Assume it fully arrives: a company is a team of Claudes with a bank account, robot hands, and a human board. Thinking is free. The only question left in the whole economy is — what does it still pay for?
Build vs. buy is decided by what the fixed cost is denominated in. Intelligence collapses every cost made of intelligence — and leaves standing every cost made of something else.
Part I The vendor list
An AGI company builds anything whose value is a computation, because computation now costs approximately nothing. It keeps buying exactly five things: atoms at scale, elapsed time (learning curves, histories, track records), legal rights that attach to entities rather than ideas, independence that by definition cannot be self-supplied, and coordination — the property of everyone-else-already-uses-this.
Everything else — most of what B2B commerce is today — doesn't get "built in-house." It dissolves into the agent. There is no artifact left to sell. The vendor ledger, verdict by verdict — open a row for the reasoning:
Why is the list complete? Because the five invoices are the practical decomposition of just three primitives — the things a superintelligence still cannot alter, no matter how smart:
Intelligence is a universal solvent — and these are the three things it can't dissolve: the universe, the past, and everyone else. Every durable business in the AGI economy is a position in one of the three. Test any candidate against them: brands (the past + everyone else), capital (a claim on future atoms, priced by time), trade secrets (thought — dissolves — unless they encode lived experiments). The list holds.
Part II The numbers
Sort today's gross world product not by sector but by what its cost is denominated in. Roughly a third of everything the world sells is priced in human thought — professional services, administration, software, media production, the great white-collar middle. A third is priced in atoms. The rest splits between human presence, coordination & trust, and rents.
Now let AGI diffuse through it. Two things happen at once, and they pull in opposite directions: the volume of cognitive work explodes by orders of magnitude, while its price collapses toward the cost of compute — which means the revenue pool shrinks even as output soars. The value doesn't vanish; it migrates to the two ends of the economy, and to the buyer as surplus. These figures are a stylized scenario, not a forecast — the point is direction and rough magnitude.
Read the blue band. That collapse is not a recession — it is the largest price deflation in history, experienced by buyers as abundance and by sellers of thought as extinction. Legal opinions, market analysis, software, diagnosis, design: output multiplies a thousandfold while the revenue pool drops by three quarters. Meanwhile every dollar that leaves the middle reappears at the ends — in the ochre of fabs, power, and robot fleets, and in the growing teal and rose bands the rest of this essay is about.
Part III Why the fab wins
Here is the repricing that matters, asset by asset. Ask one question of each: what would it cost to replicate this, if intelligence were unlimited and free? Whatever survives that question at a high price is where the economy's value concentrates — because price ultimately anchors to replacement cost, and AGI just set the replacement cost of thought to zero.
| Asset | Fixed cost, today | Replacement cost when thinking is free | Repricing |
|---|---|---|---|
| Salesforce's codebase | ~$3B of accumulated engineering | ≈ $0 — a weekend of tokens | ↓ to zero |
| A consulting partnership | Decades of recruiting & reputation | ≈ $0 — advice is tokens | ↓ to zero |
| ADP's payroll engine | Logic + 50-state licenses + liability | Logic → $0; licenses & liability unchanged | → thin shell |
| A leading-edge fab | ~$30–40B + supply chain + a decade of queue | Higher — demand explodes; the EUV machine doesn't care how smart you are | ↑ rises |
| 30 yrs down a motor learning curve | Wright's law, embodied in yield | Must be lived, not thought — the robot buys the component too | ↑ rises |
| 10 years of observed prices | Collection running since day one | Cannot be recreated at any price — time-locked | ∎ locked |
| A settlement benchmark's seat | Years of visible neutrality | Not for sale — "everyone else uses it" can't be built unilaterally | ∎ locked |
| One hour of a human's presence | Priced as labor | Rises fastest of all — supply is fixed at one life per person | ↑↑ Baumol |
This is why the fab becomes more valuable, not less. Today the market pays roughly ten times revenue for software carrying 80% gross margins, and half that for the fab that makes the software possible. That spread exists because software scaled and atoms didn't. AGI inverts it: when anyone's agents can rebuild any software over a weekend, competition drives its price to marginal cost — approximately zero — while every unit of new intelligence still queues for the same wafers, the same substations, the same learning curves. The multiple migrates from the thing that thinks to the thing that can't be thought into existence.
Commodity thinking becomes free. Frontier thinking becomes a utility bill. Trailing intelligence — human-level, a generation behind the frontier — gets crushed to the marginal cost of compute by competition and open weights. That's the solvent dissolving the middle. But the frontier stays scarce, because a frontier model's fixed cost isn't denominated in thought: it is the largest capex-and-time asset ever built. The frontier lab is a fab that manufactures cognition — which is why this page files it under atoms, and why it survives its own product.
So a tax lands on every economic act, metered like electricity: per-unit price collapsing, total bill exploding — Jevons guarantees it. The tax flows to the frontier labs, and then mostly through them: a lab's cost structure is compute and power, so the intelligence tax is largely a pass-through to fabs and substations. The labs' own margin is rent on the gap between frontier and commodity — real pricing power electricity never had (there is no "frontier grade" of electrons), but it lives exactly as long as the gap does. If capability plateaus, the rent migrates one more step down the stack, to pure atoms.
Put it all together and the income statement of every 2040 company collapses to four lines — plus everything that used to be between them:
Part IV The shape that's left
Put the pieces together and the economy takes a shape. Coase's answer to why firms exist — internal coordination is cheaper than market transactions — evaporates when both costs go to zero. What's left drawing boundaries is only minimum efficient scale of the non-cognitive asset. Smith outlives Coase.
And notice the feedback loop between the two ends: the more the atoms end concentrates — three suppliers of a critical component, a handful of compute providers — the less anyone trusts the participants' numbers, and the more the economy needs the referee end. The scale giants are structurally barred from supplying it: a fab can no more publish the neutral wafer-price index than a player can referee his own match. Neutrality is the one product the giants cannot make. Meanwhile the transaction explosion in the thin middle means every micro-trade between two agent-companies needs a reference price and a quality attestation from a third party. The coordination end doesn't survive AGI — it is multiplied by it.
And the barbell needs a bloodstream. The old middle was capital-light — you funded payroll. The ends are capital monsters — fabs, gigawatts, robot fleets, land — which makes this the most capital-hungry economy ever constructed, arriving exactly as the labor share of income collapses with the blue band. Money doesn't dissolve; it is the ur-coordination artifact, and a ×1,000 transaction explosion needs a settlement medium more than a quarterly-invoice economy ever did. Capital markets sit at the hinge of the whole shape: interest is the price of time — delivering atoms today against a claim on later is the one trade no intelligence can shortcut. The analysis of capital becomes free; the bearing of it, and the question of who owns the ochre band, becomes the politics of the 2030s.
Part V The human premium
In 1826 a string quartet took four musicians forty minutes. In 2026 it still does. While manufacturing productivity rose a hundredfold, live performance productivity rose by zero — so its relative price exploded. Economists call it Baumol's cost disease. It is about to have its final, largest episode: AGI moves every cognitive service into the fast-productivity lane, leaving exactly one good whose productivity can never improve — a human being, present, paying attention. Supply: one life per person. Non-replicable by construction.
But be honest about the boundary, because "people will always want the human touch" is mostly wrong. Bank tellers had the human touch. Travel agents had it. The premium survives on one condition only:
→ automated, no premium
→ Baumol premium, rising
It's supply-capped. Human hours don't scale, which is exactly why the price rises — and why the sector's share of spending grows (10–26% of the 2040 economy in the scenario toggle above) but can never dominate production.
It's a consumption good, not a production input. Wealth flows into it as the world gets richer; nothing downstream is built from it. It rides the boom at the ends of the barbell — it doesn't replace them.
It needs proof. A premium for humanness creates the incentive to fake humanness — performances, crafts, care, "hand-made" — at machine cost. The premium is only collectable with credible attestation. Proof-of-human becomes a coordination product like any benchmark: the rose band and the teal band need each other.
Epilogue The referee seat
Every version of this economy points the same way: more concentration in atoms, infinitely more transactions between agents, a rising premium on verified humanity — and all three raise demand for the one product neither end can supply itself: a neutral reference.
Benchmarks to settle machine contracts. Attestation that a number is true, that a datum has provenance, that a hand was human. History that was collected while it was still collectable. These are coordination assets — their fixed cost is denominated in years of visible neutrality, the only capex that can't be bought, rushed, or parallelized. Whoever holds a referee seat when the machines arrive keeps it for the same reason WTI outlived every oil trader who ever quoted it.
That is the seat worth building toward while everyone else is still selling thought.
Disclosure: this essay argues a thesis its author has a stake in — Emergent is building toward the referee seat. Read it with that bias priced in.