Try It Yourself

AI runs in buildings, and buildings need electricity. This is the interactive from my essay AI Has a Plug. Follow It. Drag the years, plug in a campus of your own, and run the checklist on a headline.

The numbers sit on the cautious end of public forecasts. The widget shows where each one comes from and how it was built.

Interactive

Follow the Plug

AI runs in buildings, and buildings need electricity. Three ways to see the part of AI you can actually touch: how big it is, how long it takes to power, and how to read the headlines.

How big is it?

How much electricity the world's data centers use, year by year, on the cautious end of public forecasts. Drag the year, or tap a bar.

2019 to 2030 · hatched bars are projections
World data centers, 2025
485 TWh
a year, or about 55 GW around the clock
45M
homes' average draw (1.23 kW each)
60
large reactors running all year
62%
of what all US nuclear plants produce in a year
World data center electricity use, terawatt-hours a year EstimateProjectionPublished figure

Show the numbers

Built from public research by the IEA and Lawrence Berkeley National Laboratory. How each number was made is under "How the numbers were built" at the bottom.

Plug it in

Pretend you're building an AI data center today. Pick a size, then pick how you'll get the power.

Your data center
300 MW
about 0.3 of a large nuclear reactor
244,000
homes' average draw
2,270
AI racks at 132 kW each
Liquid
cooling piped to the chips
How will you get the power?

Yellow squares mark new chip generations. Nvidia says it releases new data center chips on a one-year rhythm, so every year you wait, the hardware you planned around gets a newer sibling.

Run the checklist

Good operators keep a checklist for when things go wrong. Here's a five-item one for AI power headlines. Pick a sample headline and see which questions it answers.

Sample headline, written for this tool

homes' average draw
of all US nuclear capacity (98 GW)

    How the numbers were built
    • Cautious on purpose. Where public forecasts disagree, this uses the most cautious mainstream one. Most others run higher (S&P's 451 Research, McKinsey, JLL and Goldman Sachs's public research among them), so read these bars as the cautious case. The IEA's own lower scenarios exist too.
    • World. Published IEA figures: 415 TWh (2024), 485 TWh (2025) and about 950 TWh in its 2030 base case. 2019 to 2023 are filled in backward at the 12% a year the IEA reports for the previous five years. 2026 to 2029 are filled in at the steady rate that joins 2025 to 2030, about 14% a year.
    • United States. Lawrence Berkeley National Laboratory's estimates of about 76 TWh (2018) and 176 TWh (2023), filled in between at about 18% a year; its low case of 325 TWh for 2028; and about 425 TWh for 2030, calculated from the IEA's US outlook (about 240 TWh added by 2030, a 130% rise on 2024). Years between are filled in at a steady rate. The IEA's own 2024 US estimate, about 185 TWh, sits a little below this path.
    • Conversions. Terawatt-hours a year divided by 8.76 gives average gigawatts. A home averages 1.23 kW (10,791 kWh a year, EIA). US nuclear plants produce about 780 TWh a year from 96 reactors, about 0.93 GW each on average (EIA). Rack counts in panel 2 assume every megawatt goes to computers, before cooling.
    • Not counted. The IEA's figures leave out crypto mining; 451 Research's include it, which explains most of the gap between them (Our World in Data).
    Sources
    1. IEA, Key Questions on Energy and AI (2026): 485 TWh in 2025, about 950 TWh in 2030; and Energy and AI (2025): 415 TWh in 2024, 12% a year over five years, US +240 TWh by 2030; Japan comparison.
    2. Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report: about 76 TWh (2018), 176 TWh (2023), 325 to 580 TWh (2028).
    3. S&P Global, 451 Research forecast: 1,587 TWh in 2030, low-end case, crypto mining included (11.05.2025). Our World in Data on why the two differ.
    4. US Energy Information Administration: nuclear share of generation (19%, 2024) and nuclear generation (about 778 billion kWh, 2021); US nuclear fleet (96 reactors, 98,441 MW, March 2026); average home (10,791 kWh a year, 2022).
    5. Uptime Institute, Global Data Center Survey 2025: average rack density is rising slowly as more racks move into the 10 to 30 kW range; few facilities exceed 30 kW.
    6. DCD, Schneider Electric and Nvidia reference design: up to 132 kW per liquid-cooled rack (12.06.2024).
    7. Enverus, Time to Power: grid waits of five to six years in key markets (04.21.2026).
    8. Modern Power Systems, gas turbine backlog: two to three years, sometimes five or more; slots out to 2030 (02.09.2026).
    9. World Nuclear News, Crane restart: deal 2024, restart targeted for 2027, first planned for 2028 (06.26.2025).
    10. Duke Nicholas Institute via Utility Dive: about 76 GW of room for new demand that can throttle back about 85 hours a year (02.11.2025).
    11. POWER Magazine, transformers: 128-week lead time for power transformers, Q2 2025 (Wood Mackenzie).
    12. NVIDIA, Computex 2024 keynote: "Our company has a one-year rhythm" (06.02.2024).
    13. Grid Strategies via Canary Media: utilities may be overstating data center demand by as much as 40% (11.18.2025).

    Disclosure: I invest in technology companies and may hold positions, directly or through index funds, in companies mentioned in this piece. This is my personal view, not investment advice, and it does not speak for any company I lead. (Punit Dhillon)

    Read the essay: AI Has a Plug. Follow It.