The Direction Economy

Why the next divide between nations will not be who adopts AI, but who can direct it — and why the window to choose is measured in years, not generations.

I · The lag was the mercy

Every shift before this one was slow enough to survive.

Electricity took forty years to reach half of American homes. The factory system took two generations to remake who held economic power and who did not. The internet (the fast one, the one we called a revolution while it was happening) still took the better part of two decades to sort the winners from the left-behind. In every case the slowness did something we never thanked it for. It gave the people at the bottom time. Time to learn the new thing. Time to retool. The lag was the mercy. It is what let the bottom climb toward the top before the gap closed for good.

Artificial intelligence has removed the mercy.

This shift does not trickle down across a generation. It compounds in months. And it compounds in a particular way that the earlier shifts did not — the distance between a person who can direct AI to build something and a person who only consumes what it hands back widens every week, because the capability feeds on itself. The one who directs gets further ahead by directing. The lead is not a head start that the others can close. It is a separation that accelerates while you watch it.

So the question every government and every company is actually facing is not whether to build people who can direct AI. That question is already answered by the speed. The real question is how fast — because the economies that reach what I will call Direction density first may not merely lead. They may foreclose the catching-up that every earlier era still left open.

That is the claim of this paper. Everything after this is the argument for it, the data behind it, and the one decision it forces on anyone who takes it seriously.

II · What AI actually collapses

To see why Direction density matters, you have to see what AI changes underneath the economy — not in the jobs, in the structure.

In 1937 Ronald Coase asked a question that sounds too simple to have won a Nobel Prize. Why do firms exist at all? His answer became the foundation of how we understand the firm. Companies exist because using the market is expensive. Finding the right person, negotiating the price, writing the contract, coordinating the work — all of that carries a cost. When it is cheaper to do the coordinating inside an organisation than to buy it piece by piece, a firm forms. The firm grows until the cost of coordinating one more thing inside equals the cost of buying it outside. That is why companies are the size they are.

The firm was always, at bottom, a machine for coordinating minds. You hired forty people because no single mind could hold the whole problem, so you brought forty minds under one roof and spent enormous effort coordinating them. Most of the apparatus of a company is not doing the work. It is coordinating the people who do the work.

AI does not make those forty minds cheaper to hire. It makes the coordination unnecessary, because one person directing AI now holds the cognitive span that used to need forty. The writing, the analysis, the first draft of the code, the research, the synthesis — one directing mind commands all of it. The expensive thing Coase identified, the coordination of minds, collapses toward zero for the largest category of work in a modern economy.

For most of economic history we counted three factors of production. Land. Labour. Capital. The entire structure of inequality across five thousand years sits on one fact: the people with capital captured the returns, and the people with only labour captured wages, and wages and returns moved apart.

AI does something to that structure that has never happened before. It lets labour command capital-scale output without owning capital. A single person directing AI produces what used to require a team and the money to employ them. Labour and capital, the two factors we kept separate for five thousand years, begin to merge inside one person.

That is the mechanism. The firm shrinks because one directing mind can hold what the firm used to hold. And the moment that becomes possible, one question decides everything — how many people in a population can actually do it? Not how much AI it uses. How many people in it can direct AI to build.

III · Adoption is the wrong number

Here is where almost everyone measuring AI right now is counting the wrong thing.

Open any report on AI and a country. Anthropic publishes how much its model is used per head. OpenAI publishes how many people use ChatGPT. Every consultancy ranks nations by how fast they are taking up the tools. The headlines are all the same shape — this country leads on AI adoption, that one is behind.

Adoption is to directing what GDP is to distribution. It is a big number that hides the thing that matters. A country can top every adoption ranking and be a nation of people who are directed by AI rather than directing it. The usage number cannot tell the difference between a person building a system and a person accepting whatever the model hands back. Both show up as usage. Both inflate the same league table.

This is the same error the financial economy taught us to make. For fifty years we measured economies by output, by the number going up, while underneath, the gains were captured by a narrowing few and the majority's real position stagnated. The number rose. The distribution rotted. GDP saw none of it. Adoption is the same instrument pointed at AI. It sees the volume. It is blind to who is in command of it.

IV · Direction density

The right number is what I call Direction density — the share of a population that can direct AI to build, not merely consume what it produces.

The dividing line is not who uses AI. It is who directs it: who questions the output instead of accepting it, catches where the machine has quietly chosen wrong, and stays in command of work the machine carries out. The technical half (making AI do real work) is no longer the differentiator: the tooling absorbs it, quarter by quarter, and hands it to everyone as a feature. What the tooling cannot absorb is the judgment. Delegation without direction is a skilled consumer at scale. Direction is what turns delegation into building — and the only proof it is real is economic output.

Direction and delegation are independent, and the independence is the whole diagnostic. A nation can delegate massively with weak direction — millions producing with AI at speed, none genuinely evaluating what they produce. That is the most dangerous profile there is, because it looks like productivity and it is actually surrender at scale. A population is only directing when the work it hands to AI is work it still commands.

Why does this determine economic future and not adoption? Because when AI is everywhere and cheap, everyone produces. Production stops being scarce. The thing that becomes scarce is production that holds its value — output that is directed, distinct, not the same thing the machine hands to everyone else. The economic winners in an AI-saturated world are not the ones who produce the most. They are the ones who direct production toward value that sustains. And a population's economic future is the count of how many of them can do it.

This is measurable. That is the part that turns a thesis into an instrument.

V · The instrument: what the data already shows

You can begin to read Direction density out of data that already exists, if you read it correctly. Anthropic's economic data is the sharpest available signal, because Claude's users skew toward people doing technical and professional work — which is exactly the directing end of the spectrum. Israel sits at 4.90 times its population share. Singapore at 4.19. The United States at 3.69. The Gulf far lower — the UAE at 1.61, Saudi Arabia at 0.45. India at 0.22.

Read as adoption, that is just another league table. Read as Direction density, it tells a story that inverts almost everything the adoption headlines say. Put penetration and composition together and four kinds of nation appear.

The Diffused-Deep nations. Israel, the United States, Singapore, Switzerland, Canada, Germany, the Nordics. High penetration, and the work skews toward building. What makes them strong is that directing is spread through the working population rather than trapped in a small elite. One detail worth holding: the share of usage that is genuine directing is led not by the United States but by the Nordics — Sweden at 74.6 percent, the US lower at 64.7. The richest and largest is not the most directing. The index sees that.

The bifurcated-elite nations. India ranks first of any country for the share of its AI use devoted to software: 45.2 percent. Vietnam second, Egypt third. The same three hold the highest direction-weighted building quality of any major economy — India at 68.6, ahead of Sweden and the United States. The most building-intensive AI use per person on earth is not in San Francisco. It is in the emerging economies. And every one of them has near-floor penetration. India scores 199 on the bifurcation index, the highest concentration in the dataset: a sliver of top-tier directing minds on top of an enormous population that has not yet entered.

The uniformly-shallow nations. The Gulf. Real adoption, but the composition is shallow across the whole stack. There is no building elite forming. The genuine consumption case — high use, no depth anywhere.

The thin nations. Low on every axis — the base has not formed at all.

The single most important line in that data is this. The deepest directing capability anywhere is trapped exactly where its diffusion is lowest. No adoption ranking can see this, because adoption collapses the thin-but-brilliant and the broad-but-shallow into the same crude measure of volume. Direction density separates them. And the separation is where the entire opportunity sits.

I will be honest about the limit of this instrument, because the honesty is what makes it defensible. The honest scope of this index, version one, is that it measures the Direction density of a nation's serious AI-working population, using the tool that population actually builds with. It does not yet measure the whole population. Naming that boundary is not a weakness of the instrument. It is the condition of trusting it.

The evidence. This argument rests on real data. See where every nation stands in the Direction Index →

VI · What Direction density decides

Now the part that reaches past the data, because the data only tells you where nations stand today. What it leads to is not one future. It is two, and which one a nation falls into is decided by a single variable that has nothing to do with how many directing minds it has.

The real economy (making things, providing services, employing people) is worth around 123 trillion dollars. The financial economy, the economy of assets and claims, is worth somewhere above 300 trillion. For the whole of modern history the majority of human beings had access only to the first. They earned wages in the real economy. They never touched the returns of the capital economy. The wall between the two economies is the deepest structural fact the majority lives with.

AI cracks that wall, or it builds it higher. Both are possible. Which one happens is the fork.

When a person crosses into directing, when labour and capital merge inside them and they direct AI to produce capital-scale output, they earn, for the first time, something that is not a wage. They earn a return on a system they command. Multiply it across a population and you have the most democratic redistribution of economic access in human history. That is one future.

Here is the other. The AI runs on something: on models owned by a handful of companies, on compute owned by an even smaller handful. Every unit of value the directing mind produces, a slice flows upward to whoever owns the intelligence they are renting. In this future they do not enter the capital economy. They become a more productive tenant of it, paying rent to the new owners of compute. The wall does not fall. It rises higher than it has ever stood, and the people who direct (skilled, productive, even prosperous) are on the wrong side of it, paying to be there.

Same people. Same density. Two opposite civilisations. And the variable that decides which one a nation gets is not its Direction density at all. It is who owns the means of cognition those minds run on. Direction density plus distributed compute (open models, public compute) produces the first future. Direction density plus concentrated compute produces the second.

VII · The lever

The work in front of any nation that takes this seriously is two things at once, and neither one alone is enough.

The first is to build Direction density, fast — to raise the share of a population that can direct AI to build, before the separation widens past closing. This is not adoption. Handing everyone access to a tool produces consumers, not directing minds. It is the harder work of building direction in millions of people — the judgment the tooling cannot hand out with the seats.

The second is to make sure the means of cognition those people run on is not owned entirely by someone else. To distribute compute. To build on open models. Because Direction density without owned compute produces the wrong future. The people who direct get rich on someone else's land. Both, or neither works.

There is proof this can be done deliberately, and it is sitting in the data. Estonia has no business sitting among Israel and Singapore on any measure of technical capability. It sits there anyway, because it chose decades ago to build a digital society on purpose. A small nation manufactured Direction density above its station through deliberate policy. That is the existence proof.

This is where I name the place I come from. Kerala is a state of thirty-five million people, almost entirely literate, with a tradition of building human capability through public action. It is now building public compute, training open models, and putting AI into its schools. The pieces of the two-part lever are, for the first time, in one place — a population that can be raised to Direction density, and a state willing to own the compute it runs on. A nation deliberately building Direction density on compute it owns is the test of which future is reachable.

The instrument follows the argument. Direction density can be read at the level of a nation from the signals that already exist — that is the index this paper has begun. It can be measured properly only at the level of the individual, from the actual evidence of how a person directs AI.

That is the work. Build direction. Own the compute. Measure both. Do it faster than the separation widens.

The lag was the mercy, and the lag is gone. What replaces it is a choice, open now and not for long, between two futures that look identical at the start and could not be more different at the end. The nations that understand they are choosing will choose. The ones that keep watching the adoption number will discover, too late, that they were measuring the wrong thing while the thing that mattered was decided without them.