Jevons Paradox and AI Demand: The Condition Nobody States
5 min read · updated August 3, 2026
The paradox is quoted as though it said efficiency always increases consumption. It does not. It identifies a condition under which that happens, the condition is a single inequality, and whether it holds is an empirical question about a specific market rather than a law.
What the paradox actually claims
The original observation was that improvements in the efficiency of coal-fired engines were followed by more coal being burned, not less. The mechanism: efficiency lowers the effective price of the service the resource provides, a lower price raises quantity demanded, and the rise in quantity can more than offset the fall in resource used per unit.
Two things follow immediately and are routinely lost. It is a claim about total consumption, not about efficiency being useless — each unit genuinely does cost less. And it is contingent: the offset can be partial, exact or more than complete, and which one you get depends on how responsive demand is.
The contingency is why the paradox is so often misused. Quoted as a law it explains any observation in advance and therefore explains nothing: if spending rises it is Jevons, and if spending falls it is efficiency working. Stated as a condition it becomes a prediction that could turn out wrong, which is the only kind worth making. The rest of this page is the condition, and then the reasons demand for inference might or might not satisfy it.
The condition, derived
Take a constant-elasticity demand curve, which is the simplest thing that lets the question be answered exactly:
Demand: q = k * p^e with e < 0
Total spend: S = p * q = k * p^(1+e)
Differentiate with respect to p:
dS/dp = k * (1+e) * p^e
p^e > 0 always, so the sign of dS/dp is the
sign of (1 + e).
e > -1 (inelastic) -> dS/dp > 0
price falls, spend falls
e = -1 (unit) -> spend unchanged
e < -1 (elastic) -> dS/dp < 0
price falls, spend RISES
So the Jevons outcome requires |e| > 1: demand must
rise more than proportionally to the fall in price.That is the whole of it. “Cheaper inference will increase total spending on inference” is precisely the claim that the elasticity of demand for inference is below minus one. Anyone asserting it without that condition is asserting an empirical claim while presenting it as a logical one.
A worked case to make the magnitudes concrete. Assume price falls by half:
p halves: p1/p0 = 0.5 e = -0.5 : q rises 41% spend = 0.5 * 1.41 = 0.71 (falls) e = -1.0 : q rises 100% spend = 0.5 * 2.00 = 1.00 (flat) e = -2.0 : q rises 300% spend = 0.5 * 4.00 = 2.00 (doubles) e = -3.0 : q rises 700% spend = 0.5 * 8.00 = 4.00 (4x) Same 50% price cut. Four completely different conclusions about total spend, decided by one parameter.
Why AI demand is lumpy
The constant-elasticity curve is a convenience. Demand for inference is better described as a set of thresholds, and this changes the character of the argument in a way that is specific to this market.
A use case is either viable or it is not. Translating every string in a product, summarising every log line, reviewing every contract clause, personalising every page — each of these has a value per task fixed by the business, and each becomes worth doing the moment cost per task drops below that value. Nothing happens as the price falls toward the threshold; then the whole use case switches on at once and its entire volume arrives.
Aggregate demand = sum over use cases i of volume_i if cost_per_task < value_i 0 otherwise The aggregate curve is a staircase, not a line. Locally, between thresholds, demand is inelastic and a price cut lowers spend. Crossing a threshold, measured elasticity is very large and spend jumps.
This resolves an apparent contradiction in how the question gets argued. Someone measuring elasticity in a mature workload sees an inelastic response, because that workload’s volume is set by how many documents exist rather than by price. Someone watching aggregate demand sees enormous responses, because new use cases keep switching on. Both are measuring correctly and they are measuring different things: the intensive margin and the extensive margin. Which use cases are next in line is the subject of what cheap cognition changes.
Direct and indirect rebound
| Kind of rebound | Description |
|---|---|
| Direct | The same use gets used more because it costs less. A summarisation feature that was rationed becomes unlimited. This is the smallest and most predictable effect, and it is bounded by how much of the thing there is to do. |
| Indirect | The money saved is spent on something else that also consumes inference, or a different team starts a project because the budget now allows it. Larger than the direct effect and almost never measured, because it shows up in someone else's cost centre. |
| Structural | The cheap input changes what products exist at all, which creates demand that had no prior level to rebound from. This is where the largest effects live and it is unquantifiable in advance by construction. |
The elasticity that matters to you is measurable, incidentally, and hardly anyone measures it. You have run price changes without calling them that: a rate limit lifted, a model swapped for a cheaper one, a feature moved from a paid tier to a free one, a cap raised. Each is a change in the effective price of an action, and the usage response around it is an elasticity estimate for your own product. Record the before and after volumes at the time rather than reconstructing them later, and after three or four such episodes you will know whether your demand is elastic instead of arguing about whether it should be.
For your own budget, the direct effect is the one you can plan for and it has a blunt implication: a cost optimisation does not reduce spending unless something stops usage from expanding to fill it. If you cut cost per request by half and remove the rate limit that was rationing the feature, the bill does not move. That is not a failure of the optimisation. It is more product at the same price, which is usually what you wanted — but if you promised a saving, promise it as a cost per unit of work, never as a total, and keep the cap that makes the total real.
What it does not predict
- It does not say spend rises forever. The staircase has a top: once every task worth doing is being done, further price cuts reduce spend, because there is nothing left to switch on.
- It does not say the price cut was pointless. Total spend rising while cost per task falls means more work is being done. Confusing spend with cost is the most common misuse of the argument.
- It says nothing about who profits. Elastic demand can raise total spending while competition drives margins to nothing. Volume and profit are separate questions, treated in where value gets captured.
- It ignores complementary bottlenecks. If the downstream constraint is human review capacity, cheap inference cannot raise throughput past it, and demand is inelastic regardless of price.