Silicon Valley's favorite metric is measuring the wrong business
Under standard accounting rules, much of what AI startups book as "recurring revenue" is closer to gross merchandise value — and ought to be reported that way
ROY LEE, chief executive of an Andreessen Horowitz–backed startup whose official motto is "Cheat on everything," told a TechCrunch reporter in late summer that his company's annual recurring revenue had doubled in a week to $7m. He later conceded on X that he had fabricated the number; the real figure was $5.2m. Bloomberg, revisiting the episode this week, called ARR the least-trusted metric of the AI era.
Lee's $1.8m fib is small beer next to the parade. Cursor, an AI coding tool, went from zero to $100m of ARR in roughly a year and crossed $1bn by November 2025, according to SaaStr. Lovable hit $100m in under six months; Emergent, an Indian coding upstart, claims to have reached the same mark in eight. Midjourney, ElevenLabs and Decagon have posted similarly vertiginous curves. Across the sector, founders are extrapolating last month's token consumption forward twelve months and calling the result "recurring." The pressure to do so is immense: one venture partner told Fortune that the expectation, set by Cursor and Cognition, is now revenue on day one — or an awkward conversation about what, exactly, the company is for.
Principal concerns
The trouble is not that the numbers are invented (Lee is an outlier) but that the word is. Under ASC 606, the accounting standard that governs revenue recognition, a company is either a principal — setting prices, carrying inventory risk, booking gross revenue — or an agent, facilitating a transaction and booking only its net cut. Most AI "wrappers" fall closer to the agent side. They do not control the pricing of the underlying model; they cannot stop OpenAI or Anthropic from raising rates; they are, in effect, resellers of frontier-lab compute with a user interface bolted on top. What they publish as ARR is therefore closer to gross merchandise value than to software revenue. Their true top line is the take rate — the sliver they keep after paying the inference bill — and in many cases that sliver is uncomfortably thin, occasionally negative.
This article is for Vector members. Start a 7-day free trial to keep reading.
Start your free trial
