Cheap Composition · 4 of 5 · 16 min

Where the Value Goes

The economics of cheap composition, and what to do about it

Cobus Kok · September 2026

• When output gets cheap, value moves to what stays scarce: compute and distribution, which are capital, and trust and presence, which a person can build.
• The sectors that never got more productive, law, education, healthcare, government, are where the effect lands hardest, in the share of their cost that is making first versions, and the bottom rung goes first.
• The tax base moves with the value, to somewhere harder to tax, and I have not heard a finance minister answer that.

When composition becomes cheap, value moves to whatever is still scarce, and four things are: compute, trust, presence and distribution. Two of those are capital and two are positions a person can take without any, and that difference matters more than the list. The gap between a collapsing cost and a sticky price is margin, and it goes to whoever holds the trust or the distribution. Almost everything below follows from that, including the cases where it doesn't happen.

By composition I mean making the first version of anything, a draft, an analysis, a function, a lesson, which is the work AI now does. Money is where most people's decisions about it live: what to learn, what to build, what to own, what to ask for. It is also the part of the conversation that most often gets waved at, so this essay follows it.

Cognitive deflation#

Start with prices. A unit of composition, a draft, an analysis, a function, a first read of a scan, a lesson, used to cost a professional's hour. The compute for it now costs between a fraction of a cent and a few dollars, depending on how long the machine has to think, so the cost of composition, before anyone checks it, is falling toward the price of electricity.

Economists have a name for the sectors this hits hardest, and it's an old one. In the 1960s William Baumol noticed that some parts of the economy got more productive far more slowly than the rest, because they were made of human hours: education, healthcare, law, the arts, government. Their wages still had to keep up with everyone else's, so their costs rose relative to everything else, and he called it cost disease. For sixty years they kept getting dearer, and the part of them that is composition is about to get cheap, which is where AI's economic effect lands hardest. On most ways of counting, between a quarter and a third of rich-country output is cognitive and professional services of one kind or another, and the productivity growth these sectors lacked is starting to arrive, one task at a time. Not in every cost line, though. A hospital is beds, staff and liability before it is paperwork, and halving a task that is a fifth of a sector's cost saves a tenth. The claim is about the share of each sector's cost that is composition: large in law, administration, education and finance, smaller in the parts of healthcare that are hands.

Two objections deserve an answer before going further. The first is that checking isn't free, and if verifying a machine's output costs as much as producing it by hand, the deflation is smaller than it looks. That is true for work where errors are expensive and hard to find, which is why the trust-gated professions feel this last. But checking can scale differently from producing: a senior who used to review three juniors' drafts can review a great many machine drafts, and the checking itself is being automated one layer up. How many, and at what error rate, is the number every profession will have to measure for itself, and the checkers learned to check by producing, which is the point the hiring numbers below turn on. The second objection is that cheap output creates its own demand, so total spending on cognitive work may rise even as the price per unit falls, the way spreadsheets created more accountants, even as the bookkeeping clerks went. That is also true, and it is the best hope for wages. But it cuts the other way for prices. In the trust-gated sectors, price falls far more slowly than cost, because the licence and the liability haven't got cheaper. Tyler Cowen makes the related point that these are the sectors slowest to adopt anything. He is right, and slow adoption is exactly what keeps the price sticky while the cost falls. That is the margin in the opening paragraph, and the strongest form of this essay's claim: the money doesn't disappear from these sectors. It moves within them. It doesn't move on its own, though, and one worked example shows where it can go.

One service, three endings#

Take one service. A small law firm reviews a standard commercial lease for a small business; the numbers are illustrative, not measured. Before the machine, an associate spends three hours on it at $100 an hour, a partner checks it in half an hour at $300, the mistakes that get through cost $50 a lease on average, and the client pays $1,000. Now a model drafts it for a few dollars plus half an hour of the associate's time, and there are three ways it can end: the owners keep the saving, the clients get it, or checking eats it.

In the owners' ending the price holds, the firm keeps the saving, and its margin rises by half. In the clients' ending a rival signs the same review for $450 and the price follows; small businesses that never paid for a review start to, the firm handles two and a half times as many for about the same total margin, and the clients keep the rest. In the checking ending the partner can't trust a draft she didn't watch being made, checking takes her 75 minutes, the machine's new kinds of mistake slip through a little more often, and the saving is gone: the associate's hours have become the partner's.

Three questions decide which. Can the client tell a good review from a bad one? If not, they pay for the name, the price sticks, and the owners win. Can a rival sign the same work? If licences are plentiful and clients compare, the price falls toward the new cost, and if demand grows enough, total spending on the service rises anyway, as it did for accountants. Is checking a draft cheaper than making one? Only if the checker can find its errors, which she learned to do by making drafts. In the trust-gated professions the first answer is usually no and licences slow the second, so I expect the first ending more often than the second, at least at first. That is a forecast; the example only shows what decides it.

Why it doesn't show up yet#

And yet the aggregate numbers haven't moved. Productivity growth across the rich world is still crawling. A survey of nearly six thousand chief executives and finance chiefs in early 2026 found that nine in ten saw no measurable productivity gain from AI, while task-level studies find gains of 15 to 50 percent and firm-level studies find almost nothing; Daron Acemoglu's estimate for the whole economy is well under one percent of productivity over a decade. Economists have dusted off Solow's line from 1987: you can see the computer age everywhere except in the productivity statistics.

I think the paradox resolves the way it did last time. Electricity took forty years to show up in factory productivity, because factories had been built around a single steam shaft and had to be rebuilt around distributed motors. The gains from AI are stuck in the same place, inside firms built around human composition, waiting for the firm to be rebuilt. The task-level gains are real; the firm-level gains need new firms. That is why the strongest early signal isn't in productivity at all but in hiring. By mid-2026, on payroll data analysed at Stanford, employment among 22 to 25 year-olds in the most AI-exposed occupations was 19 percent behind where it would have been had it kept pace with less-exposed work, and the gap was still widening. The Stanford team is careful to call that a pattern rather than proof of cause. It is the pattern you would expect if firms were hiring fewer juniors into the work models now do.

Where the value goes#

When an input gets cheap, value doesn't vanish; it moves to the next bottleneck, and there are four.

Compute. Chips, the buildings that hold them, and above all the electricity. The largest cloud companies plan to spend around $725 billion in 2026. Power is the binding constraint, and one major cloud has reported tens of billions in orders it cannot fulfil until new data centres have the power to run. Whoever controls compute and energy sits where the value lands first.

Trust. When anyone can produce a plausible contract, diagnosis or audit, the scarce thing is someone accountable for it: a licence, a signature, an insurer, a name that can be sued. The trust-gated rows on the series' timeline of professions are the ones where the price of composition collapses and the price of accountability rises, and whole professions will reorganise around this, with fewer people doing the work and more value per signature.

Presence. Bodies in unpredictable places: plumbers, nurses, the person in the room with you. This work holds its price while the cognitive work around it gets cheap, and I expect its wages to rise as the supply of people willing to do it lags demand. That is the trades' decade, and it is a bet.

Distribution. When production is cheap, owning the customer is everything. A brand, a client list, a platform with the relationship already in place. Cheap output floods every channel with plausible content, and the value of a trusted channel rises in proportion. This is why the internet's winners are AI's winners too, and why every lab wants a consumer product.

Two of the four are capital. Compute is bought, and distribution is mostly bought or inherited: the channel, the brand, the platform. They favour whoever is already large, and they are where concentration comes from: the future in which a few firms and governments end up holding the technology. The other two are positions. Trust and presence can be built by one person without capital: a licence, a reputation, a body in a room. That split says who the value is moving toward. Owners first, then the people who are answerable and the people who are there. What is not on the list is the output itself, and the people who used to make it. That omission is the labour market's next decade.

Labour, in four splits#

The middle compresses. Cognitive work sorts into accountability at the top, which gets more valuable, and production in the middle, which gets cheap. The senior lawyer who decides what to fight over, and who carries the liability for the choice, keeps her value; the associate who drafts does not. The gap between them used to be a career ladder and is becoming a cliff.

The bottom rung goes first, and it compounds. Firms aren't firing seniors. They're not hiring juniors. But juniors are how seniors get made, and Matt Beane has shown how that apprenticeship erodes when machines let experts work without novices. Checking is learned the same way: a senior knows where a draft goes wrong because she wrote a thousand of them. So a profession that stops hiring 25-year-olds in 2026 has no 35-year-old experts in 2036, and no one left who can check the machine. It is a commons problem: every firm would rather hire the seniors someone else trained, so they all cut juniors at once, and the bill arrives in a decade, when no single firm pays it.

Status goes before income. The first thing a professional loses isn't pay. It is being the one in the room who could do the thing. A lawyer whose drafts were the best in the firm is now the person who checks a machine's, and nobody prices that. It arrives years before the payroll number moves, which is why the loudest resistance to these tools comes from the people whose standing they touch rather than the people whose jobs they take.

Capital's share rises. In the exposed sectors, the value that used to go to wages goes to whoever owns the compute, the distribution or the signature. I expect the labour share of income in cognitive services to fall the way it fell in manufacturing, and for the same reason: the work moved to machines that someone owns. Anton Korinek and Donghyun Suh reach the same place by a different road: in their scenarios, wages collapse under full automation unless something irreproducible stays scarce, and Erik Brynjolfsson's Turing trap adds that machines built to substitute for people rather than extend them take their bargaining power along with their wages. New demand can offset that, as it does in the lease example when a lower price brings in two and a half times the work. The direction is my forecast, and it is the mechanism behind the concentration scenario in the third essay.

The build-out and its fragility#

The compute bottleneck is being attacked with the largest private capital programme in history, in dollars, and it has the shape of every infrastructure boom before it. Railways in the 1840s, electricity in the 1900s, fibre in the 1990s: enormous capital raised on the promise of transformation, most of the builders bankrupt within a decade, the infrastructure surviving and then delivering the transformation on the users' schedule rather than the investors'.

Two things make this cycle more fragile than it looks. First, a lot of the money is circular: chip makers invest in labs that buy chips, clouds invest in labs that rent clouds, and the same dollar gets counted as demand three times. Second, the spend is running ahead of measured returns by a margin that only makes sense if compression is the true scenario. If capability stalls, the future the third essay calls the wall, this is the fibre bust, and the losers are whoever financed it. If it keeps climbing and the gains spread, the one it calls compression, it's the electrical grid, and the winners are whoever holds the assets after the builders are gone. The assets survive either way. The equity doesn't have to. So the value still lands with whoever owns the compute; it is just not always the people who paid to build it.

The firm shrinks and the winners grow#

Ronald Coase said the boundary of a firm sits where the cost of coordinating inside it meets the cost of contracting outside it. Agents cut both, and the direction depends on which they cut more. My bet is on the outside, because specifying and contracting are exactly what agents are good at, so the firm shrinks: fewer layers, fewer managers, and a great many very small companies doing what used to need a department. The opposite is possible, and if it happens, the concentration scenario arrives faster. At the same time platform economics get more extreme, because distribution and compute both have returns to scale. So both things happen at once: a long tail of tiny firms, a handful of enormous ones, and a hollowed-out middle of mid-sized professional services firms, regional agencies and departments. That is the same shape as the labour market, one level up.

The tax base#

Rich countries fund themselves mostly by taxing labour: payroll, income tax, and the consumption that wages support. If value moves from wages to capital across the quarter to a third of the economy that is cognitive and professional services, and the capital sits in a handful of firms that can choose where to book it, the tax base moves too, and it moves somewhere harder to tax. Social insurance has the same problem in a different form. It was built for cyclical unemployment, people who lose a job and find another, not for a rung of the ladder disappearing. Neither system has a plan, and both are about to be tested by the trust-gated professions, which are exactly the ones that pay the most tax.

Abundance, and what it leaves out#

The people building this describe the same decade as abundance, and they are not wrong about output. That is a different abundance from Ezra Klein and Derek Thompson's, which is about building more of what is scarce, and on the grid this essay agrees with them. Sam Altman expects the price of many goods to fall dramatically and the cost of intelligence to converge on the cost of electricity; Dario Amodei's Machines of Loving Grace has the cures and a century of progress in a decade. Read closely, both have already conceded the question this essay is about. Altman wrote that the balance of power between capital and labour "could easily get messed up", and floated a compute budget for everyone on Earth. Amodei's The Adolescence of Technology, in January, warns of an unemployed or very-low-wage underclass and proposes progressive taxation, aimed at AI companies if need be. So the disagreement is not about whether the pie grows. It is about three things the abundance essays leave to later.

The first is what gets cheap. Cheap composition lowers the price of everything made of composition and raises the relative price of everything that isn't: the room, the licence, the person who is there, the land under the data centre. Baumol's mechanism doesn't stop; it changes sides. Pedro Santa-Clara has argued that this is exactly how abundance reaches labour, because the human tasks that remain get dearer. That is true for the people who hold those tasks, and the split above says who they are: the answerable and the present, with owners ahead of both. For the person whose work was the composition, a dearer nurse and a cheaper draft is not abundance. It is a bill.

The second is order. Status goes before income, and the bottom rung goes before either, so the losses arrive years before any transfer does, and no transfer restores a rung. The third is timing. Taxing an enormous pie assumes the pie is still somewhere a tax can reach, and the section above says which way the base is moving; the second essay argues that the window for regulating the frontier at all is measured in months. Abundance is a claim about how much there will be. This series is about how it gets divided, and the division gets decided first.

What to do#

This is a map of where value is moving, not investment advice, which I'm not qualified to give, and your situation is yours.

If you work for a living or run a business, the moves are in the fifth essay. In short: know which part of the work is composition, move toward the signature, own something, because in the exposed sectors the value is leaving wages, and treat the junior pipeline as your problem, because in ten years it will be.

If you allocate capital. Value moves to the four bottlenecks, and the history of infrastructure booms says the assets outlast the builders. Be more suspicious of the circular money than of the technology. The biggest beneficiaries of electricity weren't the utilities but the factories that rebuilt around it, and the same is probably true here: the sectors that flip, healthcare, education, law, government services, are where the productivity is, and most of them aren't listed companies. The trades' decade is a real theme.

Questions to ask your leaders. These are the ones I'd put to a minister, a chief executive or a board.

• What is the plan for the tax base when labour income falls as a share of the economy?
• Who owns the compute this country runs on, and what happens if the owner is foreign?
• What replaces the apprenticeship in the professions that have stopped hiring juniors?
• Is our social insurance designed for a rung disappearing, or only for a downturn?
• Where is the grid capacity for the data centres you've announced, and who pays for it?
• To a chief executive: what share of your cost base is composition, and what is the plan for the people doing it?
• To anyone: are you one of the people deciding this? If not, who is, and can you see what they decide?

What I don't know#

• Whether the flip is fast or slow. The productivity paradox could last a decade, as it did for electricity, or two years.
• Whether capital's share really rises, or whether cheap composition creates enough new demand that wages hold. It did for accountants when spreadsheets arrived. It didn't for switchboard operators.
• Whether the build-out is 1900 or 1999. The assets survive either way; when they pay is the whole question for anyone holding them.
• Whether governments move before the tax base does. History says no.

The same sentence, changed subject#

The first essay said the singularity for an individual is the day the thing you're best at becomes cheap. The economic version is the same sentence with the subject changed: the singularity for a sector is the day its composition becomes cheap, and the value it held moves to whatever is still scarce. That day has a date for every sector, and for most of them it's inside this decade. The only real question is who's standing where the value lands.

Think I'm wrong? Tell me where. Follow on X · LinkedIn · RSS