AI systems run on dense computing infrastructure – data centres, chip factories, energy grids – that consume resources at every stage. Freshwater is used directly for cooling servers, indirectly in power generation, and heavily in semiconductor manufacturing, meaning every “prompt” has a physical impact beyond the screen.
Recent analyses warn that AI’s water demand is on track to reach billions of cubic metres per year by 2027, equivalent to the basic annual domestic needs of more than a billion people. A UN University report estimates that AI‑related water consumption by the end of the decade could match the minimum household water needs of about 1.3 billion people, in a world where many already face water stress.
Large AI platforms are driving this demand through hyperscale data centres that rely on evaporative cooling and water‑intensive electricity generation. In some regions, these facilities are drawing heavily on local water supplies, sometimes amid drought conditions, sparking concern about competition between AI infrastructure and basic human and agricultural needs.
The environmental impact of AI is not limited to water: it is also reshaping energy systems and land use. One assessment finds that AI‑related greenhouse gas emissions in 2025 are comparable to the annual emissions of New York City, with AI usage contributing a significant share of global CO₂ output.
Data centres powering AI could use around 945 terawatt hours of electricity annually by 2030, nearly triple the combined annual electricity use of Pakistan, Bangladesh and Nigeria. That energy carries a carbon footprint and a land footprint, especially where new generation capacity, transmission lines, and mining for critical minerals are needed to feed AI demand.
Importantly, most of this energy use comes from day‑to‑day operation, not just training headline models. One widely used AI service is estimated to process roughly 2.5 billion prompts per day, and generating a single AI image can require more than a thousand times the energy of a simple text classification, with video generation consuming even more.
The platforms with the largest AI water and energy footprints are those backed by hyperscale cloud providers. Studies and reporting point to Microsoft (Azure and OpenAI models such as ChatGPT and Copilot), Google (Gemini and Cloud AI) and Meta (Llama‑based systems and recommender engines) as major drivers of rising water use for cooling and power, due to their enormous data‑centre complexes and rapid AI expansion.
Government and civil‑society analyses warn that this growth is outpacing regulation. A UK Government Sustainable ICT report argues that AI’s water demand is likely to threaten national and global water security, especially in already stressed regions, unless usage is constrained and made transparent.
Researchers argue that AI’s resource footprint is not just a technical problem but a governance and justice issue. The concept of “digital water sobriety” has been proposed as a framework that asks which AI applications truly justify consuming scarce freshwater, where data centres are sited, and how transparent companies are about their water use.
AI growth is currently driven by corporate incentives to maximise usage – more prompts, more images, more automation – without fully accounting for the water, energy and land required. Reports warn that, without constraints, AI’s expansion could undermine climate progress by locking in high‑carbon, high‑water infrastructure just as we need to decarbonise and preserve ecosystems.
For years, “responsible AI” has focused on bias, privacy and safety, but the environmental dimension is only now being recognised as critical. United Nations and civil‑society analyses stress that every kilowatt hour used by AI carries carbon, water and land implications, and low‑carbon energy sources are not automatically low‑water or low‑land.
From a digital‑strategy perspective, this means organisations can’t treat AI as a cost‑free efficiency tool. Choosing heavier generative models, long‑form outputs, and image or video generation at scale magnifies environmental impact; conversely, using lighter models, shorter outputs, and local processing where appropriate can reduce the footprint.
A harmful trajectory – unless we change it
On its current trajectory, AI’s environmental footprint is harmful: it accelerates carbon emissions, strains water supplies and expands land use, with burdens falling disproportionately on certain communities.