How Much Power Do AI Data Centres Actually Use, And Where Does It Come From

The Ohio campus Nvidia is helping finance for OpenAI is planned at 4.25 gigawatts, with an option to reach eight. That figure appears in every write-up and almost never gets explained. So: how much power do AI data centres use, where does it come from, and which part of this is actually hard.

How much power do AI data centres use, in numbers you can picture

A gigawatt is a billion watts of continuous draw. One large nuclear reactor produces roughly that. It also sits in the same range as the average electricity demand of a city of about a million people, depending on the country and the season.

An eight gigawatt campus is therefore, very roughly, eight reactors worth of continuous supply committed to a single tenant running one kind of workload. The first 800 megawatts of that site is not due online until 2028, which tells you how much longer the power takes than the building.

At sector level, the International Energy Agency puts global data centre consumption at around 1.5 percent of world electricity, with AI the fastest growing slice. Lawrence Berkeley National Laboratory found US data centre load roughly doubling in the few years to 2023 after a long flat period. That shape, a decade of predictability followed by a step change, is what makes grid planners nervous.

Why the buildings are moving to the power, not the users

Traditional data centres sit near people, because latency matters when you are serving a web page. Training runs do not care. A model does not notice that it is being trained 800 miles from the people who will eventually use it.

That frees operators to site campuses next to generation: gas, nuclear, hydro, or a large solar and storage build. Moving the workload to the electricity is cheaper than moving the electricity to the workload, because transmission lines are slow to permit and politically difficult to route.

The real bottleneck is the queue, not the generation

The constraint is rarely whether the power exists. It is the interconnection queue, the waiting list to attach a large load or a new generator to the grid. Utility and regulator reporting consistently puts those waits in the multi-year range across the US and much of Europe.

This is why announcements increasingly bundle generation and compute into one deal, and why operators sign long contracts for output from existing nuclear plants rather than waiting on new capacity. Buying an existing connection is faster than earning a new one.

What to watch next

Efficiency is doing real work here. Each generation of accelerator delivers more compute per watt, and both optical interconnect and purpose-built silicon cut the energy cost of a useful operation. The honest uncertainty is whether those gains get banked or spent. Historically, cheaper compute has been met with demand for more compute rather than a lower total bill.

The number worth tracking is not total terawatt-hours consumed. It is how much new firm generation actually connects to the grid each year, because that is the line this buildout is pressing against.

For the longer version of where firm low-carbon power might eventually come from, Arthur Turrell's The Star Builders is a readable account of the fusion effort. Future Technology earns a small commission from qualifying purchases made through Amazon links, at no extra cost to you.