By 2030, Data Centers Powering AI Could Use Three Times the Electricity of Pakistan, Bangladesh, and Nigeria Combined

By 2030, the data centers powering artificial intelligence are projected to consume 945 terawatt-hours of electricity, nearly triple the combined annual power use of Pakistan, Bangladesh, and Nigeria, according to a new United Nations University report. The figure has been recycled as a single shock statistic. The authors argue it is the wrong lens.

The report, released on 3 June 2026 by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), quantifies AI's footprint across three categories, not one. The same electricity demand that pushes the power number to 945 TWh also pulls a water footprint equal to the basic annual domestic needs of all 1.3 billion people in Sub-Saharan Africa, and a land footprint above 14,500 square kilometers, roughly twice the Jakarta metropolitan area. Each kilowatt-hour leaves its own residue.

"AI is not just about carbon and electricity," said Prof. Kaveh Madani, the report's lead author, in the UN's release. "The hidden costs are water and land." Existing assessments, the report argues, weigh carbon heavily and almost ignore the other two, which makes the totals systematically too low.

The mechanism is physical, not metaphorical. Power plants lose water to cooling. Data halls lose more. Training and inference both run on chips, and chips require ultrapure water at fabrication. The land footprint comes from energy infrastructure siting, transmission corridors, and the surface mines that supply copper, lithium, and rare earths for the grid and the hardware. A carbon-only accounting looks at the smokestack and stops.

The UNU-INWEH team ran the numbers across the world's 20 largest data-center hubs, from the Virginia and Dublin clusters to Singapore, Inner Mongolia, and the Nordic Arctic towns that have rebranded around hyperscale tenants. The hubs diverge sharply. Some are sited on cool, dry plateaus with stressed local watersheds. Others sit beside hydroelectric surpluses. A few are inside coal belts, where marginal electricity is cheap but its carbon load is heavy. A single global average hides which communities absorb the cost.

UN News and the UN regional information centre both framed the report as a call for better measurement, not a moratorium on AI. Madani's own framing is closer to "responsible use" than to opposition, and the data backs that up: closing the measurement gap is a governance project, not a deployment project. The question is whether planners, regulators, and the industry itself will start reading the water and land columns the way they already read the emissions column.

That question is also an equity question. The communities supplying the critical minerals, hosting the substations, and absorbing the e-waste downstream are not the same communities using the models. The report does not name any one operator or jurisdiction. It points to a structural mismatch between where AI's physical costs land and where its benefits accrue, and it asks for measurement that can be acted on locally rather than averaged away globally.

Hacker News readers pushed the relative-scale counter-argument: AI's water and land load is large in absolute terms but small next to agriculture, residential cooling, or the video-streaming baseline. The report does not contest that ordering. What it contests is the leap from "smaller than farming" to "small enough not to measure." Carbon was treated the same way fifteen years ago, and the policy response was to start counting.

The next data point to watch is hub-level disclosure. The report stops at a 20-hub comparison and does not name operators. Whether hyperscale tenants begin reporting site-level water and land intensity the way they already report power-usage effectiveness is the load-bearing test of whether this report changes anything outside the citation graph.