How Much Power Does an AI Data Center Use?

A typical AI data center uses about as much electricity as 100,000 homes. The biggest ones under construction will use 20 times that, per the IEA.

Worldwide, data centers used around 415 TWh in 2024, roughly 1.5% of global electricity. The IEA expects that to reach about 945 TWh by 2030. AI is the main reason.

How much power does one AI data center use?

The IEA gives a plain benchmark. A typical AI-focused data center draws as much electricity as 100,000 households.

The largest sites now under construction will pull 20 times that. That is enough for roughly 2 million homes from a single campus.

In raw terms, a big AI campus can need several hundred megawatts, running around the clock. That steady, always-on draw is the part that strains a grid.

Sources: IEA Energy and AI, Executive Summary

Where does the power go, training or inference?

Two jobs eat the power. Training builds the model. Inference runs it for users. Both need racks of accelerators, mostly GPUs.

The IEA tracks this as "accelerated servers." It projects their electricity use to grow about 30% a year to 2030. That is the AI slice.

Conventional servers grow slower, about 9% a year. So the spike in demand is specifically the AI hardware, not general computing.

That gap is the story. In 2024 the two jobs drew roughly 415 TWh. As accelerated servers grow near 30% a year, the AI slice drags the total toward 945 TWh by 2030.

Sources: IEA Energy and AI, Energy demand from AI | IEA Energy and AI, Executive Summary

How fast is total demand growing?

Fast. The IEA has data center electricity use growing about 15% a year through 2030.

This is not new. Data center power has grown about 12% a year since 2017, more than four times faster than total electricity use. AI poured fuel on a fire that was already lit.

Two countries drive most of it. The United States and China account for nearly 80% of the growth to 2030.

The map is lopsided. In 2024 the United States ran about 45% of data center electricity, China about 25%, and Europe about 15%.

US data center use rises by about 240 TWh, up 130%. China rises by about 175 TWh, up 170%. That is a lot of new supply to find.

In the US, data centers account for nearly half of all electricity demand growth between now and 2030.

For scale, this is still under 3% of total global electricity by 2030. Data center growth is less than 10% of global electricity demand growth to 2030. Big, but not the whole story.

By 2035 the IEA sees the global total near 1,200 TWh.

Sources: IEA Energy and AI, Energy demand from AI | IEA Energy and AI, Executive Summary

Why did hyperscalers turn to nuclear?

AI data centers need power that runs 24 hours a day, every day. Solar and wind are cheap but they stop when the sun sets or the wind drops.

Firm power means electricity you can count on at any hour, not just when the wind blows or the sun shines.

Nuclear runs flat out around the clock and emits no carbon while running. It is firm, always-on power, and that combination is what the big buyers wanted.

So they went straight to reactors.

What nuclear deals have the AI companies actually signed?

These are signed agreements, not proposals. Three stand out.

Microsoft and Constellation agreed in September 2024 to restart Three Mile Island Unit 1, renamed the Crane Clean Energy Center. It is about 835 MW, on a 20-year deal, with restart targeted for 2028.
Google and Kairos Power signed on October 14, 2024 to deploy up to 500 MW of small modular reactor capacity by 2035, with the first reactor by 2030.
Amazon and X-energy are targeting more than 5 GW of small modular reactor capacity, starting with a four-unit project in Washington state.

Sources: Constellation, Crane Clean Energy Center | Google, Kairos Power agreement | Amazon, nuclear SMR projects

Will nuclear actually cover the new demand?

Not by itself, and not soon. The IEA expects renewables to add the most, over 450 TWh for data centers by 2035.

Natural gas adds about 175 TWh. Nuclear adds about 175 TWh, with the first small modular reactors coming online around 2030.

So nuclear is a real slab of the answer, not the whole answer. The first SMRs land at the end of the decade, not today.

Fusion is further out still, which we cover in when fusion will actually power the grid.

Sources: IEA Energy and AI, Executive Summary

What is the power-adjacent play for a builder?

Call it the power-adjacent play. You do not need to own a reactor to benefit from this shift. You need to reduce or route around the power problem.

Three concrete angles:

Efficiency tooling. Anything that cuts tokens, batches inference, or caches results lowers a customer's power bill directly. That is a real budget line now.
Location and timing. Software that schedules heavy jobs for cheaper, cleaner hours turns a grid constraint into a cost edge.
Materials and grid tech. Better batteries and grid materials ease the bottleneck. AI is already helping find them, as in Google's GNoME materials database.

The point is simple. When power is the scarce input, the tools that stretch each watt get valuable.

If you want help finding your power-adjacent angle, that is the kind of thing I build at nomadtechnologist.com.

Not legal, financial, or tax advice.