The candle and the token: what happens when the price of thinking collapses

In March 2023, a million tokens from the best model on earth cost thirty dollars. By the middle of this year, a model as good or better cost under fifty cents for the same million. That is a fall of roughly sixty times in a little over three years, and for a fixed level of capability the price of machine cognition is now dropping by about ten times a year. Over the last two years alone it has fallen by a factor of nearly three hundred.

Numbers that large stop meaning anything, so let me translate. We are not watching a product get cheaper. We are watching the price of thinking collapse. And thinking is not a product, it is an input, as fundamental to an economy as light or power. When the price of a fundamental input collapses, history is unusually clear about what happens next. The trouble is that almost nobody in the current AI argument is reading the right history. They are debating whether this is a bubble. The more useful question is what happened the last few times a basic input became almost free.

The price of light

Start with light, because the economist William Nordhaus already did the work, and the answer is astonishing.

For most of human history, artificial light was brutally expensive. To read this essay by lamplight in 1800 cost a meaningful slice of a day's wage in oil or tallow. Go back further and a night's decent light could cost a labourer hours of work. Nordhaus traced the price of a unit of light from Babylonian lamps to the modern bulb and found it had fallen not by a few multiples but by a factor of many thousands. Light went from a luxury you rationed candle by candle to something so cheap you leave it on in empty rooms without a thought.

Here is the part that matters. Cheap light did not destroy the value in light. It moved it. The candlemaker's trade evaporated, yes. But the abundance of light created the lit city, the night shift, the evening that belonged to ordinary people instead of only the rich, entire industries and ways of living that were impossible when light was dear. The value did not vanish when the price collapsed. It migrated off the candle itself and onto everything the cheap light suddenly made possible. Whoever was still selling candles lost. Whoever built the electrified world on top of near-free light won on a scale the candlemaker could not imagine.

Jevons, live, in 2026

Now the second piece of history, and this one is happening in front of us this year.

In 1865 a young economist named William Stanley Jevons published The Coal Question, and in it he demolished the comfortable assumption of his age. Everyone believed that as steam engines grew more efficient and burned less coal per unit of work, Britain would consume less coal. Jevons showed the exact opposite. Making coal-power cheaper made it worth using for a thousand things it had never been worth using for before, so total consumption did not fall, it soared. Efficiency did not reduce demand. It exploded it. We still call this the Jevons paradox.

Watch it repeat, almost to the letter, with cognition. Per-token prices have fallen by nearly three hundred times in two years. And enterprise spending on AI has gone not down but up, by more than three hundred percent over a comparable window, because the moment intelligence got cheap we started spending it like water. Where a workflow once made a single careful model call, an agentic system now fires ten or twenty without blinking. Cheaper cognition did not shrink the cognition bill. It detonated the demand for cognition. Jevons explained this in 1865 with coal, and we are living the identical curve in 2026 with tokens, apparently without having read him.

That combination, a collapsing unit price and an exploding total demand, is not a contradiction. It is the signature of a fundamental input becoming abundant. It happened with light, with coal-power, with steel, with computation, with bandwidth. It is now happening with thought itself.

Where the value actually goes

Underneath both stories is one economic law, and it is the most useful thing a founder can hold in their head right now. When a fundamental input becomes abundant, its own price falls toward its marginal cost, and the marginal cost of machine cognition is heading toward nothing. The value does not disappear. It migrates to whatever remains scarce alongside the newly cheap thing. It moves to the complements.

Cheap light made the candle worthless and made everything you did in the light valuable. Cheap steel made the girder a commodity and made the architecture, the brand and the distribution valuable. Cheap cognition is doing precisely the same thing as I write this. It is making the raw answer worthless and making everything cognition cannot do priceless.

So name the things it cannot do, because that list is the map of where the money is going. Judgment, the ability to know which of a thousand free and instant answers is the right one. Taste, which is judgment applied to things that cannot be measured. Proprietary data the model was never trained on and cannot get. Distribution, and the trust of a customer who chooses you. And accountability, a human being whose name is on the decision and who can be held responsible when it is wrong. As answers become infinite and free, the scarce and therefore valuable thing is no longer the answer. It is knowing which answer to trust, and being willing to stand behind it.

What founders keep getting backwards

Almost everyone I talk to about this makes one of two symmetrical mistakes, and they are mirror images of the same misreading.

The first mistake is panic. My product is mostly AI, cheap AI will commoditise it, I am doomed. If the entirety of what you sold was the cognition, then yes, the floor is falling out from under you, and it should. But the same collapse is handing you near-free access to a capability that used to cost a fortune, and the value is pooling in the complements you are perfectly placed to own. The candlemaker was doomed. The person who used cheap light to build something new was not.

The second mistake is the opposite, and just as wrong. Intelligence is almost free now, so I will simply sell lots of it and get rich. You cannot get rich selling the one thing whose price is dropping tenfold a year. Nobody got wealthy selling candles once light was cheap. They got wealthy owning what the cheap light made possible.

The only move that works is the same one in both cases. Build on top of the collapsing input, and own the scarce complement. Treat cognition as the commodity it is becoming, and pour your effort into the things that stay scarce: the data nobody else has, the distribution nobody can copy overnight, the judgment and the accountability a model cannot supply. This is not a clever new insight. It is the oldest pattern in economics wearing a 2026 costume, and it is the entire reason we build companies the way we do, keeping the compounding assets and treating the swappable parts as swappable. The model was never the moat. It is now less of a moat than it has ever been.

The last scarce thing

There is a stranger conclusion hiding at the end of this, and I find it quietly hopeful.

If cognition is becoming abundant and nearly free, then the final scarce resource in the economy is human judgment that someone can be held accountable for. The one thing an infinitely cheap oracle cannot provide is a person to trust, a name on the decision, a throat to choke when the call goes wrong. In an age of unlimited cheap intelligence, the most valuable thing a human being can offer turns out not to be intelligence at all. It is taste, judgment, and the willingness to be accountable for a decision that a machine could describe but not own.

The candle got cheap, and we built the electric world on top of it. The price of thinking is collapsing now, faster than the price of light ever did. What we choose to build on that abundance is the only question that matters, and the founders who have already understood that the value has moved off the answer and onto the judgment beside it are the ones who will own what comes next. The rest are still selling candles, and wondering why the price keeps falling.

Louis O'Connell-Bristow is a co-founder of Moonlabs, the operator-led AI incubator and academy, and previously built the home.co.uk, Homemove and homedata.co.uk stack. Moonlabs builds and funds AI-first companies whose moat is the scarce complement, not the cheap input. Site: moonlab.ventures.