Ask three analysts what an artificial intelligence company’s annual recurring revenue is and you will get three numbers, a disagreement about methodology, and a strong hint that the metric has stopped doing the job it was invented for. Competing estimates have become the norm rather than the exception, and the arithmetic behind the gaps is where the real story about this cycle is being written.
Annual recurring revenue was a useful invention. Software sold by subscription had a cash-flow profile that quarterly revenue reporting flattered badly, so operators started quoting the annualised value of contracted subscriptions instead. It worked because the underlying thing being measured was genuinely boring: a customer had signed for a year, the price was fixed, the cost of serving them was near zero, and multiplying the monthly figure by twelve told you something true about next year.
Almost none of those conditions hold for the companies the metric is now being used to value. Usage-based pricing means the monthly figure is a function of how much the customer happened to consume last month, not a contractual floor. Gross margins vary with inference cost, so revenue and profit have been decoupled in a way subscription software never managed. Contracts include credits, pilots, committed-spend arrangements that may or may not be drawn down, and in several cases revenue that flows between companies with equity positions in each other. Multiply a single month of that by twelve and you have not produced an annual figure. You have produced an extrapolation wearing an annual figure’s clothes.
A barista knows this problem in miniature. A cafe that serves two hundred cups on a Saturday in December can multiply out to an impressive annual number, and that number will be wrong in both directions depending on which month you picked. The honest operator quotes the trailing year, because the trailing year includes the quiet Tuesday in February when nine people came in. The metric is not deceptive. The choice of which month to annualise is where all the deception lives, and nobody has to say a false word to pull it off.
The insight is that a definitional vacuum is not a neutral condition. When a metric has no agreed definition, the version that circulates is the version the most interested party published first. Companies are not obliged to reconcile, because annual recurring revenue is not an accounting standard and never was. It appears in no audited statement, has no governing body, and carries no penalty for inconsistency. It is a marketing figure that the capital markets agreed to treat as an accounting one, and the agreement held only while the companies using it were small and the numbers were dull.
They are no longer small and the numbers are no longer dull, which is why markets turning nervous expresses itself first as a fight about measurement. Nervousness in a bull market looks like methodology disputes. Investors who were content to accept a reported figure start asking whether it includes compute credits, whether it annualises the best month, whether the counterparty is also a shareholder, and whether the growth rate is customers or consumption. Those are the right questions and they were always available. What changed is the appetite for asking them out loud.
The companies at the centre of this are building genuinely enormous businesses and the revenue is real money arriving in real bank accounts. The gap between the estimates is not evidence of fiction. It is evidence that an industry outgrew its instrumentation, and the instruments are now being rebuilt in public by people with opposing incentives.
Somewhere there is a spreadsheet with the correct number in it. Everybody quoting the figure in public is quoting a different cell.
Source: Newcomer, “Competing ARR estimates and fuzzy valuation math”
