Stephan Schmidt - April 4, 2026

Everyone Gets Jevons Paradox Wrong

Every time I read another Jevons paradox take on AI and developers, I want to cry. There is a paradox. But it dawned on me that everyone is applying it the wrong way.

What is the Jevons paradox?

William Stanley Jevons wrote in “The Coal Question”:

A> “It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth. As a rule, new modes of economy will lead to an increase of consumption according to a principle recognised in many parallel instances.”

Or in the words of Wikipedia:

A> “In economics, the Jevons paradox, or Jevons effect, is said to occur when technological improvements that increase the efficiency of a resource’s use lead to a rise, rather than a fall, in total consumption of that resource.”

When steam engines become more efficient, people use more coal, not less, even when they need less to accomplish a unit of work. The paradox is famously quoted in 2025 because Satya Nadella reached for it the day DeepSeek shipped: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” That tweet put Jevons on every CEO’s lips for a month.

Then the argument hops one square sideways and gets sloppy. It goes: AI makes developers more efficient, therefore by Jevons we get more software, therefore we need more developers, not fewer. This is the take I get as a reply whenever I talk about the future of software development.

It is wrong.

Before I get to my own take, here are the four positions people are actually defending out there. I went looking. The cheerful “we’ll need more developers” version is one of them, not all of them.

1. The pure optimist: more efficiency, more code, more developers. Nadella is the cleanest version. Michael Becker at HackerRank pulls the historical data — every previous “coding will die” wave (assembly, IDEs, cloud) produced more coders, not fewer. Charles Rubenfeld writes the long-bet version: “more software engineers in five years than there are today.” Sources: Nadella on X, Becker / HackerRank, Rubenfeld.

2. The price-collapse take: code gets cheap, demand explodes, jury still out on jobs. Simon Willison wrote this one early: “the price for writing code will fall, in a way that massively increases the demand for custom solutions.” He is honest enough not to claim he knows what that does to the developer headcount. Mike Grouchy adds that software demand is “closer to infinite than fixed” and that the real risk isn’t unemployment, it’s “more software than our organizations can safely operate, secure, and maintain.” Benedict Evans gives it the historical frame: 150 years of white collar automation produced more clerks and accountants, not fewer. Sources: Willison, Grouchy, Evans.

3. The reshufflers: activity explodes, the job category fragments. David Jayatillake calls this “Jevons Inversion” and writes the line everyone in the optimist camp tiptoes around: “The demand for software engineering explodes. The demand for software engineers, as we’ve known them, may not.” Jim Rutt makes the harder version: “The people who move up are not necessarily the same people who held the jobs before.” This is reshuffling, not preservation. Sources: Jayatillake, Rutt.

4. The skeptics: Jevons is the wrong tool entirely. Madhavi Venkatesan, an economist at Northeastern, says the 19th century model misses too much — externalities, disposal, energy, the whole social envelope around AI doesn’t map onto the market frame. Luccioni, Strubell and Crawford go further in a 2025 FAccT paper: Jevons is being weaponized rhetorically by both AI optimists and AI critics, and the honest move is full lifecycle analysis, not slogans. Carlos Perez takes the dark version: Jevons applies, but to compute and data centers, not developers. Sources: Venkatesan via Northeastern, Luccioni et al. (arXiv:2501.16548), Perez.

Four positions. None of them is what I think is happening.

So where is everyone wrong?

The resource is not software. The resource is not the developer. The resource is ideas. The output of the machine is software.

And here is the part that the cheerful Jevons takes refuse to look at. The machine — the thing that turns ideas into software — has historically been the software development department. Product managers, designers, developers, QA, ops. A whole org chart. That is the engine.

Now go back to how Jevons’ mechanism actually works. Steam engines get more efficient. The unit cost of coal-powered work drops. ROI on owning a steam engine goes up. So more factories buy them — entire new industries that couldn’t justify the cost before now can. The installed base explodes. Each new installation comes packaged with operators, because the engine and the human tending it were inseparable. So you get more engines, more coal and more operators. Resource use, output, and labor all rise together. The growth in the installed base swamps the per-unit efficiency gain. That’s Jevons.

Now run the same mechanism on AI software. Cost of producing software drops. ROI on commissioning custom software goes up. New “buyers” come out of the woodwork who never would have paid for software before - who used standard SaaS software - the restaurant owner with a supply chain quirk, the coach with a custom client management app, the teacher with a specific gradebook in mind. Total software produced explodes. Jevons applies. The optimist is right about that part.

Then the optimist leaps to “and so we need more developers.” This is where the analogy quietly fails. In Jevons’ coal case, every new steam engine came with a human operator as part of the package, because the machine could not run itself. In the AI case, the new buyers do not bring developer hiring with them. They never had developers to begin with, and they aren’t going to start. The “machine” they’re buying is the labor. The new wave of adoption — the part doing the heaviest lifting in the Jevons rebound — adds zero developer headcount.

And on the existing side, inside the companies that already have engineering orgs, the same cost drop applies. Output per developer rises. Marginal new projects get done with the same people, or fewer. Net headcount somewhere between flat and falling.

Both sides of the ledger push the same direction. Fewer developers, not more. The Jevons rebound is real. It’s just landing on the ideas side, not on the labor side. And the optimist position is loudest precisely from the people whose audience most wants to hear the opposite.

I am living the example. A year ago I started building things I would never have built before. A newsletter tool because the existing options didn’t fit how I work. A small SaaS for managing my coaching clients. Blog utilities. An HTML compressor. A CDN uploader. None of them are products. They are tools with an audience of exactly one — me — built because AI made building cheap enough that subscribing to someone else’s version stopped making sense.

That is not me being more productive as a developer. I am a coach who happens to know how to code. The relevant skill today is not my typing. It is knowing what I want Not more developers. More creators.

When AI removes the “typing bottleneck” (ask the CEO for what he pays developers), the new bottleneck is intent. Intent has always lived with the person who owns the problem, not the person who owns the keyboard. So even if you accept the Jevons rebound — and I do — the rebound flows to the people who hold intent, not to the people inside the dev department.

So what to do about it.

For developers: stop optimizing your “typing” by using AI. The valuable skill is no longer “I can build this.” It is “I know what is worth building.” Sit in the customer calls. Learn what makes money. The developer who survives is the developer who has become indistinguishable from the founder. Some of you were already heading there. Hurry up.

For CTOs: your job in two years is not running an engineering org the way you run one today. It is helping the idea-holders in your company become builders themselves, and operating the small specialist core that keeps the result from killing the company. Mike Grouchy named the right risk — “more software than our organizations can safely operate, secure, and maintain.” That risk is your problem now. Start building the muscle for it.

Jevons was right. The four positions in the wild are all wrong about what is actually getting consumed. The resource is not developer hours. The resource is ideas, the engine is the dev department, and AI is shrinking the engine faster than rising demand can grow it. For a growing share of users it is removing the engine entirely.

Pick carefully whose argument you are borrowing.

The three positions at a glance:

Further reading. All eleven sources behind the four positions above, in roughly the order I’d read them:

• Simon Willison, “Jevons paradox”
• David Jayatillake, “Jevons Inversion”
• Mike Grouchy, “AI Enabled Software Development and Jevons Paradox”
• Benedict Evans, “AI and the automation of work”
• Jim Rutt, “Jevons’ Paradox and the Fate of Software Developers”
• Charles Rubenfeld, “Jevons Paradox: The Most Important Idea in AI”
• Michael Becker, HackerRank
• Carlos Perez, “How AI Coding Tools Could Devour More Than They Save”
• Northeastern Global News on Madhavi Venkatesan
• Luccioni, Strubell, Crawford, “From Efficiency Gains to Rebound Effects” (arXiv:2501.16548)
• Satya Nadella on X, Jan 27, 2025