Samir Kaul, a partner at Khosla Ventures and an early backer of OpenAI, supplied both ends of the argument in a single CNBC interview, one paragraph apart, without appearing to notice.
On the way up: before the reset, a software startup could be sold to a larger company that wanted its engineers, at roughly two million dollars a head. A hundred-engineer shop cleared two to three hundred million on that logic. The price was for the team, not the product.
On the way down: fifty engineers now do what five hundred did five years ago. And the line that should have been the headline — the firm had to completely reshuffle how it valued these companies.
The $200 is the only number Kaul didn’t say. The market supplied it: a Claude Code Max subscription runs $200 a month. The same coder valued at $2 million on an acquihire is now narrated as absorbed by a $200 monthly seat. One source gave us the scarcity and the displacement. The market finished the sentence.
What $2 million a coder actually was
The diagnosis that follows applies classical price theory — Marshallian quasi-rent and the paradox of value — to this specific repricing. The economics is textbook. The application to the unicorn reset, and the claim that the venture class monetizes both the rent and its dissipation, is Luminity’s analytical contribution, not a finding drawn from any single source.
Kaul’s $2 million was not a measurement of what an engineer produced. It was a scarcity rent.
Economics has had the vocabulary for this since 1890. Alfred Marshall called it quasi-rent: the surplus a factor earns when its supply is fixed in the short run and demand runs ahead of it — a return above what is needed to keep the factor in place, and one that dissipates as supply expands. Marshall’s own examples are scarce specialized talent. The earnings of a superstar, a top surgeon, a sought-after barrister are, in his accounting, almost wholly rent.
The 2021 engineer was that factor. Capital was abundant and near-free; trained engineers were not. Acquihire bidding was not a contest over output — it was a contest to deny a competitor the team. The $2 million floor and the $3 million effective price were the rent. Almost none of it was a return on production.
Quasi-rent has one defining property: it does not survive an expansion of supply. That is not a risk attached to the price. It is the price’s expiration date, written into it from the start.
Abundance masquerading as scarcity
Adam Smith posed the puzzle in 1776: water sustains life and costs nothing; diamonds do nothing and cost a fortune. The resolution came a century later — value in use is not value in exchange, and price tracks the marginal unit under scarcity, not total worth.
The engineer was always water. Enormous value in use, the whole time. What spiked between 2019 and 2021 was value in exchange, and that was a property of supply, not of the work. AI did not lower the engineer’s value in use. It raised the supply of capable output. Fifty doing the work of five hundred is not a statement about worse engineers; it is a statement about an expanded margin. The exchange value fell to meet the use value it had always exceeded.
So the unicorns did not die of a productivity collapse. They died of rent dissipation — exactly, and on schedule, as Marshall described. The reset was overdue, not surprising. The only surprising thing is a venture class narrating a predictable expiration as a revolution it discovered.
Rent dissipates. Someone holds the receipt.
Here is the part the productivity story leaves out. The rent did not simply vanish. It was sold first.
The per-engineer and per-seat valuation logic of the boom was the basis on which inflated portfolios changed hands — and the incumbents bought them at peak. The same per-seat model that fed those acquisitions is now the one getting repriced in public: the workflow-embedded, charge-by-the-user companies are taking the hit, with the large enterprise software names down hard on the AI threat. The acquirer who paid the rent holds the receipt. The underwriter who priced it has moved on to funding the layer that dissipates it — the same capital flows that have routed more than a quarter-trillion dollars into the frontier model providers.
That is the structure worth naming. The venture class does not have to be right about the future to win the move. It collected on the scarcity by selling the rent, and it collects on the abundance by funding the thing that retires it. “Reshuffle how we valued these companies” is not a confession of error. It is a description of the product.
The $200 is not the productivity
One caution, because the symmetry is seductive. The $200 subscription does not replace the engineer any more cleanly than the $2 million ever measured one. A seat augments output; it does not retire a role at the price on the invoice. And the invoice is not fixed — metered token consumption can quietly exceed the subscription’s implied value, which is why reports of enterprises pulling back agentic-coding seats over runaway usage are not anomalies. They are the meter behaving as a meter. The bill is your architecture’s, not the model’s.
The honest claim is narrower and harder than the headline. Neither number was a productivity measurement. The $2 million was a rent priced under scarcity. The $200 is a sticker priced under abundance. The work in between barely moved. What moved was the story about how scarce it was.
The pre-ChatGPT unicorn reset is not evidence that AI made engineers cheap. It is evidence that the engineer’s price was a quasi-rent the entire time — extracted under manufactured scarcity, sold to incumbents at peak, and now retired on schedule by an expansion of supply that classical price theory said would retire it.
Abundance was always there. Scarcity was the framing. And the people who underwrite the framing know the difference between a rent and a return — which is the one thing the rest of the market is being encouraged not to ask.
