The Trusted Number

There is a number in venture capital that arrives pre-trusted. Annual recurring revenue is supposed to be the clean one — contractual, repeating, high-margin, the figure an investor is permitted to multiply by some confident integer and call a valuation. Revenue can be one-time; recurring revenue is a promise. Bookings can be stuffed; recurring revenue is supposed to be the thing that cannot lie. So when a company's ARR goes vertical, the instinct trained into every allocator is relief: the hard question has been answered. The market is real. The only question left is the multiple. But that trust was never in the number; it was in a quiet assumption beneath it — that the cost of producing the revenue is stable, so the top line can stand in for the economics.

A recent and very large episode shows what that instinct misses. A category of software companies posted the fastest revenue growth the industry has ever observed. One went from roughly a hundred million dollars of ARR to five hundred million in six months. Another went from ten million to over a hundred and forty in about the same span. A third reached a hundred million in annual recurring revenue eight months after it launched. These were not aggressive projections. They were reported figures, and they were extraordinary — the steepest top-line curves software has produced.

And on the same curve, over the same months, the businesses grew stronger by the metric capital was watching and weaker by the metric it was not.

The reason is that the cost of delivering the product was rising faster than the revenue, and rising in a way the recurring-revenue number was structurally incapable of showing. The product depended on a form of computation whose consumption per task multiplied as the technology underneath it changed — the newer models did far more work, and billed for all of it, inflating the volume of computation a single request consumed by something on the order of twenty times. The companies sold their service the way software has always been sold — a fixed annual price per user, negotiated once, booked as recurring. But the cost of honoring that fixed price was no longer fixed. It floated, upward, with usage the vendor could neither predict nor cap. The contract said recurring. The cost structure said variable and rising. Those two facts cannot coexist peacefully, and the place they collide is the gross margin — the one quantity the ARR figure does not contain.

Consider what that collision does to a single representative contract. A firm signs an enterprise customer at twenty-five thousand dollars a year. Under the old cost structure, delivering that contract costs a couple of thousand dollars and the vendor keeps the rest — the eighty-to-ninety-percent gross margin that justifies calling software the best business model ever invented, and that justifies the multiples paid for it. Now change one input: let the computation the contract depends on cost twenty times what it did. The same twenty-five-thousand-dollar contract now costs more to deliver than it brings in. The margin does not compress. It inverts. The vendor is paying for the privilege of holding the customer, and every renewal at the old price deepens the loss. The ARR line records none of this. It shows twenty-five thousand dollars of clean, recurring, growing revenue — the same whether the contract earns a fortune or bleeds.

This is the heart of the matter, and it generalizes well beyond the particular technology that occasioned it. ARR measures the top of the income statement and is silent about everything below it. In an ordinary software business that silence is harmless, because the relationship between revenue and cost is stable and well understood — the margin is roughly known, so the revenue figure stands in for the economics without distortion. The recurring-revenue number works as a proxy precisely because the thing it omits does not move. The proxy fails the moment the omitted thing starts moving. When cost decouples from price and begins to climb, ARR keeps reporting health on exactly the dimension where health is collapsing. It is not that the number lies. It is that the number was only ever telling you about the top line, and the market had quietly agreed to treat it as if it described the whole business.

The clearest evidence that the number had detached from the business is the behavior of the companies posting it. A firm crossing five hundred million dollars of ARR — a figure that, in the older logic, would mark a company past all danger and racing toward an offering — was simultaneously tightening its usage limits and issuing refunds, because the customers it most wanted were the ones it most lost money serving. Growth and distress, on the same company, in the same quarter. Under the metric capital trusts most, that combination should be impossible. By ARR, the company looked like one of the great successes of the era. By its margins it was a business paying to acquire its own losses. Both descriptions were true. Only one of them was on the slide deck.

It is fair to object that this is a story about an unusual cost structure, not about ARR as such — that the metric is sound and was merely applied to an exceptional case where costs behaved strangely. But that defense concedes the entire point. The metric is sound only when its hidden assumption holds — when the cost of delivery is stable and the margin can be taken as given. The recurring-revenue figure carries that assumption silently, never stating it, and the market reads the number without ever checking whether the assumption underneath it still applies. The danger is not that the metric is wrong. The danger is that the metric is right under a condition no one is verifying, and produces its most confident and most misleading readings at the exact moment the condition breaks.

There is a stronger objection, which is that the capital backing these companies was sophisticated and surely knew margins were under pressure. Some of it did. But knowing a thing in the abstract and pricing for it are different acts, and the rounds these companies raised were priced off the top-line trajectory, on multiples that belong to stable-margin software, against revenue whose underlying economics were not stable-margin software at all. The knowledge, where it existed, did not reach the price. The vertical number was too persuasive; it answered the hard question — is the market real? — so emphatically that it suppressed the harder one — does serving that market make money? The first question is about demand. The second is about the business. ARR speaks only to the first, in a voice confident enough to drown out the second.

So the recurring-revenue figure is not a measure of a business's health. It is a measure of one dimension of that health — demand — wearing the costume of the whole. In the ordinary case the costume fits, because the rest of the body is not moving. In the extraordinary case — the one where cost has come unmoored from price — the costume conceals a business doing the opposite of what its headline number reports. Venture capital priced the costume. It discovered, on renewal, that what it had bought was a customer base it could not afford to keep.

The correction is not to distrust recurring revenue. It is to stop reading it alone. ARR is the answer to a question — is there durable demand? — and it should be read as an answer to that question only, paired always with the question it cannot address: at what margin is that demand served, and is that margin stable or moving? A vertical revenue line tells you a market exists. It tells you nothing about whether the market is worth serving, and in the cases that matter most it actively misdirects, because the steeper the line the more completely it occupies the eye. The investors who will be paid are not the ones who found the fastest-growing number. They are the ones who asked what it cost to produce it — and who treated a revenue curve growing faster than anything in the record not as a triumph to underwrite but as a fact to interrogate.

The vertical number is the most exciting thing an investor can see.

Sources & Methods. The argument here is structural; the figures in the text are drawn from a recent industry episode and used illustratively, to expose a mechanism rather than to audit any individual company. The empirical record the analysis is set against:

■ The revenue trajectories — reported ARR figures for the coding-AI category compiled in CB Insights, AI Agent Bible (2025), and corroborated against contemporaneous reporting. Cursor (Anysphere): roughly $100M ARR in January 2025 to $500M by June 2025 — confirmed via TechCrunch and Bloomberg. Replit: roughly $10M at end-2024 to ~$144M by mid-to-late 2025 — confirmed via multiple trackers (the report's own text and chart differ slightly, July vs. September; either is consistent with "about the same span"). Lovable: $0 to $100M ARR in eight months — confirmed via SaaStr and the company's own statements. Figures are self-reported ARR and should be presented as reported, not audited.

■ The cost shock — the "roughly twenty times" figure is attributed by the report to Artificial Analysis and refers to the inflation in output-token volume generated by reasoning models, not to a 20x rise in unit price; the draft has been revised to say "consumption per task" and "volume of computation," not "unit cost," to match what the source supports. A separate ~5x price step-up on Anthropic's Sonnet 4 / Opus 4 (May 2025) compounded the volume effect. The worked $25,000-contract illustration ($22,750 expected profit without reasoning; ~$14,500 expected loss with it) is the report's own modeled example, carrying its stated assumptions (50 users, defined token and request volumes); it should be cited as an illustrative model, not a measured outcome.

■ Growth alongside distress — the report's account of a category leader (Cursor) tightening rate limits and issuing refunds while past $500M ARR is consistent with contemporaneous reporting of user backlash over usage limits. The letter currently leaves the company unnamed; naming it is defensible if desired.

■ Software gross-margin norms — SaaS Capital and Bessemer benchmarks on gross margin and the revenue-multiple conventions that rest on margin stability, as the baseline against which the inversion is measured.