A beautiful world, and nobody owns it (September 2026)

Assume the AI works, and the price of thinking falls to something near the price of electricity. Economists have already worked out what that does to output and wages, and I’ll hand you their paper. What they set aside in a single sentence is the question I actually care about: who ends up with the money? The answer is that it is not the people selling the intelligence, and the reason is the same reason the price fell in the first place.

Start from the optimistic premise, because the pessimistic one is boring. Suppose AI really does automate cognitive work. That is an enormous amount of new value entering the world. The interesting question was never whether it arrives. It’s where it lands.

I worked at that for a while from a production function, and then found that Konrad Kording and Ioana Marinescu had done it properly a year earlier, in a Brookings working paper I should have read first. Their model is better than mine in every technical respect — a nested CES with capital in both sectors, endogenous labor allocation, comparative statics, simulations, and a falsifiable empirical condition. They beat me by a year, so I can’t claim ignorance, and the honest move is to hand them the first half of this post and spend my own words on the half they explicitly set aside.

So: the production function, briefly and with attribution. Then the part nobody priced.

The part that is already settled

Write output as a CES aggregate of two inputs: $T$, the quantity of cognitive work done, and $L$, the quantity of physical work done. These are amounts of work, not amounts of money — that distinction is the whole trick below. Their prices are $p_T$ and $p_L$, so what the economy spends on each is $E_T = p_T T$ and $E_L = p_L L$.

\[ Y = \left[\beta\,T^{\rho} + \alpha\,L^{\rho}\right]^{1/\rho}, \qquad \sigma = \frac{1}{1-\rho}, \qquad \rho = \frac{\sigma - 1}{\sigma}. \]

Here $\beta$ and $\alpha$ are fixed weights on the two inputs, and $\sigma$ is the elasticity of substitution — how willing the economy is to get a job done with thinking instead of with hands. Everything below hangs off whether it sits above or below one.

Now run the experiment in three steps, because the interesting part is that two of them pull against each other.

One: the price falls, so the quantity rises. Cheap AI means $p_T$ drops, and firms respond by buying more cognition. As $p_T \to 0$ the quantity $T$ grows without bound. Physical work has no such room — hire all you like and $L$ runs into a ceiling made of people, concrete and time. So $T/L \to \infty$.

Two: quantity is not revenue. The AI sector’s income is $E_T$, and $E_T = p_T T$ is a falling price times a rising quantity. Which factor wins is not obvious, and it is the entire question.

Three: the ratio settles it. Setting each input’s price equal to its marginal product and dividing one condition by the other,

\[ \frac{E_T}{E_L} \;=\; \frac{p_T T}{p_L L} \;=\; \frac{\beta}{\alpha}\left(\frac{T}{L}\right)^{\rho}. \]

Since $T/L \to \infty$, the sign of the exponent $\rho$ decides everything — and $\rho$ is positive exactly when $\sigma > 1$:

\[ \frac{E_T}{E_L} \;\longrightarrow\; \begin{cases} 0 & \text{if } \sigma < 1 \quad (\rho < 0), \\[4pt] \infty & \text{if } \sigma > 1 \quad (\rho > 0). \end{cases} \]

Zero or everything, with nothing in between, and the switch sits exactly at $\sigma = 1$. The derivation, along with the rest of the arithmetic behind this post, is in the mathematical appendix.

Read the first case carefully, because the intuition runs backwards. If cognition and physical work are complements, then making cognition free drives the revenue of the people selling cognition to zero — not to a slower growth rate, to zero — while the economy around them gets much larger. Cheap AI in a complements world doesn’t make the AI sector big. It makes the AI sector vanish and everything it touches enormous.

Kording and Marinescu call the underlying fact intelligence saturation, and they make the case for it far better than I can. Their evidence is the interesting part: returns to IQ flatten at the top of the distribution, research productivity per researcher has been declining for decades, the ICT revolution delivered enormous cost declines and modest growth, and measured LLM productivity effects so far are small. You build things through physical means; intelligence can at best make you maximally efficient at doing so. The recipe does not cook the meal.

Everything else in their paper follows, and it is worth reading in the original rather than my summary:

In the complements case ($\sigma < 1$), the physical world eats the boom. Output stalls against concrete, permits, trucks and time. This is Baumol’s cost disease wearing a new hat — the sector with slow productivity growth absorbs a rising share of spending, which is why a string quartet costs more every decade. Aghion, Jones and Jones got there first for AI specifically: what limits you is not what you do well, but what is essential and stubbornly hard to improve.
Cheap cognition pushes labor into the physical sector, because intelligence tasks automate first. Kording and Marinescu model the reallocation endogenously and give the condition under which it happens.
The wage effect is genuinely ambiguous, and this is where their treatment is much better than anything I had. Two forces fight: a scale effect (automation raises output, which raises wages) and a reallocation effect (more workers crowd into physical work, which lowers them). Their baseline simulation has wages rising and then falling as automation deepens, and they extract a testable necessary condition — wages cannot fall unless the intelligence sector’s employment share falls first. If you want a number for what happens to the electrician, use theirs, not mine.
Substitutability makes it worse. Wage declines steepen at high automation when the two sectors’ outputs are more substitutable, which is the $\sigma > 1$ case: output explodes, demand for cognition turns elastic (the Jevons paradox — cheaper resources raise total spending on them), and human wages flatten against a capable substitute.

Two things I’d add to that, one small and one that turns out to matter.

The small one: $\sigma$ isn’t a property of an economy, it’s a property of a kind of production, and the two cases above are running simultaneously in different halves of the same economy. Framing a house, staffing an emergency room, installing a heat pump, looking after somebody’s mother — cognition helps and does not replace, $\sigma < 1$. Generating a film, a game asset, an illustration, a news summary, a block of boilerplate — cognition is most of the job, $\sigma > 1$. (Kording and Marinescu’s own framing anticipates this: they note you could relabel their two sectors as the “human domain” and the “machine domain” and the model survives.)

The one that matters is what none of it prices.

What the production function doesn’t price

Notice the shape of that whole argument. It takes the falling price of intelligence as given — in their paper, $K_I$ and the automatable share $\alpha_I$ are exogenous parameters, set to illustrate a scenario — and works out the consequences for output and wages.

That is the right way to write that paper. It also means the entire result is driven by a number nobody in it derives. Why does the price of intelligence collapse? What stops the people selling it from holding the line? And when it does collapse, where does the surplus actually come to rest?

Their limitations section names this in a single sentence: “a lack of competition among AI providers may limit AI development and deployment, or lead to strategic choices that diverge from socially optimal development paths.” That’s the whole treatment of market structure, filed under things the model doesn’t do.

I think that sentence has it backwards, and the rest of this post is the argument. The risk isn’t too little competition among AI providers. It’s that there is so much — structurally, unavoidably, for reasons that have nothing to do with antitrust — that nobody selling intelligence can hold a margin at all. Their exogenous falling price is the output of a market-structure argument, and once you run that argument you get the answer to the who-gets-the- money question their model was never built to ask.

The fringe sets the ceiling

There isn’t one kind of intelligence on sale. There’s closed intelligence and open-weight intelligence, and the open kind drags the price of the whole category toward the cost of compute and electricity.

Give the closed model a genuine productivity advantage $\lambda > 1$: it does $\lambda$ units of useful work per unit the open model does. Then a rational buyer pays a premium up to — and not one cent past — that advantage:

\[ p_{\text{closed}} \le \lambda \cdot p_{\text{open}} \approx \lambda \cdot c_{\text{compute}}. \]

That’s limit pricing, a result from the 1940s, and it’s brutal here because $\lambda$ is not a moat you built. It’s a lead you’re defending against a pack that redistributes your homework every twelve to eighteen months. The open fringe doesn’t have to beat you. It only has to set your ceiling.

Four ways the margin dies anyway

Grant the labs a real and durable $\lambda$. They still have to satisfy a profit equation, and four structural forces push it to zero.

Inference bloat. Staying smarter than a heavily optimized open model tends to mean more parameters — enormous mixtures of experts spanning multiple nodes. If serving cost scales worse than the intelligence advantage it buys, the firm is structurally insolvent: it cannot reach the limit price without losing money on every query. The premium is real and the unit economics still don’t close.

The Red Queen’s amortization window. Open weights trail by roughly 12–18 months. To keep $\lambda$ from decaying to one, the labs must train the next frontier model, with fixed costs marching from $1B to $10B to $100B. That buys an 18-month window in which to recover a $100B training run at thin margins — and then do it again, larger. You can win every generation of the race and go bankrupt on schedule.

The compute tax. Intelligence is, at bottom, a software wrapper around silicon and energy. If the upstream layers are concentrated — foundries, accelerator designers, the grid — they can see the same inelastic demand you do, and price into it. Every dollar of surplus the labs generate is a dollar the suppliers can reach. Joel Spolsky’s old advice was to commoditize your complement; the labs’ complement is a fab with a two-year lead time and no substitute, and the commoditizing is running in the other direction.

The router. Nobody sends every task to the most expensive model. They build an orchestration layer that routes the easy 95% — classification, extraction, summarization, boilerplate — to a local open-weight model at bare-metal cost, and calls the expensive API only for the residual. That starves the frontier of exactly the high-volume, low-thought traffic it needs to amortize the fleet.

One clarification before the objections, because it’s the difference between an argument and a slogan. “Profit goes to zero” in the competitive sense means economic profit — the excess over the cost of capital. It does not by itself mean the capital is lost; a firm at zero economic profit has recovered its investment at a perfectly ordinary rate of return. What makes this case sharper is that marginal cost sits nowhere near average cost. Serving a token is cheap; the training run that made the token possible was not. When those two numbers differ by orders of magnitude, “price falls to marginal cost” and “the fixed costs come back” stop being jointly satisfiable, and the industry has to either not build at that scale or eat the loss.

Reliability is the best objection, and retries answer it

The router draws the sharpest pushback, and it deserves full strength. João Sedoc led with it: reliability compounds. Chain a per-step success rate across a fifty-step agentic workflow and

\[ 0.95^{50} \approx 7.7\%, \qquad 0.99^{50} \approx 60.5\%. \]

Four points of per-step reliability is an eightfold improvement in end-to-end success — the difference between a demo and a deployment. If only the frontier model clears that threshold, willingness to pay for it is discontinuous rather than marginal, the router sends everything to the expensive endpoint after all, and $\lambda$ is worth far more than the limit price allows.

The objection is right about the value and wrong about who has to pay for it.

Retries collapse the cliff, and retries are cheap. If you can detect a failed step, one retry takes a 95% step to 99.75%, and $0.9975^{50} \approx 88\%$ — comfortably better than the 60% you get from a perfect single-shot 99% model. In coding agents you very often can detect it: compilers, type checkers, linters and test suites are failure oracles you already own.

The extreme version is the one I’ve been living in. On the Lean project the oracle isn’t a linter, it’s a kernel: a proof either type-checks or it doesn’t, no false positives and no judgment call. Against an oracle that good, per-attempt reliability nearly stops mattering — six failed tactic rounds cost minutes of wall clock, and the attempt that finally compiles is exactly as correct as a first-try success would have been. That is an unreliable model emitting verified output, bought with cheap inference against a perfect checker rather than with a better frontier model. l3m, the coding agent built on that machinery, is the same trick turned on the agent’s own sandbox.

Two smaller cracks in the arithmetic. Independence is doing enormous work: the $p^{50}$ model treats step failures as independent draws, and they aren’t — the same model meeting the same ambiguous spec with the same blind spots fails in correlated clusters. And a threshold doesn’t repeal the ceiling, it paints a staircase on it: cross 99% and the workflows needing 99% light up, then the next tranche wants 99.9% for another order of magnitude of compute.

The sting is in what turns out to be scarce instead. Once the retry loop is cheap, the bottleneck moves off proving and onto stating: saying precisely what you want becomes the expensive step, and that is a human design cost no quantity of compute relieves. Reliability gets bought with scaffolding and specification — and neither of those is something a frontier lab can sell you exclusively.

Both halves, one ending

Lay the two halves side by side and the symmetry is the finding.

	$\sigma \le 1$ (complements)	$\sigma > 1$ (substitutes)
Output	stalls at 2–3×	explodes to 6×+
AI’s income share	falls	rises
Who’s scarce	hands, concrete, permits	silicon, energy, fabs
Wages	rise, then possibly fall	flat, offset by deflation
How the labs lose	no volume	no margin

The left column is Kording and Marinescu’s. The right column’s top half is theirs too. The bottom row is mine, and it is the row that does not vary: where cognition complements hands, the labs lose on volume, because demand saturates against a physical wall; where it substitutes, they lose on margin, because the fringe sets the ceiling and the suppliers take the difference.

The two halves disagree about almost everything — how rich we get, who works, what a wage means — and agree completely about the one thing the investment case needs them to disagree about.

That’s an uncomfortably robust conclusion, so it deserves the strongest objection I can find.

Follow the money

Everything so far says the labs don’t capture the surplus. That leaves the question the title is making a promise about: does anyone? It would be a thin sort of abundance if the value simply evaporated.

It doesn’t, and the reason is an accounting identity rather than an argument. Output is income. The value of what gets produced equals the sum of the wages, profits, rents and interest paid out to produce it — that is how the quantity is constructed, not a claim about it. Six times the output means six times the income, and somebody has it.

Which is why “the firms will spend it and it circulates back to people” is true and yet does no work. Circulation tells you the money doesn’t vanish. It says nothing about whose income it becomes: a dollar of profit spent on more compute is income to a foundry. The distribution question is settled by factor shares and by nothing else, and here the model has something exact to say.

Labor’s share of income is constant if and only if $\sigma = 1$. That is the special property of Cobb-Douglas, and it is why the intuition that a richer world must be a richer workforce feels so solid — in a Cobb-Douglas world it is simply true, automatically, at any level of output. Off that knife edge the share moves, in the direction we already computed: as cognition gets cheap, labor’s share rises when $\sigma < 1$ and falls when $\sigma > 1$.

So the complements half — the disappointing one for output — is the good one for workers.

Two doors, both leading to the household

Run each half of the economy through its own regime — physical at $\sigma < 1$, digital at $\sigma > 1$ — and the household wins twice, through two different doors.

In the physical half, people win as workers. $\sigma < 1$ means labor’s share rises as cognition cheapens, and the wage is $\alpha Q/L$. The bottleneck factor captures the rent, and the bottleneck is hands.

In the digital half, people win as consumers. $\sigma > 1$ means labor’s share falls, so this is not a wage story. It is the story of the four horsemen above: marginal cost near zero, an open fringe setting the price, producer surplus collapsing, and the gain landing as consumer surplus instead.

The word for the second door is appropriability: not how much value you create, but what fraction of it you manage to charge for. It is low in the digital half for two independent reasons, and only one of them is about competition.

The structural one needs no competitor at all. Sell somebody a video game for $70 and they play it for two hundred hours. You created an enormous amount of value and charged for a sliver of it, and to do better you would have to know each buyer’s willingness to pay and be permitted to charge exactly that. Uniform pricing on a good that costs nothing to copy leaves most of the surplus on the consumer’s side of the line, and this is true of an unchallenged monopolist. It is why Nordhaus, measuring how much of the social return to innovation innovators actually keep, came out at about 2.2%.

The competitive one is the entire first half of this post: an open fringe pushing price toward marginal cost, which shrinks even the sliver.

So the claim is not that digital goods resist private ownership. Steam, copyright and the App Store appropriate extremely well, and the firms running them are enormous. It is that “enormous” and “most of the value created” are wildly different quantities, and the gap between them is where everyone else lives.

And then the part that makes this more than a tidy symmetry. Baumol says the sector with the slower productivity growth absorbs a rising share of total spending. That sector is the physical one. So the half of the economy where labor’s share is rising is also the half whose share of the economy is rising. The composition effect and the within-sector effect point the same direction, and labor’s share of total income is helped twice over.

Which means the 6× is an undercount

There is a measurement problem sitting underneath all of this, and it cuts in the optimistic direction.

$Q$ in this model is output at market prices, and so is GDP in the national accounts: the rectangle $p \times q$, not the whole area under the demand curve. Consumer surplus is excluded by construction — it is the part nobody was charged for, so there is no transaction to record. For wheat that exclusion is a rounding error, because price sits close to value. For the digital half it is most of the story. Wikipedia’s contribution to measured GDP is approximately zero.

So the “6× output” in the substitutes half is an undercount of the welfare gain, and the size of the undercount is governed by exactly the appropriability we were just discussing. Low appropriability means price sits far below value, which means a large consumer surplus, which means a large gap between what people got and what the statisticians recorded.

Notice that this is the same fact as the labs’ profit problem, seen from the other side. Price near marginal cost is simultaneously why producers capture little and why the accounts record little. There is one phenomenon here, not two: the failure to monetise and the failure to measure are the identical gap, read once by a CFO and once by a statistical agency.

Which has a consequence worth stating plainly, because it cuts against the natural reading of the labor-share literature. As appropriability falls, the labor share becomes a worse proxy for how people are doing. Someone watching labor’s share of GDP decline and concluding that workers are losing may be carefully measuring the half of the economy that is shrinking, while the other half quietly delivers unpriced gains to the same people in their capacity as consumers. That is not an argument for ignoring the labor share. It is an argument that in a low-appropriability economy it stops being sufficient.

One honest scope note while we’re here. A CES production function is a better description of the physical half than the digital one anyway. Near-zero marginal cost with large fixed costs means increasing returns, and under increasing returns marginal-cost pricing does not cover the fixed costs, so the competitive machinery that makes factor shares meaningful starts to creak. That is another reason to run the two halves on separate models rather than averaging them into one $\sigma$: use the production function where the production function works, and the surplus-and-appropriability framing where it doesn’t.

Does it survive contact with arithmetic

Enough of it does. Labor income is $s_L \cdot Q$ — a share times a level — so the question is only whether the share falls faster than output grows. Starting from a labor share near 0.6 with output at 6×, workers are absolutely worse off only if the share falls below 0.1, a more-than-sixfold collapse. Halve the share instead and labor income still triples. Output growth multiplies against share decline, and 6× is a great deal of headroom.

So no, I don’t think the optimism is wishful. But it is a bet on the factor share rather than a consequence of the GDP number, and three things keep it honest.

The bounding case is real and it has a name. Leontief’s horses were a factor of production through an era of spectacular output growth. The horse population collapsed anyway, because their marginal product fell below their maintenance cost. “The economy got much richer” and “this factor did well” are independent propositions. $\sigma > 1$ everywhere, permanently, is the horse scenario, and the only reason it isn’t this post’s forecast is the sectoral split above.

Consumer surplus is not a paycheck. The digital half arrives as things being cheap rather than as income. Distributionally that is good news, because consumption is spread far more evenly than income is. It is still cold comfort to the illustrator, whose wage was the concentrated thing and whose compensation is now a diffuse benefit shared with everybody. Aggregate improvement is entirely compatible with brutal individual loss, and the losses land precisely on the people the $\sigma > 1$ sector used to pay.

The boundary moves. $\sigma$ is a technology, not a constant of nature. Every advance in robotics pushes another slice of the physical half across the line into the substitutable column. That is the transition problem from the next section wearing different clothes, and it is the thing I would actually worry about.

The objection worth taking seriously

$\sigma$ is not a constant of nature. It’s a technology, and it drifts upward as robotics improves. So the honest path isn’t complements or substitutes — it’s complements, then a transition, then substitutes. And transitions are exactly where fortunes get made. During the crossover, whoever supplies the newly-substitutable cognition is selling into demand that is becoming elastic while competitors haven’t arrived yet. The two-regime table is too clean; the money is in the smear between the columns.

Worse for me: where cognition substitutes, the capital share of income rises, and the labs are capital owners. Shouldn’t they collect?

The answer is that not all capital collects, and the distinction is durability. Rents accrue to factors that are scarce, non-reproducible and long-lived. Model weights are none of those. A frontier checkpoint is a depreciating asset with something like an 18-month economic half-life, reproducible by anyone with the recipe and the compute, and getting cheaper to reproduce every year — that is the Red Queen trap restated as a balance sheet. A fabrication plant is a thirty-year asset with a two-year lead time and four plausible operators on Earth. A gigawatt of firm power is worse. When the capital share rises, it rises toward the durable capital, and the weights are the most perishable capital in the stack.

So the transition is real, and I’d expect it to be lucrative — the way the fiber buildout was lucrative for about four years. It just isn’t a franchise.

The bibliography, both sides

Grouped by which brick each supplies. The economics here is all borrowed; the arrangement is what I’d defend.

The production function and the bottleneck.

Kording & Marinescu, “(Artificial) Intelligence Saturation and the Future of Work” (Brookings Center on Regulation and Markets working paper, November 2025). The paper this post should be read against, and which does the first half of it properly: a nested CES with physical and intelligence sectors, each with its own capital and a freely-flowing labor supply; complementarity between the sectors producing intelligence saturation; and automation of intelligence tasks pushing workers into physical work. Their wage result is the one to quote — the effect is ambiguous, decomposing into a scale effect and a reallocation effect, non-monotonic in their baseline (wages rise, then fall), with a falsifiable necessary condition for wages to fall at all. Ours differs by: endogenising the one thing they hold exogenous. The falling price of intelligence is a parameter in their model and the conclusion of an argument in this post, and the argument turns out to answer a question their framework doesn’t pose — not what happens to output and wages, but who ends up holding the surplus. Where they list “a lack of competition among AI providers” as a factor that might moderate the impact, I argue the opposite sign: competition among providers is structurally unavoidable and total, which is why their exogenous price falls and why nobody selling intelligence keeps the gains. Their result and this one compose; they do not compete. (link)
Baumol, “Macroeconomics of Unbalanced Growth” (AER, 1967). The cost disease: when sectors have unequal productivity growth, the slow sector absorbs a rising share of spending. Ours differs by: casting the physical world as the slow sector and the AI boom as the fast one, which turns a story about symphony tickets into a story about who captures an automation windfall.
Aghion, Jones & Jones, “Artificial Intelligence and Economic Growth” (NBER w23928, 2017). The closest prior work to the complements case, and it got there first: growth is limited “not by what we do well but rather by what is essential and yet hard to improve.” Ours differs by: carrying the argument past growth into factor shares, and asking what the bottleneck implies for the revenue of the firms selling the cheap input. (link)
Acemoglu & Restrepo, “The Race Between Man and Machine” (AER, 2018). Task-based framework: automation displaces labor, reinstatement of new tasks restores it, and the balance sets factor shares. Ours differs by: using a blunter CES two-factor model on purpose, because the question here is the sign of one derivative rather than the path of employment. (link)
Acemoglu, “The Simple Macroeconomics of AI” (NBER w32487, 2024). TFP up ~0.55% over a decade, an order of magnitude below the consultancy forecasts — an argument that total surplus is modest. Ours differs by: taking no position on the size of the pie and arguing about its division; if Acemoglu is right the labs do worse still, so the arguments compound. (link)
Moravec, Mind Children (1988). Moravec’s paradox: the hard problems are easy and the easy problems are hard, with sensorimotor competence the stubborn one. Ours differs by: pricing the paradox. If Moravec holds, $\sigma \le 1$ persists, and the paradox stops being a curiosity about robots and becomes the reason the AI sector’s revenue saturates.

Factor shares: where the money actually lands.

Karabarbounis & Neiman, “The Global Decline of the Labor Share” (QJE, 2014). Documents a ~5 percentage point fall in the global labor share since 1980, attributes it to the falling relative price of capital goods, and estimates $\sigma \approx 1.25 > 1$. This is the substitution case already measured, on the previous technology: a cheaper substitutable input, an elasticity above one, and labor’s share giving way. Ours differs by: arguing the elasticity is sectoral rather than economy-wide, so the same mechanism that lowers labor’s share in the digital half raises it in the physical half. (link)
Kaldor’s stylized facts (1957, 1961). The empirical regularity that factor shares are roughly constant over long horizons — which is exactly the Cobb-Douglas, $\sigma = 1$ case, and exactly why the intuition “richer economy, richer workers” feels like arithmetic rather than a bet. Ours differs by: treating constancy as the knife edge it is, and asking what happens on either side of it.
Autor, Dorn, Katz, Patterson & Van Reenen, “The Fall of the Labor Share and the Rise of Superstar Firms” (QJE, 2020). The competing explanation: the labor share fell because activity reallocated toward low-labor-share superstar firms, not because of an elasticity above one. Ours differs by: needing the elasticity story to be at least partly right. If concentration rather than substitution is doing the work, the post’s substitution mechanism weakens and its antitrust implications get much stronger. (link)
Leontief, on horses (1983, and Machines and Man, 1952). Horses were a factor of production across an era of enormous output growth, and their population collapsed because their marginal product fell below their maintenance cost. The canonical demonstration that a rising tide does not lift a factor that is being substituted rather than complemented. Ours differs by: treating it as the bounding case rather than the forecast — the horse outcome requires $\sigma > 1$ everywhere, and the argument here is that half the economy is on the other side of one.

Who captures the surplus of a general-purpose technology.

Nordhaus, “Schumpeterian Profits in the American Economy” (NBER w10433, 2004). Measures the share of social returns to innovation captured by innovators and finds it around 2.2%. This is the single most load-bearing empirical number behind this post. Ours differs by: supplying the mechanism — limit pricing against an open fringe plus upstream supplier power — rather than the measurement. (link)
Brynjolfsson, Collis, Diewert, Eggers & Fox, “GDP-B: Accounting for the Value of New and Free Goods in the Digital Economy” (NBER w25695, 2019). Proposes a measure that counts the consumer surplus from free and new goods, on the grounds that GDP records the transaction and therefore misses the entire welfare gain from anything priced near zero. Ours differs by: using the same gap as an argument about market structure rather than about statistics. Their unmeasured surplus and this post’s unappropriated surplus are one quantity, which is why a sector can simultaneously transform daily life and fail to show up in either GDP or anybody’s earnings. (link)
Teece, “Profiting from Technological Innovation” (Research Policy, 1986). Where the word appropriability comes from, and the framework this post keeps reinventing: whether an innovator captures the value depends on the appropriability regime and on who owns the specialised complementary assets. When the regime is weak and the complements are held by somebody else, the innovator loses and the complement-owner collects. Ours differs by: not at all, really — the labs are Teece’s textbook loser. Weak appropriability from open weights, complementary assets consisting of leading-edge fabrication and firm power, held by others. The compute tax above is his result with a 2026 subject.
Bresnahan & Trajtenberg, “General Purpose Technologies: ‘Engines of Growth’?” (J. Econometrics, 1995). Defines the GPT pattern: pervasive, improving, spawning complementary innovation, with value realized downstream. Ours differs by: asking specifically where in the stack the rent settles when the GPT’s inputs are more concentrated than the GPT itself.
Spolsky, “Strategy Letter V” (2002). Smart companies commoditize their complements. Ours differs by: running it backwards. The labs’ complement is leading-edge fabrication and firm power, which are the least commoditizable things in the economy, so the strategy is being executed on them. (link)

Limit pricing and the fringe.

Bain, “A Note on Pricing in Monopoly and Oligopoly” (AER, 1949) and Modigliani, “New Developments on the Oligopoly Front” (JPE, 1958). The origin of limit pricing: an incumbent’s sustainable price is capped by what keeps entry unattractive. Ours differs by: the entrant already exists and publishes its weights, so the limit price binds immediately instead of as a deterrent.
Gaskins, “Dynamic Limit Pricing” (JET, 1971). The incumbent optimally lets share erode over time while harvesting margin. Ours differs by: the erosion rate being set by an open-weight release cadence rather than by capital accumulation, which makes it far faster and largely outside the incumbent’s control.

The neighbors arguing the other way.

Dylan Patel (SemiAnalysis) on Dwarkesh, 2026. Two labs absorb most incremental global compute and hold the resulting position. The stalking horse for this post. Ours differs by: granting the compute forecast entirely and attacking the claim that compute converts into durable rent. (link)
Aschenbrenner, Situational Awareness (2024). AGI on a short timeline, trillions in capex, nationalized clusters. Ours differs by: being agnostic on the timeline. the substitutes half is his world, and the labs still don’t capture it. (link)
Brynjolfsson, “The Turing Trap” (Daedalus, 2022). Argues the choice between automating and augmenting humans is a choice, with distributional consequences. Ours differs by: treating the distribution as an outcome of market structure rather than of a design decision — in both regimes here, nobody chose the split.

The moral

The optimistic reading of this post is the correct one, and I want to be clear about it: a world where the price of cognition collapses is a world that gets much richer, and most of that wealth reaches ordinary people. It just reaches them as cheaper everything rather than as bigger paychecks, which is how general-purpose technologies have always paid out and is a genuinely poor fit for how we measure whether people are doing well.

What the analysis will not support is the equity story. Between a physical world that refuses to scale on demand, an open fringe that caps the price, upstream suppliers who can see your margins, and a router that sends them the boring 95%, the value created by AI is going to be extremely hard to appropriate. The chokepoints that survive are not the models. They are energy, physical infrastructure, and raw compute — the parts you can’t fork.

Which leaves a pleasant irony to end on. The most likely outcome of the race to build god in a datacenter is a world of extraordinary abundance, delivered at cost, by companies whose shareholders would have done better in utilities.