The Invisible Trillion · The Energy Question, Honestly

Okay, But What About Training?

Here is where the numbers get more serious, and where I have to be honest about what we don’t know.

The AI models you use — ChatGPT, Claude, Gemini, whichever — were trained before anyone ever talked to them. Training is the expensive part. It involves running thousands of specialized chips, nonstop, for weeks or months, processing trillions of words of text.

The companies that train these models do not, for the most part, publish their energy numbers. We have estimates.

A model the size of GPT-4, the best public estimate is that training used roughly 50 gigawatt-hours of electricity. A gigawatt-hour is a million kilowatt-hours, so this is 50 million kWh. A big number.

Here is what 50 million kWh is, in context:

• About what 5,000 American homes use in a year.
• About what one large airport uses in a year.
• About what 60 fully-loaded transatlantic flights use, total.
• About one-seventh of the electricity used in Las Vegas in a single day.
• About 0.0002% of global annual electricity consumption.

Here is the other thing you need to know about training. It happens once. The model you talk to today was trained a year ago, and the energy was spent then, and it’s not being spent again. Every conversation any person has with the model from now until it’s retired comes out of that fixed training cost, spread across hundreds of billions of conversations.

Amortized across the conversations it will eventually have, the training cost of a single query is very small. Much smaller than the inference cost — the cost of actually generating the response you see. The headlines about AI training’s energy use are real, but they are a one-time capital expense, not a recurring cost.

Data Centers, Globally

Now zoom all the way out. All the data centers in the world — every Google search, every Instagram photo, every streaming show, every AI model, every email, every spreadsheet stored in the cloud, every video call, every game, every bank transaction, every hospital record — together use about 1 to 1.5% of global electricity. Current estimates are around 300–400 terawatt-hours a year, out of a global total of around 28,000.

For comparison:

• Residential lighting worldwide: about 4%.
• Cement production: about 2%.
• Aluminum smelting: about 1%.
• Cryptocurrency mining: about 0.5%.
• Air conditioning: about 10% and rising fast.
• Livestock agriculture (all energy inputs): about 15%.

Data centers are real. They use real energy. But they are not, in 2026, remotely the largest use of electricity on Earth, and they are not obviously the fastest-growing, and they are not going to “boil the oceans,” which is a phrase you may have encountered online and which does not correspond to any number anyone can produce.

Where the number is rising fast is AI specifically. The best estimates suggest that AI-related data center load could grow from maybe 10% of total data center energy today to 30% or 40% by 2030. That’s the part of the story that has the energy analysts actually worried. Whether that growth is fine or alarming depends on two things: how fast the grid decarbonizes (how much of that electricity comes from the sun, wind, and atom rather than gas and coal), and how fast AI chip efficiency improves (which, historically, has been very fast — AI chips have gotten roughly 30% more efficient per year for the last decade).

Both of those trajectories are uncertain, and honest people disagree about them. If you hear someone claim certainty in either direction, they are probably selling something.

All 19 chapters of The Invisible Trillion · Library · All courses