Why AI is straining the power grid: data centers, explained
The rise of AI has turned a quiet piece of infrastructure into a central character in the energy story: the data center. These warehouse-scale buildings full of servers have always used significant electricity, but AI is changing the scale of the demand.
Why AI needs so much power
Traditional computing tasks — serving web pages, storing files — are relatively light work. Training a large AI model is different: thousands of specialized processors run flat-out for weeks or months, and each of those chips draws far more power than a conventional server processor. Even after training, answering queries — inference — runs continuously at high intensity across data centers worldwide.
Cooling compounds the demand. All of that computation becomes heat, and removing it takes yet more energy, which is why efficiency metrics and cooling design have become competitive battlegrounds among operators.
The grid-level effect
Utilities in data-center-heavy regions now field connection requests measured in hundreds of megawatts — the scale of small cities. That is reshaping planning in several ways: grid operators are re-forecasting demand growth after years of flat consumption, new capacity and transmission are being built with data centers explicitly in mind, and operators are signing long-term power contracts, including with nuclear and renewable providers, to lock in supply.
Siting has changed too. Where data centers once clustered near internet exchange points, access to abundant power increasingly decides where the next campus goes.
What to watch
The pressure is pushing in two directions at once: more demand, and more efficiency. Chipmakers are racing to deliver more computation per watt, operators are experimenting with advanced cooling, and grids are adding capacity. Which side moves faster will shape both the cost of AI services and the pace at which the technology can keep scaling.
The bottom line: AI’s growth is no longer just a story about software. It is a story about physical infrastructure — and electricity is its most binding constraint.