AI is starting to look less like a software story and more like a power, water and permitting story. The newest UN numbers put data centers on the scale of national infrastructure.
The AI boom now has a physical footprint big enough to compare with countries. That is the point of a new United Nations University Institute for Water, Environment and Health report, which puts hard figures behind something investors, founders and local officials have been feeling for months: the next bottleneck for AI may not be the model, the chip or the customer. It may be the grid connection.
According to the Associated Press, global data centers used 448 terawatt-hours of electricity last year, more than all but 10 countries. The power behind that activity produced about 189 million metric tons of carbon dioxide, or 208 million U.S. tons, and required about 4.5 trillion liters of water to generate. That is not a rounding error in the climate debate. It is a new industrial load appearing at the same time utilities are already trying to electrify vehicles, buildings and manufacturing.
The report projects data center power use will roughly double by 2030 to around 940 terawatt-hours, approaching 3% of projected global electricity demand. AI accounts for about one-fifth of data center energy use today and is expected to reach 40% by the end of the decade. That shift matters because AI demand is not simply more cloud storage or more video streaming. It is a heavier class of computing, driven by training, inference, image generation, coding agents and enterprise systems that run continuously.
The industry often talks about model training because it is easier to dramatize. A huge frontier model takes enormous clusters, months of planning and large upfront capital. But the larger long-term bill is likely to come from everyday use. Once AI is built into search, office software, customer service, advertising, medical workflows and developer tools, the meter runs all day.
That is why hyperscaler sustainability pledges are about to face a harder test. Microsoft, Google, Amazon and Meta have spent years telling customers that cloud computing can be made cleaner through renewable energy contracts, efficiency gains and better cooling systems. Those efforts are real. The problem is that efficiency does not automatically reduce total consumption when the product becomes cheaper, faster and more widely used. The same pattern has appeared across technology for decades. Better infrastructure often creates more demand for infrastructure.
For AI startups, this changes the cost structure in a quiet but serious way. A company building on rented GPUs may think of infrastructure as a line item from a cloud provider. In reality, that line item is connected to substations, transmission queues, water rights, construction permits and political tolerance in the communities where the machines sit. If those inputs tighten, compute does not just become more expensive. It becomes less predictable.
That will influence which startups can scale. Teams with efficient models, smaller inference footprints and smarter routing may have a real advantage over companies that assume unlimited compute will always be available on demand. This does not mean the smallest model always wins. It does mean capital efficiency in AI will increasingly include energy efficiency, not just headcount discipline and customer acquisition costs.
Local constraints become global strategy
Reuters reported that the UN researchers also expect data center water consumption to reach 9.3 trillion liters by 2030, while carbon emissions rise to 399 million metric tons. The land footprint is forecast to expand sharply as well, from 2,664 square miles last year to more than 14,500 square kilometers by 2030. Those figures turn AI into a siting problem. Where a data center is built can matter as much as what chips it uses.
This is where national comparisons can be useful, but also slightly misleading. AI will not run out of electricity everywhere at once. The pressure will show up in specific places first: Northern Virginia, parts of Texas, Arizona, Georgia, Oregon, Ireland, Singapore and other regions where cheap land, tax incentives, fiber access and power availability have attracted large projects. In those places, the question becomes practical. Who gets the next megawatt? Who pays for new transmission? How much water can a facility use during a dry year?
That creates a political risk the AI industry cannot brush aside. Local communities may not care whether a data center supports a medical model, an ad platform or a chatbot. They will care if electricity bills rise, if water permits look too generous, or if promised jobs do not match the scale of public concessions. For years, software companies benefited from being treated as clean, weightless growth engines. AI infrastructure makes that harder to maintain.
There is still a stronger version of the AI argument. AI can help optimize grids, improve weather forecasting, reduce waste in logistics and speed up scientific work that supports cleaner industries. But those benefits do not erase the immediate resource demand. They have to be measured against it. A useful technology still needs responsible deployment, especially when its costs are concentrated locally and its benefits are spread globally.
The practical takeaway is simple. AI infrastructure is becoming a boardroom issue, a utility issue and a public policy issue at the same time. Investors should ask not only who has the best model, but who has durable access to power. Founders should treat compute efficiency as a product advantage. Governments should demand better disclosure before approving projects that lock in decades of resource use. The next phase of AI will be built in data centers, but it will be decided just as often in planning offices, water districts and power markets.
Also read: Brookfield is turning the AI boom into an infrastructure wager • Coralogix raises fresh funding as AI agents reshape observability • Sitecore is buying Scrunch as AI search becomes a marketing budget
