessing AI's environmental cost cannot be limited to carbon emissions alone — water footprint and land footprint must be included in the evaluation, because low carbon emissions do not equate to low environmental impact.
UN Report Projects 2030 Data Center Resource Consumption Levels
According to the report, 2025 global data center electricity consumption is estimated at 448 terawatt-hours (1 terawatt equals 1 trillion watts). This electricity consumption generates approximately 208 million tons of carbon dioxide, roughly equivalent to Argentina's emissions last year, and producing this energy consumed approximately 4.5 trillion liters of water.
By 2030, associated water consumption from these data centers is projected to reach 9.3 trillion liters, equivalent to the annual basic living water needs of 1.3 billion people in sub-Saharan Africa. Water resources are primarily consumed in data center cooling and indirect water consumption at power plants.
If global data centers were viewed as a country, their future electricity consumption would rank among the world's highest. By 2030, data centers will account for nearly 3% of projected global electricity consumption, increasing to 945 terawatt-hours — nearly three times the combined annual electricity consumption of Pakistan, Bangladesh, and Nigeria — and will generate nearly 440 million tons of carbon dioxide.
AI Energy Consumption Concentrated in User Interaction Phase
Currently, 20% of data center energy consumption is caused by AI, but by 2030 this proportion will increase to 40%.
The report states that the more complex the AI task, the higher the energy consumption. Currently, the main energy-consuming phase is not training large models, but user interaction with AI and the generation process, accounting for 80% to 90% of AI's overall energy consumption.
Miriam Aczel, a United Nations University environmental policy researcher and collaborator on the study, stated: "What surprised us most is that the most environmentally friendly choice from a carbon emissions perspective often ends up causing more serious damage to water resources or land."
Different AI tasks have vastly different energy consumption. A typical chat query consumes approximately 200 times the energy of a basic text classification task, generating an AI image consumes approximately 1,450 times the energy, and generating a short video consumes energy equivalent to 200,000 spam classification tasks.
The report found that reducing the word count in requests by 30% can lower AI energy consumption by approximately 25%. ChatGPT alone processes 2.5 billion prompts per day, with an annual electricity consumption of approximately 383 gigawatts (1 gigawatt equals approximately 1 billion watts).
Research Reveals Environmental Trade-offs Between Carbon and Water Footprints
The research found that for every 1 kilowatt of electricity consumed by AI, carbon emissions are generated, water resources are consumed during cooling and power generation processes, and land resources are occupied by energy infrastructure and supply chains. These three environmental footprints do not always change in sync. For example, switching from coal to bioenergy can significantly reduce carbon footprint, but may simultaneously significantly increase water and land demand.
Vladimir Smakhtin, Director of the United Nations University Institute for Water, Environment and Health, stated that this report does not oppose AI, but calls for responsible use of AI. In planning, environmental assessment, and community consultation, the true costs of carbon, water, and land must be fully incorporated to ensure the technology revolution develops sustainably and equitably within planetary boundaries.
Smakhtin also noted that although some companies claim their data centers use renewable energy, this means other places will use relatively less clean energy. He further stated: "AI is not just something virtual. We are talking about something with physical properties that produces actual impacts. (AI) has infrastructure and is using energy. Behind all these operations is a large amount of hardware support. Although we don't see smoke coming from devices on our phones and computers — it looks very clean — elsewhere, people are suffering."
SpaceX Adds Water Resource Risk Warning in IPO Filing
On June 1, SpaceX added new wording to the "Risk Factors" section of its initial public offering (IPO) application, stating that obtaining water resources is as important as SpaceX ensuring power supply, obtaining processors, and other critical resources. Water resources are also a key consideration in data center site selection, development, and operations.
Previously, SpaceX mainly emphasized to investors that its data centers are primarily constrained by "obtaining electricity at reasonable prices, lengthy construction cycles, and material shortages."
It is currently unclear why SpaceX added this clause about water, or why it was initially omitted. Tech media Tech Crunch reported that SpaceX is currently in the pre-IPO stage, during which the U.S. Securities and Exchange Commission (SEC) has been sending "comment letters" to the company seeking clarification or supplementary details about the document — SEC inquiries may have prompted this change.
In March this year, Oracle and OpenAI abandoned plans to expand an AI data center in Texas, United States. The plan was originally part of the large-scale U.S. data center project "Stargate."
Besides SpaceX, some tech giants including Microsoft, OpenAI, and Oracle have stated in recent months that they are completely abandoning evaporative cooling to conserve water.
Google Announces Water Replenishment Strategy for 2030
On June 3, Google stated in a blog post that by 2030, its server cluster water replenishment will exceed water consumption. The company is working to reduce environmental impact by actually increasing water supply in communities where data centers are located, investing in local water supply infrastructure, seeking alternative water sources to power the company's facilities, and fully disclosing the company's water usage.
Google parent company Alphabet recently stated it plans to raise $80 billion through stock offerings to fund data center construction.
Ben Townsend, Google's global infrastructure and sustainability director, stated that Google is accounting for indirect water consumption to the extent possible and investing in waterless renewable energy (renewable energy technologies that consume almost no water resources during power generation or hydrogen production, including wind energy, solar photovoltaic energy, etc.).
In 2024, the Lawrence Berkeley National Laboratory, affiliated with the U.S. Department of Energy, predicted in a report that if hyperscale data centers rely heavily on evaporative cooling, they could consume up to 33 billion gallons (approximately 125 billion liters) of water by 2030.
Compared to other high water consumption industries, this figure is comparable or lower. A single hydraulic fracturing well can consume 1.5 million to 16 million gallons of water. However, in areas where water resources are already scarce, this still poses risks, especially in summer, because data center cooling demands often surge simultaneously with municipal water use.
Researchers Call for Industry Transparency on Data Center Operations
Aczel and Smakhtin both pointed out that one problem faced when conducting this research is that many companies and institutions are not transparent about the energy consumed by their data centers and AI systems, and are not even clear about the specific locations and sizes of these systems.
Priscilla Johnson, an independent consultant who served as Microsoft's water resources strategy director from 2017 to 2020, stated that companies can be encouraged to develop better design solutions that simultaneously reduce water and energy consumption. "This industry must accept the challenge and design something smarter and cleaner," Johnson said.
Caleb Marks, president of the National Artificial Intelligence Association, emphasized that AI is rapidly integrating into people's daily lives, bringing many benefits such as improving work efficiency and reducing poverty. The return on investment in AI development has transformative impacts on the world, making it very worthwhile to develop.
Josh Levi, president of the U.S. Data Center Alliance, further stated that the AI industry takes its environmental impact issues very seriously. "We will continue to work with policymakers, local communities, and industry partners to ensure that as data centers scale up, their development process is responsible, transparent, and meets current best practice standards."
However, Smakhtin raised a universal paradox: when things become more efficient, their usage frequency increases, and total energy consumption rises significantly. This occurs even when individual steps in the usage process are more efficient.
FAQ
What did the UN report project about AI data center water consumption by 2030?
According to the June 3 report from the United Nations University Institute for Water, Environment and Health, by 2030 global data center associated water consumption is projected to reach 9.3 trillion liters, equivalent to the annual basic living water needs of 1.3 billion people in sub-Saharan Africa. Water resources are primarily consumed in data center cooling and indirect water consumption at power plants.
How can users reduce AI energy consumption according to the report?
The report found that reducing the word count in requests by 30% can lower AI energy consumption by approximately 25%. Vladimir Smakhtin, Director of the UN University Institute for Water, Environment and Health, stated that people can reduce AI's enormous energy demand by expressing themselves more concisely in queries — users must be very precise and brief, as adding polite phrases like "please" makes a significant difference in energy consumption.
