Before the middle of the eighteenth century, most people lived in near darkness once the sun went down. By 2000, according to the economic historians Roger Fouquet and Peter Pearson, lighting services in the United Kingdom cost less than one three-thousandth of what they had cost in 1800. Nobody spent less on light as a result. Per capita use rose 6,500-fold, and total consumption rose 25,000-fold. The candle gave way to gas, gas gave way to the filament, and every time the price of an hour of light fell, people found more rooms, more streets and more hours to fill with it.
William Nordhaus, who later won a Nobel Prize for other work, drew a lesson from the same history that anyone reading an AI invoice should keep close. In a 1996 essay titled Do Real-Output and Real-Wage Measures Capture Reality? The History of Lighting Suggests Not, he showed that conventional price indexes had followed the price of the fuel, the tallow and the kerosene and the kilowatt, rather than the price of the light the fuel produced. Because lamps kept getting better at turning fuel into light, the indexes vastly overstated what light had actually cost. The error was not a rounding problem. It came from counting the input and calling it the output.
The Fuel Got Cheaper
AI budgets now have their own version of this problem. CloudZero's November AI Economics Pulse, published Tuesday, tracks what companies actually pay per million tokens across the models and routes they use. The average fell to $0.69 in October from $0.86 in September, a 20% drop in one month and 34% since July. Over the same months, the median company in its panel spent more on AI every single month. CloudZero's own summary is the whole story in a sentence: every efficiency gain becomes headroom, and usage fills it.
That is the lighting curve running at the speed of a software release. Accenture's survey of 750 senior executives found the same reflex stated as intent. Nine in ten said cheaper tokens would lead them to expand use cases rather than cut spend, and asked what they would do if prices fell another 25%, only one in twenty said they would bank the savings. Respondents expect token consumption to grow 78% over the next two years, and even after a 19% price decline, they expect aggregate token spend to rise from roughly $2.5 billion to nearly $3.6 billion.
None of this is irrational. Cheap light was one of the great gifts of the industrial economy, and cheap inference may prove to be one of the gifts of this one. The rebound is not the problem. The problem is what gets measured while it happens.
Counting Fuel and Calling It Light
The same Accenture research found that less than one dollar in five of enterprise token spend can be traced to a quantified financial outcome: revenue influenced, cost avoided, or productivity turned into a figure finance can act on. Only 35% of companies can calculate the cost per business outcome even for their largest AI use case, and just 7% charge token costs back to the teams that consume them. Cloudera's State of Enterprise AI survey found that 89% of respondents have limited or redesigned AI use cases because of rising token consumption, while only 19% continuously monitor token and inference costs, and 58% say their AI initiatives are generating about as much value as they cost.
Read together, those numbers describe organizations making the mistake Nordhaus caught in the price statistics. They know the price of the fuel to the fourth decimal place. They do not know the price of the light. A falling rate card looks like efficiency, a rising invoice looks like waste, and neither tells a leader whether a single brief, proposal or service answer got cheaper or better.
For a marketing leader, the light is not abstract. It is a campaign brief that went from request to approved draft, a proposal that reached the buying committee a day sooner, a product page that answers the questions an assistant will ask about it, a renewal that was saved. Each of those has a cost today, in hours and in agency fees, and each can be priced again once agents are doing part of the work. That second price, the cost of one accepted unit of work, is the only number that says whether the AI program is buying light or simply burning fuel.
Meter the Room, Not the Candle
The practical shift is in what gets a budget line. A token budget is a candle budget: it caps the fuel and says nothing about which rooms deserve to be lit. The reflex when the bill surprises finance is a curfew. In CloudZero's survey of 252 finance leaders, 27% named new approval gates as their main change for the coming fiscal year, and the teams leading with gates killed or paused AI initiatives at a rate of 20%, against 7% for teams that tied funding to return. A curfew darkens the useful rooms along with the wasteful ones.
The better instrument is a meter on each room. Name the workflows where agents do real marketing work, give each one an owner, and attach its token cost to the output it produced: cost per approved asset, per qualified lead routed, per page shipped and kept. Then let the falling price do what falling prices have always done, which is to make more work worth doing, while the meter shows which work it was.
The households of 1800 rationed candles because light was dear. Their descendants stopped thinking about the price of light at all, and lit everything. Most companies sit somewhere in between with AI, watching the price of the fuel fall and the bill rise, and mistaking one number for the other. The fuel will keep getting cheaper. Whether the house gets any brighter is a question only the meter on the room can answer.
