The Idiot Index of Intelligence

There is a way of seeing the world that, once learned, you cannot switch off.

It works like this. Take anything manufactured — a car, a rocket, a server, a sentence — and refuse to ask what it costs. Ask instead what it is made of. Price the raw materials and nothing else: the metal, the silicon, the energy. Now imagine you could wave a wand and assemble those materials into the finished thing for free, with no labor, no margin, no supply chain, no waste. The number you are left with is the floor — the irreducible cost set by physics alone.

Then compare that floor to the actual price. The ratio between them is one of the most clarifying numbers in existence, because it measures something precise: how much of what you pay is real, and how much is inefficiency wearing the costume of necessity.

When the gap is enormous, it is not a fact about the materials. It is a verdict on the people making the thing.

This is first-principles reasoning applied to cost, and its power is that it ignores the question everyone else is asking. The crowd reasons by analogy — what do comparable things cost, what has this always cost, what is everyone else charging? First-principles reasoning throws all of that out and asks only what the laws of nature actually demand. The entire distance between those two answers is opportunity.

I want to take this lens, which has reorganized one industry after another, and point it somewhere it has never quite been aimed before. Not at a rocket or a car. At a thought.

What is the true cost of intelligence — and how much of what we pay for it is real?

The method, and what it has actually done

Before intelligence, look at what this way of seeing has built, because its track record is the reason to trust it.

Consider Henry Ford in 1908. The automobile of his era was a hand-built luxury, assembled by small teams of craftsmen working a car at a time, priced accordingly. Ford's insight was not to build a better car by analogy to existing cars. It was to decompose the act of making one into its simplest constituent motions, then rebuild the process around the materials and the physics of assembly. When the moving line opened at Highland Park in 1913, it cut the time to build a Model T from 728 minutes to 93. The price followed the same logic downward — from around $850 at launch to $290 by the mid-1920s, in an era when rivals stayed expensive. Ford had found the gap between what a car cost and what a car required, and he closed it. Fifteen million cars later, he had not just built a company; he had created the middle-class motorist as a category of human being.

A century later the same move rebuilt spaceflight — and this time someone gave the gap a name.

When Elon Musk set out to lower the cost of reaching orbit, he refused the industry's central assumption: that a rocket costs what rockets have always cost. That, he argued, is reasoning by analogy, and in his words it really doesn't illustrate what the true potential is. So he asked the rude, literal question instead — what is a rocket actually made of? Aluminum, titanium, copper, carbon fiber. He priced the raw materials, then ran a thought experiment he calls the magic wand number: if you had the materials stacked on the floor and could wave a wand to arrange the atoms into a finished rocket at zero cost, what would the rocket cost? That number is the floor, set by physics and nothing else.

For rockets, the magic wand number turned out to be brutal in its implication — well under 5 percent of the prevailing cost, in some cases closer to 1 or 2 percent. Which meant 95 to 99 percent of what everyone paid for a rocket was not physics at all. It was inefficiency. As Musk put it, if the raw material is only one or two percent of the finished product, the manufacturing must be wildly inefficient — and he could see enormous room for improvement.

Out of that he built a general-purpose tool and gave it a deliberately rude name: the Idiot Index. Take any part, divide its finished cost by the cost of its raw materials. A high ratio is not a fact about the metal; it is a verdict on the design or the process. His own favorite example was a half-nozzle jacket on the rocket that cost $13,000 — and was made of $200 worth of steel. An Idiot Index of sixty-five. His standard was that every engineer should know the best and worst parts in their system by that measure at all times. The name carries the whole philosophy in two words: if the ratio is high, you're the idiot, because the gap between price and material floor is a measure of your own waste, not nature's.

Closing that gap is the reason it now costs a fraction of what it once did to reach orbit, and the reason an entire space economy has bloomed in the room the old prices used to occupy.

Two industries, a century apart, transformed by the same refusal. Ford never used Musk's vocabulary, but he was computing the same ratio — the gap between what a car cost and what a car required — and going to war on everything in between. Both ignored what the thing "should" cost. Both asked what it was made of, found a floor far below the going price, and closed the distance. The lesson is not about cars or rockets. It is that the gap between an item's price and its material floor is almost always larger than anyone believes, and almost always closable — if you reason from the physics up instead of from the market down.

So let us be willing to do exactly that. Let us compute the Idiot Index of intelligence.

The bill of materials for a thought

Start where the method always starts: with the raw materials. What are the physical inputs to a unit of machine intelligence — to one answer, one generated paragraph, one solved problem?

Strip away the mystique and there are only three. Compute — the silicon doing the arithmetic. Energy — the electricity flowing through it and the cooling that carries the heat away. And data — the training corpus, paid for once and amortized across every query that follows. That is the entire bill of materials for a thought. Beneath the wonder of it, an AI's output is a defined quantity of matrix multiplications, drawing a defined number of joules, on a chip that wears out by a defined fraction each time it runs.

The magic wand number for intelligence, then, is what those inputs cost if you could rearrange the electrons for free — the energy and the silicon's wear for a single inference, with no company, no margin, no waste between the power plant and the answer.

And here intelligence reveals something no rocket or car ever did. That floor is not fixed. It is collapsing, on one of the steepest curves in the history of technology.

The numbers strain belief. The price to run a model at GPT-3.5's level fell from $20 per million tokens in late 2022 to about $0.07 two years later — a 280-fold drop. Stanford's AI Index confirms the collapse across every performance tier; Epoch finds the fastest segments falling roughly 200-fold per year. At seven cents per million tokens, a single dollar buys around eleven million words — more than most people read in a decade. The cost of a thought is racing down toward the cost of the electricity required to have it, and the electricity is the only part physics genuinely insists on.

This is the most underabsorbed economic fact of the era: the magic wand number for intelligence is approaching zero, and it is getting there fast.

The paradox that tests the whole idea

If the cost of intelligence is collapsing, our spending on it should be collapsing too. It is doing the exact opposite, and the contradiction is where the Idiot Index proves its worth.

Across the same span that token prices fell 99.7 percent, enterprise AI spending tripled. The average corporate AI budget climbed from $1.2 million in 2024 to $7 million in 2026. The raw material got three hundred times cheaper and the bill went up. Inference now devours around 85 percent of enterprise AI budgets, and most of that spend sits not in the model at all, but in the orchestration, retrieval, retries, and scaffolding wrapped around it.

At first glance this looks like the method failing. The atoms got cheaper; the product got pricier; the whole framework inverted.

Look again, because the lens has not failed — it is doing precisely its job, pointing straight at the waste. When the raw material of thinking becomes nearly free, we do not buy the same amount of thinking for less. We buy wildly more of it. A single agentic task now burns fifty to five hundred times the tokens of a simple query from a few years ago, looping and retrying and calling itself in sprawling cascades. The cost per thought cratered, so we let the machines think extravagantly, often wastefully — and roughly three-quarters of the resulting bill is the overhead around the intelligence rather than the intelligence itself.

That is an enormous Idiot Index, hiding in plain sight. It is Musk's $13,000 nozzle jacket made of $200 of steel — except now it is a seven-figure AI bill resting on a far smaller base of actual compute and energy. The floor fell away, and we poured the savings straight into new layers of waste. The inefficiency did not vanish. It migrated — out of the model, where everyone was watching, and into the architecture nobody was.

The lens did not break. It told us exactly where the waste went to hide.

The analogy trap, which is the real enemy

Here is why this matters far beyond a cost-accounting exercise, and why I think it is the single most valuable habit of mind for anyone building right now.

Musk's deeper point was never really about ratios. It was about which kind of reasoning you trust. The deep error is never the high price itself. It is the reasoning that accepts the high price as natural — reasoning by analogy. Almost everyone thinks about AI this way. They benchmark against last year's AI budget, against a competitor's deployment, against what software "ought to" cost. And analogy is exactly the wrong instrument when the underlying floor is moving two-hundred-fold a year, because every comparison points at a world that has already ceased to exist by the time you finish the sentence. Pricing this year's intelligence against last year's is like pricing a reusable rocket against the Space Shuttle, or a Model T against a hand-built carriage. The comparison feels rigorous and teaches you nothing true.

The first-principles question is harder and far more generative: if a thought costs almost nothing, what becomes worth thinking that was unthinkable before? Not "how do I run my current workload more cheaply" — that is the analogy trap again, lovingly optimizing the thing you already do. The real question is which entire categories of problem crack open once intelligence is functionally free. What do you attempt when you can afford to be wrong a thousand times, simulate every branch, and put a PhD-grade reasoner on a problem far too small to have ever justified one?

This is the exact fork Maxime Labonne and I kept circling: once intelligence is cheap enough to run privately on the device in your pocket, the question stops being "what does AI cost" and becomes "what does a world saturated with nearly free intelligence even look like." A floor falling to zero does not merely shrink a line item. It redraws the map of what is worth doing at all.

Following the floor all the way down

Trace the logic to its end and it arrives somewhere genuinely vertiginous.

If the Idiot Index of intelligence is the gap between what a thought costs and the raw energy of computing it, then the whole industry is in a furious race to drive that ratio toward 1 — to close the gap until the only thing left to pay for is the physics itself. Custom silicon is one front of that war: when Midjourney moved its inference from general-purpose GPUs to Google's TPUs, its monthly bill fell from $2.1 million to under $700,000. The next generation of chips targets another tenfold cut. Each advance strips away one more layer of everything that is not physics, pressing the cost of a thought closer to the cost of the electricity beneath it.

And that is the thread tying this to everything else worth watching. Once the only irreducible input is energy, the entire competitive frontier of intelligence quietly becomes a contest over energy. It is why the same companies racing to lower the price of a token are simultaneously trying to build data centers in orbit and beam solar power down from the sky. They have followed the Idiot Index down through every layer of waste and found, sitting at the very bottom, a single cost that refuses to fall further: the joules. Strip away all the rest and what remains is thermodynamics. The magic wand number for intelligence, in the limit, is an energy bill.

That is the whole game, said plainly. We are watching the cost of thinking get decomposed, stratum by stratum, down toward its physical floor — and discovering that the floor is energy. Which is why the race for cheap intelligence and the race for abundant energy are turning out, underneath, to be the same race.

Where the lens must be put down

Every powerful instrument has a domain where it works and a border past which it turns destructive, and the discipline lies entirely in knowing where that border runs.

The Idiot Index has an obvious next move, and we will not be able to resist making it. If the ratio of finished price to material floor measures waste, then sooner or later we will aim it at human labor. A knowledge task an AI performs for pennies of energy, but for which we pay a salaried professional, will start to look — in the cold arithmetic of the index — exactly like that $13,000 nozzle jacket made of $200 of steel. The framework that reorganized factories and rocketry does not pause to ask whether the inefficiency it has found is a nozzle, an assembly step, or a person's livelihood. It only measures the gap.

This is precisely where the instrument reaches its limit — where reasoning from physics, taken too literally, curdles into its own species of stupidity. Because the Idiot Index measures cost. It is utterly silent on worth. It can name every raw material in a thought and tell you nothing about the value of the human being having one. Its entire power comes from stripping away everything that is not physics — and its entire danger is that a life, a vocation, a sense of being needed by others was never reducible to physics to begin with. Ford, for what it is worth, half-understood this: in the same breath as the assembly line he doubled his workers' wages, grasping that the efficiency was only worth building if the people it displaced could still afford to live inside the world it created.

So here is where I land. The Idiot Index is the best tool I know for understanding what intelligence actually costs, where its waste is hiding, and why the race for cheap thought is becoming a race for cheap energy. Use it without mercy on your architecture. Use it on your assumptions, your pricing, your sense of what is "normal." Reason from the floor up, never from last year's bill down.

But hold its border clearly in view. The number that made cars and rockets cheap is a magnificent measure of cost and a catastrophic measure of worth. The discipline of the coming decade is not learning to calculate the Idiot Index of everything — almost anyone can do that, and many will. The discipline is knowing, with conviction, which things deserve to be measured this way and which are precious precisely because they were never efficient to begin with.

The magic wand number for a thought is approaching zero. The magic wand number for a person was always a category error. Holding both of those truths at once — without letting the first one quietly erase the second — is the actual work of the years ahead.

On Masters of Automation I keep returning to this: the cost of intelligence is collapsing toward the cost of energy, and what we choose to do with that gap is the defining question of the decade. If reasoning about AI from first principles is your kind of question, the newsletter is where it continues. All signal, no spam.