essay
published
november 2026
reading time
6 min
subject
energy / intelligence

intelligence too cheap to meter.

A token is a small quantity of electricity turned into reasoning. Anything that makes that conversion cheaper changes what the world can afford to try.

listen / 8:18

the cost of a thought

A token is a small amount of computation, and computation is electricity converted into an answer. So the cost of a token is decided in three places: how much electricity the silicon needs for a given result, how much of that electricity becomes useful work, and what the site pays for the electricity in the first place.

The first two are public. They appear in model releases and benchmark tables, and in the habit of announcing that a smaller model now matches the large one it replaced. The third is private. It is settled in a utility tariff, signed years before anyone asks a model a question, and it holds a large part of the remaining cost of intelligence. That is the part we work on.

what the curve actually is

Each fall in the cost of a token moves a different set of uses across the line from unreasonable to routine. Reading an entire codebase, running an agent through a long loop, checking every document in a file room, translating a conversation while it happens: at one price these are demonstrations, at another they are ordinary product decisions.

The pattern is older than machine learning. When the cost of a basic operation falls by a large factor, people stop rationing it and start spending it on things that were not worth attempting before. The consequences that matter arrive later, when the habits of scarcity stop making sense and somebody points the new budget at a question nobody had priced as affordable.

a joule is a decision

Electricity enters the cost of intelligence at the site, not at the lab. Two facilities running the same model can pay very different amounts for the same work, because they sit on different rate classes, carry different load shapes, and signed their power contracts years apart. Both operators will describe the price of intelligence as a property of the model.

Efficiency at the model layer is a public achievement, and it compounds for everyone at once. Efficiency in the power contract is a private one. It is negotiated once per site, and then it applies to every improvement upstream for the life of that contract.

more science

Science is a search with expensive steps. Every hypothesis costs something to test, and the boundary of what gets tested sits wherever that cost meets the budget of the lab doing the testing. Lower the cost of an attempt and hypotheses that were not worth the electricity get attempted. Most of them fail, which is what a search is supposed to look like.

A field that can afford many failed attempts learns at a different rate from one that can afford a few. The difference does not show up in any single paper. It shows up a decade later, in the results that exist because somebody could afford to be wrong.

A field that can afford more failed attempts learns faster than one that cannot.

cures

Medicine is where this matters most, and where it is hardest to say precisely. Drug discovery, trial design, molecule screening and the reading of a literature too large for any person to hold are all searches over spaces too big to walk by hand. In each case the binding constraint is not the quality of the question. It is how many attempts the budget permits.

Make each attempt cheaper and the field gets more shots at the same problem. That is the chain we work on: cheaper power, cheaper computation, more attempts, and eventually a compound in a vial that nobody could have afforded to look for.

abundance

Abundance gets used carelessly, so here it is in concrete terms. It means a research group in a smaller economy can run the experiments a well funded one runs. It means a clinic can afford to check a diagnosis twice. It means the cost of trying something unproven stops being the reason not to try it.

None of that requires intelligence to be free. It requires intelligence to be cheap enough that the decision to use it stops being interesting, the way nobody now weighs the cost of a search before typing one.

the machines themselves

The useful way to describe a language model is not as a store of text but as a device for producing a plausible next step. Given a partly specified problem it proposes a continuation, and the proposal is sometimes good enough to be worth checking. That sounds like a small thing to build a decade on. It is also the first machine that returns something other than an exact answer. It returns a candidate.

A candidate is the unit of research. Before a result there is a guess, and before a guess there is somebody willing to spend an afternoon on it. A machine that generates candidates for the price of electricity changes the supply of the cheapest input into science. That is not the same as doing the science, but it is close to the beginning of it.

There is a stranger point underneath. Nobody designed these systems to reason about proteins or tariffs. The reasoning showed up because the arithmetic that predicts a next word also turns out to be useful for holding a problem in place, which is the part that should make anyone cautious about predicting what the next round will be good at.

what the future could hold

The futures worth arguing about are the boring ones described precisely. A biologist who can ask what is known about a protein family and get the state of the literature instead of a list of links. An engineer whose design is evaluated against physics before anything is fabricated. A clinician whose first read of an image comes with a second opinion that never gets tired and never rushes.

Further out the shape stays the same at a larger scale. Experiments that run in simulation until only the promising ones need atoms. Materials found against a target rather than by proximity to whatever is already known. Drug candidates whose failures happen in a model instead of in a person. None of these is a single breakthrough. Each is a change in how many attempts a field can afford, and that is how fields have always advanced.

There is a version of the future where intelligence stays expensive because power stays expensive, and the benefits collect in the places that could already pay for electricity. Which of those arrives depends less on model quality than on the cost of the substrate underneath the model. That cost is not fixed. It is a matter of contracts, rate classes and arithmetic, and it is being settled right now for sites that will run for twenty years.

the part that stays expensive

Cheap thinking does not make atoms cheaper. A trial still takes years to recruit, a fab still takes years to build, a molecule still has to be made and tested in something wet. What falls is the cost of deciding what to try and the cost of being wrong early, which is why the effect appears as a change in the rate of attempts rather than in the price of any finished product.

That distinction matters to anyone planning a decade out. The binding constraint on intelligence will not be the price of a token for very long. It will be power, fabrication, regulation, and the human attention required to check what a machine proposes. Power sits at the front of that list, and it is the one item that can be fixed with paperwork and arithmetic rather than a new factory. Which is why we spend our time there.