Jevons paradox and the AI infrastructure bet: the 160 year old idea behind hundreds of billions in spending

Jevons paradox is an observation from 1865 about coal, and it is currently doing more work than almost any other idea in the argument over AI infrastructure spending. Hundreds of billions of dollars are being committed on the strength of it, mostly without anyone naming it out loud.

It is worth understanding, because it either makes the current spending rational or it is the exact assumption that turns out to be wrong.

What Jevons actually noticed

William Stanley Jevons was an English economist looking at coal consumption during the Industrial Revolution. Steam engines were getting steadily more efficient, using less coal to do the same work, and the reasonable expectation was that coal consumption would fall.

It rose. Sharply.

His explanation is the part that survived. Making something cheaper to use does not reduce how much of it gets used. It makes entirely new uses economically viable, and those new uses swamp the savings. Efficiency is not a brake on consumption. Under the right conditions it is an accelerant.

That is the whole idea. Cheaper coal did not mean less coal. It meant more industries that could afford to run on coal.

Why anyone is applying this to compute

Substitute intelligence for coal and you have the argument being made, implicitly, on every large technology earnings call right now.

The cost of a unit of machine intelligence has been falling fast. If Jevons holds, that does not mean we need less compute. It means whole categories of work that were never worth automating suddenly become worth automating, and demand rises to consume everything that gets built and then asks for more.

If that is true, the companies building data centers at the current pace are not overspending. They are underspending, and the ones who hesitate will spend the next decade renting capacity from the ones who did not.

That is why Alphabet, Meta, Amazon, Microsoft and, on the supply side, Nvidia can absorb visible investor discomfort on earnings calls and keep going. They are not ignoring the question. They have answered it with Jevons, whether or not they use the name.

The seven year problem

There is a structural reason this cannot be decided later.

A data center takes something like seven years to build, from site through power through commissioning. Some get done faster, but that is the shape of it.

Which means the capacity you will need in 2033 has to be committed now, on a forecast rather than an order book. There is no version of this where you wait for demand to prove itself and then respond. By the time the demand is legible, the window to serve it has closed.

This is the same shape as any capacity decision an operator has ever made, just with more zeros and a longer fuse. You commit ahead of certainty or you do not compete.

Google is the reference case people point to. Roughly 20% lifetime return on invested capital, built substantially on infrastructure bets made years before the demand that justified them showed up. That track record is a real argument. It is not a guarantee.

The counterargument nobody should wave away

The obvious objection is dark fiber.

During the dot-com era, telecom companies laid enormous quantities of fiber optic cable on exactly this reasoning. Demand for bandwidth would be effectively unlimited, capacity took years to build, therefore build now. The logic was sound. The capacity was real. Much of it sat unused for years and a lot of companies did not survive the wait.

The lesson from that episode is not that the reasoning was stupid. It is that being directionally right about demand does not save you if you are wrong about the timing, and that the people who eventually profited from the fiber were mostly not the people who paid for it.

So the honest position is that the current buildout could be right about the destination and ruinous about the schedule. Those are different questions and the second one is the one that bankrupts people.

What would have to be true

Strip it back and the entire bet rests on a single assumption: that there is no practical ceiling on demand for intelligence.

Coal had no ceiling for a long time because there was always another industry that could be mechanized. The question is whether intelligence behaves the same way, or whether it saturates.

The case for saturation is that the useful applications are not infinite, that quality stops improving in ways anyone will pay for, or that the work simply runs out.

The case against is where I land, and it is fairly simple. Right now AI is mostly being pointed at work that already exists. It drafts things people were already drafting, reviews things people were already reviewing, answers questions people were already answering. Almost none of the current usage is work that did not exist before.

That is what the early phase of a Jevons dynamic looks like. Coal did not create new industries on day one either. It got cheaper, it displaced existing effort, and then the new industries arrived because the input had become cheap enough to build on.

If that pattern holds, the current demand is the floor and not the ceiling, and the buildout is not big enough.

If it does not hold, a great deal of concrete and silicon is going to sit idle for a long time.

The 5 things I keep coming back to

1. Efficiency does not reduce consumption, it relocates it. This is the piece most people get backwards. Every time inference gets cheaper, the correct expectation is more total compute consumed, not less. If you are forecasting your own AI costs on the assumption that falling prices mean falling bills, you are going to be wrong in the same direction as everyone else.

2. The lead time is doing the arguing. Seven years to build means the decision cannot wait for evidence. When you see a capital commitment that looks reckless against current demand, check the lead time before you judge it. Long lead times force commitment ahead of certainty, and that is a feature of the asset rather than a failure of discipline.

3. Being right about direction is not the same as being right about timing. Dark fiber was correct about bandwidth demand and still destroyed the companies that funded it. Those are separable outcomes and only one of them shows up in your survival.

4. The whole thing rests on one assumption, so name it. No practical ceiling on demand for intelligence. That is the load-bearing claim. Anyone can hold a view on it, but nobody should hold a view on the spending without knowing that this is the thing they are actually taking a position on.

5. Current usage is a replacement pattern, not an expansion pattern. Almost everything AI is doing today is work that already existed. That is what makes me think we are early rather than late, because the expansion phase, if it comes, has not started yet.

FAQ

What is Jevons paradox? The observation, made by economist William Stanley Jevons in 1865, that improving the efficiency with which a resource is used tends to increase total consumption of it rather than reduce it. He documented it with coal during the Industrial Revolution, where more efficient steam engines were followed by sharply higher coal use.

How does Jevons paradox apply to AI? As the cost of machine intelligence falls, uses that were never economically viable become viable, and total demand for compute rises rather than falls. This is the implicit argument behind current AI infrastructure spending, and it is why falling inference costs are not expected to reduce data center demand.

Why are tech companies spending so much on AI infrastructure? Because data centers take roughly seven years to build, so capacity for the early 2030s has to be committed now on a forecast. Waiting for demand to prove itself means missing the window entirely. Google's roughly 20% lifetime return on invested capital, built on early infrastructure bets, is the precedent most often cited.

Is AI infrastructure spending a bubble? It depends on whether demand for intelligence has a practical ceiling. The comparison people reach for is the dot-com dark fiber buildout, which was correct about long-run bandwidth demand and still ruined many of the companies that funded it. Being right about direction and wrong about timing produces the same outcome as being wrong.

What would prove the AI infrastructure bet wrong? Demand saturating. That would mean the useful applications turn out to be finite, or quality stops improving in ways anyone will pay more for. The counter-signal is that current AI usage is overwhelmingly replacing work that already existed rather than creating new categories, which suggests the expansion phase has not started.

Also mentioned

• William Stanley Jevons and The Coal Question, 1865
• Alphabet, Meta, Amazon and Microsoft, and the capital commitments under discussion
• Nvidia, on the supply side of the same trade
• The dot-com era dark fiber buildout, and who eventually profited from it
• Seven year data center lead times, and what they do to decision-making

Read the original

This piece is adapted from The 160 year old idea behind the AI infrastructure boom, published on the 20 Minute COO newsletter on 4 August 2026.

Between Two COO's is hosted by Michael Koenig. Subscribe on Apple Podcasts, Spotify, or wherever you listen.