NodalZen

Blog · News

How much electricity do data centers and AI use

· 6 min read · power, grid, forecasts

What the US national lab and the International Energy Agency say about data center electricity use today and where it may be heading.

Headlines about AI and electricity can feel like a blur of big numbers. This post sets out what two of the most cited sources actually say: a US national laboratory report on American data centers, and the International Energy Agency's reports on the whole world. We also explain why their forecasts come as ranges, not single numbers.

A quick word on units

Electricity use over time is measured in watt-hours. A kilowatt-hour (kWh) is what a 1,000-watt microwave uses in an hour. A terawatt-hour (TWh) is one billion kilowatt-hours. For scale, the average US household bought about 10,791 kWh in 2022 (US Energy Information Administration).

The United States: the Berkeley Lab report

In December 2024, the US Department of Energy released a report from Lawrence Berkeley National Laboratory on data center energy use. According to the department's summary (US Department of Energy, December 2024):

  • US data centers used about 58 TWh in 2014.
  • That rose to about 176 TWh in 2023, or about 4.4% of all US electricity.
  • By 2028, the report projects 325 to 580 TWh, which would be about 6.7% to 12% of US electricity.

Put another way, by our own arithmetic using the household figure above, 176 TWh is roughly what 16 million average US homes buy in a year.

The 2028 range is wide. The high end is almost double the low end, which is an honest signal of how uncertain the next few years are.

The world: the International Energy Agency

The International Energy Agency (IEA), an intergovernmental energy body, published a major report called Energy and AI in April 2025. News coverage of that report said the IEA projected global data center electricity use of about 945 TWh in 2030, more than Japan uses today, with AI as the main driver. It also said data centers would account for nearly half of US electricity demand growth to 2030 (RCR Wireless News, April 2025).

In 2026 the IEA followed up with a report called Key Questions on Energy and AI. As summarized by Our World in Data (Our World in Data, July 2026):

  • Data centers worldwide used about 485 TWh in 2025.
  • Data centers focused on AI used about 155 TWh of that, roughly one third.
  • In the IEA's base case, global data center use reaches about 945 TWh by 2030, about 3% of the world's electricity, with AI-focused sites making up about 465 TWh.

Our World in Data calculated that the 2025 total was about 1.5% of global electricity generation, and AI-focused data centers about 0.5%.

Why the estimates disagree

Not everyone counts the same way. The same Our World in Data article notes that the Energy Institute and S&P Global put 2025 data center use at about 790 TWh, roughly 60% higher than the IEA figure (Our World in Data, July 2026).

Gaps like this come from a few honest sources of doubt:

  • What counts as a data center. Studies can draw the line differently, for example around small server rooms.
  • How hard the chips work. A GPU that sits idle half the day uses far less than one running flat out.
  • Building efficiency. The overhead for cooling and power losses, measured by PUE, changes the total a lot.
  • Company secrecy. Many operators do not publish their electricity use site by site.

Small share, big local impact

A few percent of global electricity may not sound like much. The pressure is local. Data centers tend to cluster in particular regions, and a large campus asks for all of its power at one connection point, which the local grid has to be able to deliver. That is why grid connection waits and equipment shortages, rather than the global total, are often the real limit on building. We cover that in a separate post on the grid bottleneck.

How to read future headlines

When you see a new number, ask three questions. Is it for the US or the world? Is it a measurement of a past year or a forecast? And if it is a forecast, who made it, in which year, and what range did they give? A forecast is a scenario, not a promise.

Try it in NodalZen

Size shows how the power of a single GPU cluster adds up, and Run turns that power into energy use and cost per million tokens. Simulations lets you test how growth or delays change the plan.

Sources

Learn the basics Open the planner