Alden House. Parent Center

The Invisible Trillion · The Energy Question, Honestly

The Skeptic’s Checklist

When someone tells you the energy cost of a technology, here are the questions to ask, in order:

  1. What’s the unit? If they don’t tell you Wh, kWh, MWh, or GWh, they probably don’t know.
  2. Per what? Per query? Per user? Per year? Per training run? A number without a “per” is meaningless.
  3. Compared to what? Five kWh is a lot compared to a phone charge and nothing compared to a flight. Every claim needs a baseline.
  4. Which data center? Powered by what? In what climate? Using what cooling?
  5. Is this the total, or the marginal? The total cost of all data centers is a big number; the marginal cost of your specific use is a small one. Don’t confuse them.
  6. Who’s the source, and what do they want you to conclude? An environmental group estimating high, an AI company estimating low, and a utility grid operator estimating carefully have very different incentives. Read accordingly.

If a claim survives all six questions, it is probably worth taking seriously. Most claims, unfortunately, don’t survive the first two.

What Comes Next

The next chapter is about water; the one after that, mining and e-waste. Each has the same structure: real numbers, honest uncertainty, and a careful separation of the parts that are a problem from the parts that only feel like one.

Several chapters later, the book reaches the parts this one has been gesturing at — the specific costs imposed on the specific people who live near the infrastructure, the architecture that would do less damage if the industry chose to use it, the financial structure of the buildout and who is funding it, and the question of whether the current execution is producing outcomes worth the cost. Those chapters are darker than this one, because the more you learn about the specifics, the harder it becomes to read the global aggregates as the whole story.

One more thing worth flagging, which the rest of the book will come back to: what if the scarce resource isn’t the electricity at all? What if it’s the political legitimacy that gets spent every time a community is forced to host infrastructure it did not consent to? What if the durable cost is not measured in watt-hours but in the corrosion of the systems we rely on to make collective decisions? Hold the question. The book will come back to it.

A note on sources: the per-query AI energy figures are based on published estimates from the International Energy Agency, academic studies (particularly those by Alex de Vries and the Hugging Face team), and leaked or disclosed figures from the major AI companies when available. The 50 GWh training figure for GPT-4-class models is a midpoint of public estimates; the real number could be half or double that. The comparison figures (flights, burgers, AC) are from standard life-cycle assessments. When in doubt, I’ve used the higher end of plausible estimates for AI and the middle of the range for everything else, so that the comparison is conservative in AI’s disfavor. Even with that bias, AI’s per-query energy use is small compared to almost anything else a person does in a day.

Questions to talk about

No answer key. These are for arguing over, with someone who read the same chapter.

  1. The author says most of what you have heard is exaggerated or too vague to be useful — not because people lie, but because energy is invisible. Why does invisibility produce bad numbers specifically?
  2. He gives you a unit first, 'otherwise the numbers are just shapes'. Take an energy claim you have heard and try to put it in watt-hours. Can you?
  3. He deliberately used the highest plausible estimates for AI and middles for everything else, biasing against his own conclusion — and AI still came out small per query. Why is arguing against yourself persuasive?

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