Artificial intelligence is often presented as a digital revolution unfolding in the cloud. In reality, it depends on an expanding network of energy-hungry data centres whose demand for electricity, water and land is beginning to reshape local economies, infrastructure planning and the geography of technology itself.
Artificial intelligence is still commonly described as something that happens “in the cloud”, as though it were detached from the physical world. In practice, that cloud rests on a fast-growing network of data centres, transmission lines, substations, cooling systems and industrial-scale construction projects. The public discussion around AI still centres on chatbots, image generators and productivity gains. Far less attention is paid to the material system required to run those tools at scale. The result is a widening gap between the image of AI as weightless software and the reality of AI as an increasingly resource-intensive industrial sector.
The scale of the electricity demand is already clear. A 2024 report by Lawrence Berkeley National Laboratory found that US data centres consumed 176 terawatt-hours of electricity in 2023, equal to 4.4% of total US electricity consumption. The same study projected that, depending on the pace of AI deployment, US data centre demand could rise to roughly 325 to 580 terawatt-hours by 2028, or between 6.7% and 12% of projected national electricity use. That translates into about 74 to 132 gigawatts of demand. The International Energy Agency now treats data centres as a major driver of electricity growth through the end of the decade, with the United States expected to account for a large share of that increase.
The financial outlay behind this expansion is equally striking. Reuters reported in January that Meta expects capital expenditure of between $115 billion and $135 billion in 2026, largely to support AI infrastructure. The company’s own investor materials show capital expenditure of $72.22 billion in 2025, giving a sense of the speed at which spending is rising. Reuters also reported in February that the biggest technology groups are planning an AI spending surge of about $600 billion in 2026. This is no longer ordinary IT investment. It is large-scale industrial build-out, with capital commitments more commonly associated with utilities, transport or energy infrastructure.
That build-out is beginning to affect local electricity markets. A Bloomberg investigation found that wholesale electricity prices in some areas near major US data centre clusters were as much as 267% higher than five years earlier. Price pressures vary by region and cannot be attributed to a single cause in every case, but the broader pattern is becoming harder to ignore. When several large facilities arrive in one place, the cost of grid reinforcement and new generation does not remain confined to the technology companies. It filters into local energy systems and, ultimately, into customer bills.
Water is the other critical input. Many large data centres still rely on water-intensive cooling systems, and the pressure is greatest where new AI facilities are being developed in already dry regions. The World Resources Institute says that two-thirds of all US data centres built or in development since 2022 are located in water-stressed areas. The Environmental and Energy Study Institute notes that a large data centre can consume up to 5 million gallons of water per day, equivalent to the water use of a town of 10,000 to 50,000 people. For local communities, that turns AI from a distant technological trend into a direct issue of water allocation, planning consent and environmental risk.
There is also a growing tension between the AI expansion and the climate commitments made by the large technology firms themselves. In its 2024 Environmental Sustainability Report, Microsoft said its total emissions were up 29.1% from its 2020 baseline, largely because of the materials and construction needed for additional data centres and hardware. The IEA’s wider assessment suggests that electricity generation for data centres could rise from 460 terawatt-hours globally in 2024 to more than 1,000 terawatt-hours by 2030 in its base case. Even if efficiency improves, the overall scale of demand is moving sharply upwards.
The geography of the industry is changing as well. The largest AI training clusters no longer need to sit close to major cities in the way traditional latency-sensitive server infrastructure once did. Instead, companies are increasingly looking for places with cheaper land, large power availability and access to water. At the same time, inference services — the systems that generate rapid responses for users — still benefit from being closer to population centres. The result is a two-layer geography of AI: giant remote training sites on one side, and smaller user-facing facilities nearer cities on the other. The IEA points to this broader restructuring as part of the new energy and industrial landscape being created by AI deployment.
AI is often marketed as frictionless, instant and immaterial. The evidence points in another direction. It depends on an expanding physical system that consumes electricity on the scale of national infrastructure, water on the scale of towns, and capital on the scale of historic industrial booms. The question is no longer whether AI will change the economy. It already is. The more important question is whether governments, regulators and local communities are prepared for the real cost of the infrastructure required to support it.
