Every time an AI prompt is sent — a question to a chatbot, an image generated, a video produced — it draws on more than electricity. According to a new United Nations University report published on 3 June 2026, the global data centres powering artificial intelligence could consume 945 terawatt-hours of electricity by 2030, alongside water and land footprints large enough to affect millions of people. For users in Singapore and across Southeast Asia, where data centre capacity is expanding rapidly under tight government oversight, the findings carry particular weight.
What does the UN report say about AI’s environmental footprint?
The report, Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, was produced by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), marking the institute’s 30th anniversary. Led by Professor Kaveh Madani, the 2026 Stockholm Water Prize laureate, the investigation argues that AI’s environmental cost has been systematically mismeasured by focusing on carbon emissions alone.
Global data centres consumed an estimated 448 terawatt-hours of electricity in 2025 — comparable to the 11th largest national electricity consumer in the world, behind France and ahead of Saudi Arabia. By 2030, that figure could nearly double to 945 terawatt-hours, with an associated water footprint of 9.3 trillion litres (equal to the basic annual domestic water needs of 1.3 billion people in Sub-Saharan Africa) and a land footprint exceeding 14,500 square kilometres, roughly twice the size of the Jakarta metropolitan area.
Crucially, the report finds that these three footprints — carbon, water and land — do not move in the same direction. Switching electricity generation from coal to bioenergy, for instance, can cut carbon emissions by an average of 70 per cent while increasing the associated water footprint more than thirtyfold and the land footprint a hundredfold. “Low-carbon” infrastructure, in other words, is not automatically “low-water” or “low-land”.
Why does sending a single AI prompt matter?
Public debate has tended to focus on the energy required to train large AI models — estimates suggest GPT-4’s training run consumed between 50 and 70 gigawatt-hours of electricity. But according to the UN report, this framing is now outdated. Once a model is deployed, inference — the everyday running of the model to answer user prompts — accounts for 80 to 90 per cent of total AI energy use. ChatGPT alone is estimated to process around 2.5 billion prompts a day, translating to roughly 383 gigawatt-hours of electricity annually for a single product.
Per-query impact also varies enormously depending on the task. A typical conversational query is around 200 times more energy-intensive than a basic text classification task. Generating a single AI image can require around 1,450 times that baseline — enough to power a 10-watt LED bulb for 17 minutes — while a short, high-complexity AI video can consume as much electricity as 200,000 spam-classification queries, with an associated water footprint of around 4.1 litres, close to two days’ drinking water for one person.
For comparison, Google has published its own per-prompt figures for its Gemini assistant: a median text prompt is estimated to use 0.24 watt-hours of energy — about the same as watching television for nine seconds — and around 0.26 millilitres of water, or roughly five drops. Google says efficiency improvements reduced the energy footprint of a typical Gemini prompt by a factor of 33, and its carbon footprint by a factor of 44, over a single 12-month period.
The gap between Google’s per-prompt figures and the UN’s aggregate projections is not necessarily a contradiction — both can be true at once. A tiny footprint per prompt, multiplied across billions of daily queries and a rapidly growing user base, is exactly the dynamic the UN report describes.
Does AI efficiency actually reduce its environmental impact?
This is where the report’s central warning comes in: the rebound effect, also known as the Jevons paradox. As AI models become cheaper and more efficient to run, they tend to be used more often — and the report argues that without explicit limits on token usage, image resolution or default output length, efficiency gains at the per-query level are easily overtaken by sheer growth in volume.
“A lot of people think that the environmental footprint of AI reduces, as technology improves and processes become more efficient. But that is only a partial picture of the overall problem,” said Professor Madani. “More efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains.”
The report also points to a growing concern around AI-generated images and video, which it describes as an emerging environmental pressure point given how energy-intensive these formats are relative to text.
What does this mean for data centres in Singapore and Southeast Asia?
The UN report highlights several site-level cases where AI infrastructure built to serve global users creates concentrated local pressure. In Ireland, data centres accounted for 21 per cent of total metered electricity in 2023 — more than all urban households combined — prompting the national grid operator to pause new approvals around Dublin until 2028.
Singapore has faced a version of this tension for years. Data centres already account for more than 7 per cent of the country’s electricity consumption, which led the government to impose a moratorium on new data centre developments between 2019 and 2022. Since then, capacity has been allocated selectively: the second Data Centre Call for Application (DC-CFA2), launched by the Economic Development Board and the Infocomm Media Development Authority in December 2025, makes at least 200 megawatts of new capacity available to operators that meet strict efficiency and sustainability requirements, including a Power Usage Effectiveness (PUE) target of 1.25 at full load and a minimum of 50 per cent power from green energy sources.
Digital Realty, which operates data centres across the region including in Singapore, shared that it achieved 100 per cent renewable energy coverage in Singapore through direct retail energy agreements, supported by locally generated biomass energy, regionally sourced renewable energy credits and on-site solar. Globally, the company reported a Power Usage Effectiveness of 1.38 and a Water Usage Effectiveness of 0.59 in 2025 — a 15.7 per cent improvement on the previous year — with 97 per cent of water supplied across its Asia Pacific portfolio coming from non-potable sources.
These figures compare favourably with the industry-wide average WUE of around 1.8 to 1.9 litres per kilowatt-hour. However, independent researchers have noted that fewer than a third of data centre operators globally even track water usage metrics, and that reported figures rarely distinguish between direct water use — which places immediate pressure on local supplies — and indirect water embedded in electricity generation.
How are AI companies responding to growing infrastructure demands?
The scale of AI’s infrastructure needs is also reshaping how companies source their compute capacity — sometimes in ways that complicate previously straightforward sustainability claims. Apple, for instance, has maintained 100 per cent renewable energy coverage across its own data centres since 2021, powering Private Cloud Compute (PCC), the system underpinning Apple Intelligence.
At WWDC 2026, however, Apple announced it was expanding Private Cloud Compute beyond its own infrastructure for the first time, working with Google and Nvidia to run more demanding Apple Intelligence workloads — including agentic tool use and complex reasoning — on Google Cloud servers powered by Nvidia’s Blackwell B200 chips. According to reporting on the announcement, Apple had attempted to run its newer Siri model on its own PCC hardware first, but found it ran too slowly in testing.
Google has its own sustainability commitments for its data centre fleet, including a goal to run on 24/7 carbon-free energy across all operations by 2030. But the shift illustrates a broader dynamic described in the UN report: as AI workloads scale and get distributed across third-party infrastructure for performance reasons, the environmental accounting becomes harder to trace — even for companies, like Apple, with a long-established renewable energy record.
What should AI users and businesses take away from this?
The UN report is careful not to frame its findings as a case against AI itself. “It is a call for using it responsibly and addressing its unintended impacts proactively to make it sustainable and equitable,” said Professor Madani. Among its recommendations: governments should integrate AI infrastructure into energy and water planning and require standardised environmental footprint reporting; companies should treat model selection and default settings as footprint decisions; and users and organisations should adopt “fit-for-purpose” use — choosing the lightest model and lowest-energy format suited to the task at hand.
For everyday users, that might mean defaulting to text rather than image or video generation where either would do, or being more deliberate about how often AI tools are used for tasks that don’t strictly need them. For enterprises across Southeast Asia weighing AI adoption at scale, the report’s findings suggest that questions about where infrastructure is sited, how it is powered, and how its water and land use are measured — not just its electricity bill — are becoming part of the cost of doing business with AI.
