Writing · Energy
AI runs on electricity
Data centres will use about twice as much power by 2030. Microsoft has bought 20 years of power from one reactor at Three Mile Island, and Google and Amazon back small reactors. AI now waits on the plug.

For most of the past decade the scarce input of artificial intelligence was chips. Increasingly it is electricity. In April 2026 the International Energy Agency put the use of data centres at about 485 terawatt-hours in 2025, 17% more than a year earlier, while the world’s demand for electricity grew by 3%. It expects the total to roughly double by 2030, to about 950 TWh, slightly more than Japan uses in a year. AI is the main reason.
The averages hide the real problem, because data centres cluster. A typical AI data centre uses as much power as 100,000 households, the IEA says, and the largest ones being built will use 20 times as much. A grid can add a city’s worth of demand over decades. It struggles to add it in two years at a single substation.
The cloud signs for reactors
The biggest buyers have drawn their own conclusion. In September 2024 Microsoft signed a 20-year agreement with Constellation Energy to restart Unit 1 at Three Mile Island in Pennsylvania, an 835-megawatt reactor closed for economic reasons in 2019. Renamed the Crane Clean Energy Center, it is due back on the grid in 2027.
Less than a month later Google agreed to buy power from a fleet of small modular reactors to be built by Kairos Power, up to 500 MW, with the first unit planned by 2030. Two days after that Amazon anchored a financing round of about $500m for X-energy, with a goal of more than 5 gigawatts of small reactors in America by 2039.
Companies that replace their products every year or two are signing for twenty years of power. That tells you how scarce they think firm, clean power has become.
Two kinds of founder
For energy, the customer has changed. A data centre wants clean power around the clock, close to where it stands, and it will pay for certainty. Small reactors, heat recovery, storage and the software that runs them now have a buyer who signs for decades.
For software, energy is now part of the product. A model that does the same work with a tenth of the compute is cheaper to run and easier to place. That turns efficiency into a sales argument.
Both sides are in our portfolio, from computing that runs on clean energy to a compact reactor designed for data centres. And we now ask every AI founder a question that would have sounded odd five years ago: where does your power come from, and what happens when it costs twice as much?
Sources
- International Energy Agency: Key questions on energy and AI, April 2026
- International Energy Agency: Energy and AI, April 2025
- Constellation Energy: Crane Clean Energy Center
- Engineering News-Record: Three Mile Island unit eyes $1.6B revamp and restart in Microsoft deal, September 2024
- E&E News: Three Mile Island restart closes in on NRC approval, June 2026
- Google: New nuclear clean energy agreement with Kairos Power, October 2024
- X-energy: Amazon invests in X-energy, 16 October 2024
More writing
RoboticsRobots learn from peopleLanguage models learned from trillions of words. Robots have no such library yet. In robotics the scarce thing is now data on how people work.
CapitalEfficiency is the new moatDeepSeek wiped $589bn off Nvidia in a day by doing more with less. The lesson for founders is older than AI: whoever needs the least money is hardest to catch.
FoundersA business plan is a test of judgementEisenhower said plans are worthless but planning is everything. We read business plans for the planning. Here is what we look at first.