Layer 1: Energy

The bottom of the cake is the least glamorous and the most important: electricity. Huang calls energy the binding constraint on AI, the one limit that the others ultimately run into. To see why, you need one idea about how AI is different from older software.

Answers are made, not fetched

Most software you use retrieves things. When you open a web page or play a song, a computer finds a file that already exists and sends it to you. Serving it takes a little energy, but the file was made once, long ago.

AI works differently. When you ask a model a question, there's no stored answer to look up. The model generates the reply, token by token, doing billions of calculations for each one. Huang puts it simply in his essay: "Every response is newly created."

So every answer costs energy at the moment it's made. Ask the same question twice and you pay twice. Multiply that by billions of questions a day, plus AI agents that work for minutes or hours on a single task, and electricity becomes the raw material of intelligence.

How much power are we talking about?

The International Energy Agency (IEA) estimates that the world's data centers used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of all the electricity the world uses. Not all of that is AI, since data centers also run email, streaming and websites. But AI is the fastest-growing part. In the IEA's base case, data-center use more than doubles to around 945 TWh by 2030, just under 3% of the world's electricity, with AI as the main driver.

The growth is also concentrated. The IEA expects the United States to account for the largest share of the increase, and there data centers make up nearly half of all growth in electricity demand to 2030.

Why energy is the constraint

The other layers can be scaled with money and engineering: design a better chip, order more servers, write a better model. Power is harder:

• It's slow to build. New power plants and transmission lines take years to plan, approve and construct. A data center can go up faster than the grid connection it needs.
• It's local. A data center needs a large, steady supply at one site. Plenty of power elsewhere in the country doesn't help.
• It comes with heat. Almost all the electricity a chip uses ends up as heat, which has to be carried away by cooling systems that use more power and often water.

That's why AI companies now sign long-term deals for power, from solar and wind farms to nuclear plants, and why "where can we get power?" often decides where a data center is built.