AI Is Not Virtual When the Data Center Needs Water
Reuters reported Wednesday that data centers are expected to consume twice as much power and water by 2030 as artificial intelligence demand accelerates, citing the United Nations University Institute for Water, Environment and Health. The report says data centers consumed 448 terawatt-hours of electricity last year, more than Saudi Arabia, with AI accounting for roughly a fifth. It also estimated 4.5 trillion liters of water use and 189 million tons of carbon dioxide emissions. By 2030, annual data-center power demand is projected to reach 945 TWh, roughly Japan’s current scale, with AI accounting for 40 percent; water use is projected at 9.3 trillion liters. The institute’s director, Kaveh Madani, warned that AI is physical infrastructure: power generation, cooling, transmission, chips, minerals, land, and water — not just software. The political issue is no longer whether AI is exciting. It is who pays when the hype cycle becomes a utility bill, a water fight, and a permitting battle.
The useful way to read this story is to stop pretending AI lives in the cloud. The cloud is a marketing word. The real object is a building with substations, cooling equipment, diesel backup, fiber routes, land contracts, water rights, and a political strategy for pushing the cost of all that onto somebody else.
That is why the UN numbers matter. A data-center industry that already uses more electricity than Saudi Arabia is not an abstract innovation sector. It is an industrial load. If the projection is even close, the next phase of AI will compete with households, factories, farms, and local governments for grid capacity and water. The public debate still talks as if the only policy question is whether America will beat China, or whether one model can reason better than another. Those questions are real. They are not the whole story.
The hidden incentive is simple. AI companies want the upside to be private and the bottlenecks to be socialized. They want fast interconnection, cheap power, favorable tax treatment, relaxed permitting, and a public that believes a chatbot is weightless. But when a new data-center cluster needs transmission upgrades, backup generation, cooling water, or a new rate structure, the bill rarely stays inside the venture-capital pitch deck. It moves into rate cases, land-use fights, utility planning, and local politics.
This is where the media frame gets slippery. The industry sells inevitability: AI is coming, so build first and ask questions later. Politicians like that frame because it lets them pose as pro-growth without pricing the tradeoff. Utilities like it because large-load customers can justify massive capital spending. Local officials like it because ribbon cuttings look better than explaining whose water gets constrained or whose electric bill rises. Everyone has a reason to describe the infrastructure race as progress. Far fewer have an incentive to describe it as a transfer.
That does not mean AI should not be built. It means adults should stop using software language to hide physical scarcity. If an AI campus needs enough electricity to change a region’s load forecast, then citizens deserve to know the rate impact before the deal is locked. If cooling demand stresses local water supplies, the public should know before zoning is treated as a formality. If companies want tax breaks for facilities that employ relatively few people after construction, the subsidy should be argued in daylight.
The second-order consequence is political. AI infrastructure will become one of the places where ordinary people discover that “innovation” often arrives as a bill they did not vote on. The backlash will not come because people hate technology. It will come because they were told the future was virtual while the costs were very physical.
The better policy test is not whether a project has the word AI attached to it. It is whether the project can carry its own infrastructure burden without quietly dumping the cost on households and small businesses. Make the companies disclose projected power demand, water demand, backup generation, and ratepayer exposure. Make utilities show who pays for grid upgrades. Make local governments publish the tax-abatement math.
AI may become enormously productive. But productivity is not magic. It has an energy balance sheet. It has a water footprint. It has a land footprint. If the people building this system believe the benefits are as large as they say, they should have no problem putting the real resource cost in front of the public before the concrete is poured.
Source: Reuters, June 3, 2026.