Musing in Public

Jevons Grows Faster Than Moore Shrinks

The most predictable forecast in technology is that everything gets cheaper. The computer that once filled a room now fits in a pocket. Storage once needed its own line in the budget. Now it is cheap enough for people to hoard thousands of photos they will never look at again.

Most mainstream predictions assume AI will follow the same curve and costs will fall. I’m not so sure.

Two economic ideas are usually invoked as shorthand for falling technology costs. Moore’s Law says each unit of computing gets cheaper as chips improve. Jevons Paradox observes that when something useful becomes cheaper, we find more uses for it and often consume more in total.

The usual simplification assumes that the two combine into cheap abundance. Moore lowers the price. Jevons spreads the benefit, which encourages more production, and everyone gets plentiful flat-screen TVs. This time, I think the interaction may work a little differently.

Jevons will cost Moore. Sorry, I couldn’t resist.

Demand for AI may grow so quickly that total usage outruns the traditional decline in unit cost. The tokens get cheaper, but the machine consumes so many more of them that the sticker shock remains.

There are plenty of fancy charts showing the falling price of tokens and how efficiently the latest models complete fixed tasks. I guarantee the average vibe coder or executive is thinking about neither. They are thinking about the price of the completed project. They want the app built, the bug fixed, the report written, and the meeting summarized so thoroughly that perhaps everyone can finally stop having it.

A token is an ingredient. The bill arrives for the meal.

That distinction matters because the price of the ingredient can fall while the size of the meal grows much faster. The old AI interaction was a question followed by an answer. The new one begins with a request, reads your files, searches the codebase, drafts a plan, writes the code, runs it, discovers it does not work, reads the error, revises the plan, edits seven files, breaks an eighth, fixes that, runs the tests, and asks a second model to review the whole thing before disagreeing with the review and going around again. It costs more because it does more, and doing more is the point.

The simple version of Jevons keeps consumption attached to the same activity. The lightbulb becomes so useful and affordable that you put one in every room. Eventually, every room is lit and your demand for lightbulbs plateaus. Another hundred bulbs would probably make the house uncomfortably bright except at Christmas.

AI behaves differently because its success expands our ambition. New tasks worth attempting get added to the existing list. Existing tasks also grow as capabilities improve. Then every successful project leaves behind code, data, decisions, bugs, and maintenance that require more AI work later. It is classic technical debt, priced in tokens instead of developer hours.

Picture asking the model to build a web-based tic-tac-toe game. Three minutes later, it exists. Your next thought is probably not to commission another forty-nine tic-tac-toe games. You reach higher.

“Recreate my favorite childhood arcade game.”

That works too, so now add multiplayer, accounts, saved progress, a mobile version, achievements, custom levels, a marketplace, and an admin console. Then ask a second model why the first one believes purple gradients are a substitute for taste.

The first success changes your estimate of what is possible, and a bigger vision changes what you want. A more efficient car might lead you to drive farther, but you are still driving. A more capable model leads you to reconsider where driving could go, build three new roads, hire something to inspect them, and start sketching an airport.

The demand does not stay attached to the task that prompted it. It continues to climb higher and higher.

Intelligence has no satiation point.

Every answer suggests three new questions, and every system you build leaves something new to decide about. Once the arcade game becomes a real program, every feature adds more code. More code means more places for bugs to hide, which justifies more tests, which generate more logs, which someone, or rather something, has to read. Every future code review also has to understand a larger system than the review before it. The codebase keeps growing, so even when the model becomes more efficient at reading each file, there are more files to read.

The machine builds the house, enlarges the house, and wins the maintenance contract.

Then the agents arrive. One agent builds the feature. Another reviews the code. Another runs the tests. Another checks security. Each reads the codebase independently, calls tools, checks results, and often revisits the same material several times. As far as I can tell from the logs, parallel productivity requires every robot to read the employee handbook for itself.

Dell offered a nearly perfect illustration. In its account of a 2026 keynote by chief operating officer Jeff Clarke, the company said model prices had dropped roughly 80% over the previous year while the use of reasoning tokens had risen 320 times.¹ A fifth of the price but at 320 times more volume is still a 64x increase in the bill. In other words, demand growth far outran the lower unit cost.²

Once people learn to use agents well, they begin using them for everything. That is why the next surge in AI consumption may come from habit change rather than headcount. Each successful project reveals another use. Each new use leaves behind more work to maintain. Each improvement in capability makes the next project easier to justify.

The faster the flywheel spins, the higher the bill.

Sources