There’s a sentence written in 1865 that explains your AI subscription bill better than anything published this year.
It was written by a 29-year-old economist named William Stanley Jevons, sitting in Manchester, living off savings he’d brought back from Australia, convinced as usual that nobody would read him. He was writing about coal. The sentence goes like this: “It is a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption of it. The very contrary is the truth.”
That sentence, the Jevons Paradox, is why the unit cost of AI intelligence has fallen by a factor of roughly 1,000 in three years, and yet you are paying more to access it every month than you were paying before. Not despite the cost drop. Because of it.
The Day Two Announcements Changed Everything
On January 21, 2025, the day after his inauguration, Donald Trump announced Stargate: $500 billion in committed investment to build American AI infrastructure. The message was this: AI is a power race, and power is bought in gigawatts.
By the most precise coincidence, the night before that announcement, a Chinese research lab nobody outside the industry had heard of published a model called DeepSeek R1. Open-weight, free, capable of matching the best OpenAI models in reasoning — and built for a reported $5.6 million in compute, against the $100-plus million that GPT-4 had cost. The weekend that followed broke the internet. DeepSeek’s app displaced ChatGPT at the top of the American App Store, then did the same in fifty other countries.
The Sunday night before the markets opened, Marc Andreessen called it the AI’s Sputnik moment. On Monday, January 27, Nvidia lost nearly 17% in a single session — roughly $600 billion evaporating in one day, the largest single-session destruction of market value ever recorded for a listed company. The NASDAQ fell 3%. Jensen Huang’s net worth dropped more than $20 billion on paper.
The market’s logic was clean: if frontier intelligence costs 20 times less to produce, then the mountains of GPUs are oversized. Then the $500 billion of Stargate is a monument to excess. Then sell.
But that Sunday night, before the carnage started, Satya Nadella posted a single line: “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.”
Same facts. Diametrically opposite reading. One of them was wrong in a way that cost hundreds of billions of dollars to discover.
Manchester, 1865
To know which reading was right, you need to understand what Jevons actually observed.
Britain in 1865 dominated global industry and trade. Steam-powered cotton, forged steel, propelled ships. Everything ran on coal. And in this context, a young economist noticed something nobody was asking about: every new generation of steam engine consumed less coal per unit of useful work. Engineers were celebrating. The logic seemed airtight: better engines meant less coal was needed.
Jevons looked at the numbers instead. Since the start of the century, Britain’s population had doubled. Coal consumption had multiplied eightfold. The country was extracting close to 100 million tonnes per year, growing at over 3.5% annually, and the pace was accelerating.
He wrote that sentence. Then he explained the mechanism: efficiency lowers the price of useful work. A lower price makes steam profitable in places where it previously wasn’t — pumping deeper mines, pulling longer trains, equipping entire factories. Every time an engineer saved coal, he didn’t reduce England’s bill. He expanded the territory where steam was worth using. As Jevons put it elsewhere in the same book: “The new applications of coal are of an unlimited character.”
The book found an audience. Gladstone, then Chancellor of the Exchequer, invited Jevons to meet him. John Stuart Mill cited the work in Parliament. A royal commission was created to inventory coal reserves. In under a year, an unknown economist had set Parliament into motion.
His conclusion, though, turned out to be wrong in its specifics. Jevons believed England was condemned — that the coal would run out and British greatness with it. He knew about petroleum and electricity but badly underestimated their capacity to reduce coal dependence. He was so convinced paper would also run scarce that he stockpiled it. His children still had not exhausted his reserves fifty years after his death, during which he had been scrawling his notes on the backs of old envelopes.
What the Evidence Actually Says
The Jevons Paradox didn’t stay a Victorian curiosity. After the oil shocks of the 1970s, economists dug it back up, renamed it the rebound effect, and spent decades testing it against data.
Their verdict is more nuanced than either the market panic or Nadella’s triumphant tweet would suggest.
The direct rebound is real, but usually partial. Adding insulation to your house will make heating more affordable, enabling you to potentially extend heating times or increase the thermostat setting. But that rebound eats 10 to 30% of the efficiency gain — not 100%, and certainly not 150%. You don’t triple your heating hours because your boiler became twice as efficient. Your need for warmth has a ceiling. Once the room is comfortable, a cheaper furnace gives you no reason to push to 35 degrees.
Economists, including Steve Sorrell at the University of Sussex, have reviewed a century and a half of data and concluded that a full rebound — where efficiency actually increases total consumption rather than just partially offsetting the savings — is the exception, not the rule. The conditions for it are specific and rare.
When a complete recovery happens, it’s almost always accompanied by the same trio of conditions. The resource must be at the heart of the cost structure. Demand must be far from saturation. And each efficiency gain must open genuinely new uses rather than just making existing ones cheaper.
The steam-coal system of 1865 satisfied all three. And that is the question that decides both your monthly invoice and the entire capital expenditure of this decade.
Three Conditions, All Met
The resource must dominate the cost. When coal powered over 90% of Britain’s energy, and its price flowed directly into every unit of force the steam engine produced, saving coal made the whole system cheaper in a way that immediately translated into more machines, more activity, more coal.
Inference compute today plays a structurally similar role. It accounts for roughly two-thirds of the cost of delivering AI output at scale. When token costs fall, the cost of each additional query falls directly with them. And when organisations reinvest those savings into more queries — which the data shows they do — total compute demand goes up, not down.
First condition: met.
Demand must be far from saturation. Your living room has a comfortable temperature. Once you’ve reached it, cheaper heating gives you no meaningful reason to go higher. But where is the thermostat for intelligence? At what point does a person, a company, a government decide it has enough AI?
Sundar Pichai disclosed the numbers at Google I/O 2026: in May 2024, Google processed 9.7 trillion tokens per month across its AI products. By May 2025, that had risen to 480 trillion. By May 2026, it exceeded 3.2 quadrillion — a 330-fold increase in two years, and a sevenfold increase in the most recent twelve months alone. This happened while token costs were falling sharply and models were getting more capable. If demand were saturating, volumes would plateau as prices dropped. The opposite is happening.
Second condition: met.
Efficiency must open new uses, not just make old ones cheaper. This is the sentence Jevons wrote, I asked you to remember: the new applications of coal are of an unlimited character. Steam efficiency didn’t make existing factories run at a lower cost. It made steam viable in trains, ships, pumping stations, and urban mills that had been impossible before.
AI is playing the same move. At $60 per million tokens, a deep research workflow — where an agent searches, reads, compares, and synthesises hundreds of sources before producing a report — was too expensive for a consumer subscription. At current prices, it’s a standard feature. Reasoning models spend more tokens thinking before answering. Agentic systems multiply those steps across hours of autonomous work. The market’s perception that DeepSeek R1 would eliminate the need for GPUs was mistaken; this new generation of model was engineered to invest more compute in each query to enhance its reasoning abilities. The efficiency didn’t eliminate the spending. It made a vastly more expensive way of producing intelligence affordable.
Third condition: met.
Three for three. The room is never warm enough.
