Data Centres and the Real Cost of AI: Electricity, Water, Land

The physical layer nobody sees

Cloud is a marketing word. There is no cloud. There are buildings, full of racks of specialised processors, drawing electricity off national grids and pumping waste heat into the air. Every interaction with an AI system is a round trip to one of those buildings. The energy, water, and land that the buildings consume are real, measurable, and increasingly material to the politics of where AI is built and who pays for it.

The International Energy Agency's *Electricity 2024* report put data-centre consumption at roughly two per cent of global electricity use and rising sharply, a figure updated in the IEA's *Energy and AI* analysis. In some countries the share is much higher. Ireland's data-centre sector consumed about 21 per cent of national electricity in 2023 according to Ireland's Central Statistics Office, and is on a trajectory to exceed a quarter of national demand by 2027 on current projections. The Netherlands, Denmark, and Singapore have all imposed moratoriums or stringent siting restrictions on new data-centre construction. Northern Virginia, which hosts the densest concentration of data centres on the planet, is in active negotiations with utilities about how to add capacity without expanding fossil-fuel generation.

These are not edge cases. They are the leading edge of a problem that will reach more jurisdictions as AI inference scales. For the model architecture that drives most of this inference load, see our explainer on large language models.

How much electricity an AI query actually costs

Training a frontier model is the line item that captured early public attention. Training a state-of-the-art language model takes tens of megawatt-hours of electricity over several weeks. That number is real but it is paid once, amortised over the model's deployment lifetime, and is now a small fraction of total AI energy use.

Inference is where the cost lives. Each user query consumes a small amount of electricity, but the volume of queries is now in the billions per day across the major providers. The best published estimates, including de Vries (2023) in *Joule*00365-3), put a single chatbot interaction at roughly an order of magnitude more energy than a single web search. Image generation costs more again. Video generation costs more again. The marketing pressure to integrate generative AI into more products multiplies the inference load.

The implication is that the energy cost of AI is now driven by how many times models are used, not how many times they are trained. Improvements in model efficiency, of which there have been many, are racing against growth in deployment volume, and deployment volume has so far been winning.

Water, which gets less coverage but matters more locally

Cooling a data centre consumes water. The dominant method, evaporative cooling, uses freshwater that is then released as vapour. Estimates vary widely by site and by climate, but research from Shaolei Ren and colleagues at UC Riverside ("Making AI Less Thirsty", 2023) finds that a frontier model training run can consume hundreds of thousands of litres of water. Inference workloads at scale consume vastly more in aggregate.

The local impact concentrates in dry regions where data centres have been sited cheaply. Several facilities in the western United States, in northern Chile, and in parts of Spain are now in active disputes with regional water authorities. The disputes are not about absolute volumes that would be alarming in a wet country. They are about how much of a scarce local resource a single facility is permitted to consume in a place where agriculture and municipal supply are already constrained.

For humanitarian observers, the relevant pattern is that data-centre siting decisions are usually framed as economic-development questions and resolved without participation from the communities whose water and electricity costs will be affected. The infrastructure looks technical. The decisions are political.

Where the carbon footprint actually lives

If a data centre is powered by renewables, its operational carbon footprint can be very low. If it is powered by the grid in a coal-dependent jurisdiction, its operational footprint is high. The trend in the industry is toward renewable power purchase agreements, often with new dedicated capacity, but the accounting is contested. A data centre that buys renewable energy certificates on paper while drawing power off a fossil-heavy grid in practice is not actually decarbonised. It has bought a financial instrument.

The more honest accounting includes three additional components. The embodied carbon of the chips themselves, which are manufactured in highly energy-intensive fabrication plants concentrated in Taiwan, South Korea, and increasingly the United States. The embodied carbon of the building itself, which is non-trivial because data centres are concrete- and steel-intensive structures. And the displaced emissions, which are the increased fossil-fuel use elsewhere on the grid when a data centre's renewable supply is being consumed.

Aggregate published figures from the major cloud providers improved through the mid-2020s and then plateaued or reversed as AI demand outran new renewable capacity additions. Google's 2024 Environmental Report, for example, showed total emissions up roughly 48 per cent against the 2019 baseline, with AI cited as a leading driver. Microsoft's 2024 Environmental Sustainability Report showed a 29 per cent increase against its 2020 baseline. Both companies remain publicly committed to net-zero targets that depend on technologies, principally direct air capture, that do not yet exist at the scale required.

Who pays

The cost of running data-centre capacity at the scale AI now requires is not paid only by the companies that own the data centres. It is paid by ratepayers in the jurisdictions where the data centres are sited, through electricity tariff structures that often distribute infrastructure-upgrade costs across all customers. It is paid by water users in dry regions whose supply is reduced. It is paid through the air-quality impacts of the gas peakers that are increasingly being built to cover data-centre demand. And it is paid through the opportunity cost of the renewable capacity that goes to running AI inference instead of decarbonising other sectors of the economy.

None of this is an argument against AI. It is an argument for paying attention to the full ledger, and for being suspicious of efficiency claims that count only the cost inside the data centre and not the costs that flow outward from it.

What this means for humanitarian work

Three concrete implications for the humanitarian sector.

First, AI is not free. Every query has a marginal cost in electricity and water that the user does not see. Tools that wrap AI into routine workflows multiply that cost across every team member. Procurement decisions about AI-enabled software should ask the vendor for an estimate of the per-query energy and water cost, and most vendors will not be able to provide one. That is itself useful information.

Second, siting matters. Cloud providers offer the choice of which region to host workloads in. Choosing a region powered by a cleaner grid is a real lever. It is also the case that many humanitarian organisations are constrained by data-protection requirements to host inside particular jurisdictions, which limits the choice. The trade-off is worth being explicit about.

Third, the climate-displacement connection is direct. The same AI infrastructure that helps humanitarian responders forecast displacement contributes, at the margin, to the emissions trajectory that drives climate-induced displacement. The sector cannot honestly claim to take climate seriously while expanding its AI footprint without measuring it. Several of the major operational agencies are now publishing AI energy budgets alongside their broader sustainability reporting. The practice should spread. Our guide to understanding displacement covers how climate-driven displacement is measured and reported.

What a sensible operational discipline looks like

There is no single answer, but the operational patterns that hold up across the agencies doing this work seriously look broadly similar. Track the energy and water consumption of AI workloads as a budget line, not an externality. Prefer smaller specialised models over frontier models where the task does not require frontier capability, because the energy cost per inference is often an order of magnitude lower. Cache aggressively, because every cached answer is a query not made. Run batch workloads on a schedule that aligns with renewable availability on the chosen grid. And include the AI footprint in the same emissions reporting as travel and office operations, because excluding it is a choice that flatters the numbers.

Further reading

The International Energy Agency's *Electricity 2024* report and follow-on *Energy and AI* analysis have the cleanest global numbers on data-centre demand. Ireland's Central Statistics Office publishes the most transparent national-level data-centre electricity figures, which are useful as a leading indicator for jurisdictions earlier in the curve. Shaolei Ren and colleagues at UC Riverside have published the most-cited peer-reviewed work on the water footprint of AI. The Green Web Foundation maintains an ongoing dataset on the carbon intensity of cloud regions that is useful for siting decisions.

The honest summary is that the physical layer of AI is significant, growing, and unevenly distributed. The promotional narrative of AI as a weightless cloud-based service is, in 2026, a marketing position that survives only because most users never encounter the building their queries reach.

Sources

• International Energy Agency, *Electricity 2024* and *Energy and AI*. Primary source for the global data-centre electricity share quoted above.
• Central Statistics Office Ireland, Data Centres Metered Electricity Consumption 2023. Source for the 21 per cent Ireland figure.
• de Vries, A., *The growing energy footprint of artificial intelligence*00365-3), Joule, 2023. Source for the per-query energy comparison.
• Li, Yang, Islam and Ren, *Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models*, 2023. Source for the water-footprint estimates.
• Google, 2024 Environmental Report. Primary source for the 48 per cent emissions figure.
• Microsoft, 2024 Environmental Sustainability Report. Primary source for the 29 per cent emissions figure.
• Green Web Foundation. Carbon-intensity dataset referenced for siting decisions.
• Lawrence Berkeley National Laboratory, *2024 United States Data Center Energy Usage Report*. U.S. national reference for data-centre demand projections.