The story in one line

Every time you ask an AI a question, it costs a fraction of what it did a year ago. That collapse in price feels like progress and it is. But it hides a second story unfolding in the opposite direction: training the models behind those cheap answers now costs hundreds of millions of dollars, consumes staggering amounts of compute, and draws ever more power from the grid. Falling prices and rising consumption aren’t a contradiction. They’re the same trend, and we’re only watching half of it.


Chart 1 — The price of using AI fell off a cliff

The cost to use frontier grade AI is racing toward the cost of the electricity it runs on. This is the half of the story everyone celebrates.


Chart 2 — But building models keeps getting more expensive

Three variables at once — time, cost, organisation, and compute (size) — showing the frontier pulling away into nine-figure territory only a handful of labs can fund.


Chart 3 — The divergence, side by side

The same calendar years, opposite directions. The headlines track the left panel; the strategic risk lives in the right one.


Chart 4 — Compute (and the power behind it) keeps exploding

Time x compute x power in one view: each new frontier model needs more compute, and more compute means more megawatts.


Chart 5 — Jevons’ trap: more efficient and more total demand

Efficiency rising has not reduced consumption , it has unleashed it. That is the blindsided beat: the cheaper intelligence gets, the more of everything (compute, energy, water, capital) the world pours into it.