The artificial-intelligence boom is often described as a contest over chips, models and software. Increasingly, it is also a contest over electricity.
Across the United States, companies are building or planning large data centres to train and operate AI systems, provide cloud services and store the world’s growing volume of digital information. These facilities can run around the clock. Their appetite for power is forcing utilities, regulators and local communities to ask a difficult question: how can the grid support rapid digital growth without shifting unreasonable costs or environmental burdens onto ordinary customers?
From flat demand to rapid growth
For much of the period from the mid-2000s to the early 2020s, U.S. electricity use changed relatively little. Efficiency gains helped offset population and economic growth. That pattern is now changing. The U.S. Energy Information Administration has linked rising demand to the commercial sector, including data centres, and to industrial expansion.
The Department of Energy says data centres used about 4.4% of total U.S. electricity in 2023. A Lawrence Berkeley National Laboratory analysis cited by the department estimated that the share could rise to roughly 6.7%–12% by 2028. The wide range shows how uncertain the trajectory remains, but even the lower end represents a major planning challenge over a short period.
Why AI uses so much power
Traditional websites and business software already require servers, cooling systems and network equipment. Generative AI adds more demanding workloads. Training a large model can require clusters of specialised processors operating at high intensity, while serving millions of user requests creates a continuing “inference” load.
The electricity demand does not end with the computer chips. Data centres also need cooling, backup systems and power-conditioning equipment. Operators have improved efficiency, but the number and scale of facilities are growing so quickly that total consumption can still rise.
The grid cannot be expanded overnight
A data centre can move from proposal to operation faster than a large transmission line or power plant can be planned, approved and built. That mismatch is becoming one of the central infrastructure problems of the AI era.
Utilities must forecast how much electricity a proposed facility will actually use and when it will come online. If they build too little, reliability can suffer. If they build too much and a project is cancelled, other customers could be left paying for underused infrastructure.
Local debates therefore extend beyond technology. Communities want to know who will pay for new substations and transmission, how much water cooling systems will require, how many permanent jobs a facility will create and whether household electricity bills will rise.
Clean energy—and the return of old fuels
Technology companies have announced major clean-energy purchases, and data centres can provide long-term demand that supports new wind, solar, storage, geothermal or nuclear projects. Flexible operation and better batteries may eventually allow some computing work to shift toward hours when clean power is abundant.
But the near-term picture is not automatically green. The EIA has warned that faster-than-expected data-centre demand could increase fossil-fuel generation in some regions. Grid congestion can also delay renewable connections. The climate impact will depend not only on how much electricity AI uses, but on where and when that electricity is produced.
Why Europe should pay attention
Europe faces many of the same questions: how to attract digital investment, maintain reliable power and meet climate goals. The American experience is becoming a large-scale test of different approaches to grid connections, cost allocation and corporate energy procurement.
It also matters economically. If abundant power allows the United States to build AI infrastructure faster, that could strengthen American cloud and technology companies. If grid bottlenecks raise costs or slow construction, other regions may gain an opportunity. Energy policy is becoming technology policy.
A better measure of the AI race
The AI industry often measures progress in model size or computing performance. A more revealing measure may be whether new infrastructure delivers broad value without weakening affordability, reliability or climate commitments.
That will require transparent demand forecasts, stronger transmission planning, clear rules on who pays for upgrades, efficient cooling and investment in low-carbon generation. The companies building the future of computing may discover that their most important partner is not another software firm. It is the power grid.