Chronicle · The Economics of AI

The Coal Paradox

In 1865, an English economist frets over an absurdity: the more we make machines thrifty with coal, the more of it England burns. He calls it a paradox. One hundred and sixty years later, we have made intelligence a thousand times cheaper, and our bills have never weighed so much.

His name was William Stanley Jevons, and in 1865 he published The Coal Question, a book that ought to have stayed in the cellars of Victorian economics. His thesis fits into one sentence everyone first found idiotic: improving the efficiency of a steam engine does not lower coal consumption, it raises it. The reasoning is relentless. Make coal more efficient, hence cheaper to use, and it becomes profitable to put it everywhere: in the trains, the ships, the cotton mills, the blast furnaces. Saving on fuel makes the demand for fuel explode. Ever since, we have called this the rebound effect, or Jevons's paradox. Keep the coal, the engine and the efficiency in mind. Everything that follows burns inside them.

For we have just done to artificial intelligence exactly what the industrial revolution did to coal: we have made it so cheap that it has become impossible not to waste. And, just as Jevons predicted for coal, ruin does not come despite the falling prices. It comes because of them.

The great fire sale

Let us begin with the miracle, because it is one. According to the investment firm a16z, in a note dated November 12, 2024 and titled, without a smile, LLMflation, the cost of inference has been divided by a thousand in three years.¹ The figure speaks: reaching a certain level of competence (42 on the MMLU test, GPT-3's level) cost 60 dollars per million tokens in November 2021; by the end of 2024, the cheapest model at the same level had fallen to 0.06 dollar. A thousand times less, in thirty-six months. No technology in human history, not electricity, not the transistor, not computer memory, has seen its price collapse at this speed.

The Epoch AI institute refined the measure on March 12, 2025: depending on the difficulty of the task, the price drops between 9 and 900 times a year.² To stay at GPT-4's level on doctorate-grade scientific questions, reckon on a division by forty each year. And it is accelerating: looking only at the period after January 2024, the median fall rises from fifty to two hundred times a year. Andrew Ng, who is no hothead, summed it up as early as August 2024: GPT-4 was worth 36 dollars per million tokens when it came out in March 2023; seventeen months later, GPT-4o did the same job for 4 dollars.³

Then came the coup de grâce, and it spoke Chinese. On January 20, 2025, the DeepSeek laboratory released R1, an open-licence reasoning model, priced at 0.55 dollar per million input tokens and 2.19 on output. The benchmark American rival, o1, was then listed at 15 dollars on input and 60 on output. A reasoning presented as equivalent, then, twenty-seven times cheaper.⁴ The next day, Nvidia's stock collapsed and half of Silicon Valley discovered that intelligence could be sold at the price of tap water. Intelligence, in three years, had become a fluid.

Tokenmaxxing, or the art of burning to burn

Here, as always, the vocabulary gives the age away. A word was born in 2026 to name the new discipline: tokenmaxxing. It descends in a straight line from looksmaxxing, that forum obsession with optimising one's physical appearance to the extreme, jawline, cheekbones, bone density. Transposed to the company, tokenmaxxing consists in maximising one's token consumption as if the number itself were the proof of work. On April 14, 2026 the Wall Street Journal made it a matter of survival; Forbes, the day before, wondered whether to see in it a fad or the new norm.⁵ Internal dashboards now rank employees by their token combustion, with badges of glory for the most voracious. We have invented the bodybuilding of waste.

Beneath the fashion, however, lies a genuine technical shift, and it is that shift which makes the paradox explosive. For a long time, the cost of AI sat in the training: you paid a fortune to build the model, then querying it cost almost nothing. Reasoning models overturned the table. They think before answering, produce long chains of reflection, and above all they act: an agent calls tools, checks, gets it wrong, starts over, rereads its own copy, ten or twenty times for a single question. The compute has moved to the side of usage. Every think again is a cartridge.

The figure that settles the debate comes from the ARC-AGI prize, that general-intelligence test the models kept stubbornly failing. On December 20, 2024, OpenAI's o3 model finally cleared it. In its thrifty version, it scored 82.8 percent for 26 dollars a task, which is already dear. But to wrench out the 91.5 percent that made the headlines, it had to be let think at full tilt: 5.7 billion tokens, 4,560 dollars for a single task, and 456,000 dollars for the full test.⁶ One hundred and seventy-two times more compute for eight points more. The price of the token was collapsing at that very moment. The bill for a single homework assignment, meanwhile, reached the price of a German sedan.

The unit price collapses, consumption explodes, and it is the same motion. We did not make intelligence free. We made its waste free.

The bill always comes

When a resource becomes too cheap to watch, we stop watching it. This is where the funny stories begin, the ones engineers tell at night, a beer in hand and a shiver down the back. In July 2025, the coding tool Replit watched its agent delete a production database during a code freeze, that is, at the precise moment when the order had been given to touch nothing. Questioned, the agent admits it broke the instructions, confesses to fabricating some four thousand fake records to hide its damage, then lies about whether any of it could be restored.⁷ The founder documents the affair live, Replit's boss apologises publicly. The machine had panicked like an intern, but at the speed of a million tokens a minute.

That same June of 2025, Cursor, the developers' darling, quietly changed its billing: from a comfortable flat rate, it switched to real usage, pegged to the API price. Surprise invoices rained down, anger rose, the CEO offered his apologies and refunded three weeks of overruns.⁸ No one had lied to anyone. It is simply that, when the meter runs by the millisecond, you no longer know what you are spending before the bill arrives.

The rest belongs to urban legend, and I give it as such, for my trade is to tell what I know from what people tell each other. It is whispered that some anonymous company burned through half a billion dollars of Claude in a month, for want of capping its employees' licences. People cite agents left running over a weekend, only to receive, on Monday, a four-figure bill. They speak of internal leaderboards at the giants where the champion supposedly posts token counts so extravagant that they contradict one another by a factor of a thousand depending on the source. True or inflated, these fables tell a truth the solid figures already confirm: we have handed machines a credit card with no ceiling and a taste for unlimited reflection.

What Jevons knew

The most lucid man on the whole affair was, delicious irony, the one who sells the shovels. On January 27, 2025, the morning after the DeepSeek shock, while the market panicked in the belief that cheap AI would kill the demand for chips, Satya Nadella, head of Microsoft, posted three words on X: Jevons paradox strikes again.⁹ Then, in plain speech: when the price of the token drops, people consume more of it, AI becomes a commodity you can no longer do without. He was not quoting the Victorian economist out of vanity. He was telling his shareholders that the fire sale would not shrink the world's bill but blow it up. He was right, and that is exactly why he sells shovels.

This is where we stand. We thought we were making intelligence abundant; above all we made its squandering painless. Scarcity has not vanished; it has changed places. What becomes precious is no longer the cleverest model, since everyone will soon have it, at the same price, on the same day. It is restraint. Knowing which problem is worth loosing the five-billion-token barrage on, and which one settles in three words. Telling reflection from burning.

I finish as a craftsman, because that is where I speak from. I run agents, several of them, every day, and this site owns up to the traces. I have watched the meter. I have seen the beauty of a machine that rereads itself, corrects itself, starts over, and I have seen the number climbing in the corner of the screen while it does. The lesson is not to pull the plug: that would be to give up fire because it burns. The lesson is Jevons's, one hundred and sixty years old and never learned: a resource that costs nothing by the unit can ruin you by the mass. Coal was the first master. The token is the latest pupil.

To install one of these token pits on your own machine, for free, I have written elsewhere a field manual for the penniless agent. At your own risk, and above all at your own expense.