Artificial intelligence is often presented as a sustainability opportunity.
It can improve energy forecasting, optimise industrial processes, identify methane leaks, model climate risk and help organisations use resources more efficiently. Those applications are real and potentially significant.
But AI’s environmental footprint cannot be assessed at the level of a software product alone.
It depends on physical infrastructure: data centres, power systems, cooling technologies, transmission networks, water supplies, backup generation, semiconductor supply chains and construction materials.
The sustainability question is therefore not simply whether AI uses “too much” energy.
It is whether the infrastructure supporting AI can be developed within local environmental limits, with credible reporting, fair allocation of costs and meaningful benefits for the communities that host it.
That is the water–energy–AI nexus.
The local dimension matters
Global estimates can highlight the scale of AI infrastructure growth, but they can also conceal the issue that matters most for local stakeholders: where demand occurs.
A facility located in a region with abundant low-carbon electricity, robust grid capacity and low water stress presents a different sustainability profile from one located in a constrained grid zone or water-stressed catchment.
The same is true of cooling design. Water-based cooling may reduce some energy requirements in certain conditions, but it can create material local water implications.
Air cooling, closed-loop systems, reclaimed water and heat-reuse approaches each involve different trade-offs.
UNEP has highlighted that AI data centres face rising energy demand, while water impacts differ considerably depending on cooling approaches and local operating conditions.
This is why broad corporate claims such as “100% renewable” or “water positive” should not substitute for site-specific evidence.
A credible sustainability strategy must be able to answer: What is the facility’s incremental load? Where does electricity come from at the relevant time and location?
What water is withdrawn, consumed or returned? Does the project compete with households, agriculture or ecosystems? What resilience benefits—or risks—does it create for the local grid?
A new test for data-centre projects
Data-centre development should be assessed through an infrastructure lens, not only a corporate emissions lens.
A practical local infrastructure test has five parts.
1. Grid impact
A project should disclose its expected electricity demand, connection timetable, backup-power arrangements and plan for meeting demand with additional low-carbon generation or flexibility. Corporate renewable-energy procurement is useful, but it does not automatically demonstrate that a local grid can accommodate new load without additional fossil generation or network strain.
2. Water stewardship
Operators should distinguish between withdrawal, consumption and discharge. They should also explain the water source, local water-stress context, seasonal variation, cooling system and contingency arrangements during drought conditions.
3. Carbon and materials
Operational emissions matter, but so do construction emissions, diesel backup systems, equipment replacement cycles and electronic waste. A lifecycle perspective is essential as facilities expand and hardware turns over quickly.
4. Community and economic value
A development may create jobs, tax revenue, heat-reuse opportunities or grid investment. These benefits should be specific, measurable and proportionate to the infrastructure demands being placed on the local area.
5. Governance and transparency
Boards and investors need comparable metrics, clear accountability and independent assurance where claims are material. Sustainability information should be available before—not after—planning controversy, water restrictions or grid delays.
What corporate cloud buyers should ask
This is not solely a question for hyperscalers and operators.
Companies buying cloud and AI services are becoming part of the demand signal that drives infrastructure growth. Sustainability teams should therefore include digital infrastructure in responsible-procurement and climate-risk discussions.
A practical supplier assessment can ask:
Where are the workloads hosted, and what is the carbon intensity of electricity in those locations?
What is the provider’s water-use effectiveness and water-replenishment approach at site or regional level?
How does the provider manage capacity in water-stressed regions?
What evidence supports renewable-energy and carbon claims?
Does the provider report embodied emissions, hardware life cycles and e-waste practices?
Can the customer choose regions or workload configurations with lower environmental impact?
Is AI use being measured against a clear business or sustainability outcome?
That final question is important.
Not every AI workload delivers enough value to justify its resource use. Organisations should distinguish between use cases that meaningfully improve efficiency, reduce emissions or strengthen resilience, and those that simply add computational demand without a clear benefit.
The goal is not to stop digital innovation. It is to apply the same investment discipline that businesses would expect for any other energy- and infrastructure-intensive activity.
From efficiency claims to infrastructure accountability
The strongest data-centre sustainability strategies will move beyond headline efficiency metrics.
Power Usage Effectiveness remains useful, but it does not answer whether local water resources are protected, whether renewable procurement is additional, whether a grid connection causes wider system costs or whether the economic benefits are fairly shared.
Similarly, a good water metric must go beyond total volume. A litre used in a water-abundant region is not equivalent to a litre consumed in a drought-prone catchment. Location, timing, source and local stress all matter.
This creates a role for investors and lenders as well. Funding decisions should assess the physical constraints around each proposed facility, not only the projected demand for AI capacity. A project with a compelling technology narrative but weak water, grid or community planning may face delays, higher costs, regulatory risk and reputational pressure.
The opportunity
AI can become part of the climate solution. But it will only retain public legitimacy if its infrastructure is planned responsibly.
The winning model will not be “AI at any cost.” It will be infrastructure designed around clean power, water stewardship, grid flexibility, transparent environmental data and local benefit.
For boards, investors, cloud buyers and policy-makers, that is the question to ask now:
Can this data centre grow within planetary limits—and within the practical limits of the place where it operates?
Sources
United Nations Environment Programme — How to make AI data centres more sustainable
United Nations Environment Programme — AI has an environmental problem: here’s what the world can do about it
United Nations Environment Programme — The environmental impact of the full AI lifecycle needs to be assessed