AI revenue headlines now rest on definitions the buyer has to ask for

Usage pricing breaks the recurring assumption on both sides of the ledger. A customer metered in tokens can spend heavily one month, pull back the next, or switch to a cheaper model, and none of that requires renegotiating anything ^[6]. The vendor's own cost moves on the same meter, because it pays a model provider or a cloud company every time a customer uses the product ^[7]. A twelve-month projection built on that base inherits variance in the revenue line and in the gross margin at once, which is the variance a signed annual subscription used to remove, and the reason ARR functioned in the software era as an estimate of yearly sales from paying customers ^[1]. Matt Murphy of Menlo Ventures states the old bargain plainly: with ARR there is a signed contract, and you know what you have signed up for for the year ^[5].

Run rate is candid about the extrapolation and quiet about the base. It takes a month and multiplies by twelve, without claiming the revenue recurs ^[8]. So a $100 million run rate is a month of roughly $8.3 million ^[1], and each additional $1 million in whichever month gets chosen arrives in the headline as $12 million ^[2]. Pocket, which makes an AI recording device, says it crossed the $100 million run-rate threshold, and founder Akshay Narisetti says he avoids ARR because tokens make revenue too unpredictable to expect it to recur ^[9]. He also says he never annualizes one large month with a spike ^[10], a choice the vendor makes on his own, not one the metric forces on him.

The substitution can also read as more accurate rather than as a degradation, and the labs' own usage supports that reading. OpenAI describes subscription sales as ARR while calling its newer advertising business an annualized revenue run rate ^[11], and Anthropic has generally described overall sales as run-rate revenue ^[12]. Calling consumption revenue recurring would be the larger misstatement. What has disappeared is the shared definition behind the term: the same three letters now cover signed subscriptions, deals not yet closed, token consumption, hardware sales and one good month ^[2].

What the record does not contain is a base rate. Array Ventures general partner Shruti Gandhi calls it ARR inflation, says there is no accountability, and reports that founders she has called out answer that everybody else is doing it ^[3]. Greg Isenberg of Late Checkout wrote that the number of VC-backed AI companies lying publicly about ARR is unsettling ^[17]. Alexander Niehenke of Scale Venture Partners reads the padding as late-bull-cycle behaviour, comparable to the tail end of 2021, when founders hold the leverage ^[4]. The one named founder who has admitted misstating the figure is Cluely's Roy Lee, to a TechCrunch reporter ^[15], and private startups face far less scrutiny than public companies, with revenue claims rarely verified ^[16]. That leaves the actual prevalence of padding unmeasured, even for the investors describing it as common.

For a decision this quarter that gap matters less than it sounds, because the diligence response is the same whether the practice is universal or rare: the number is worth what its definition is worth. The questions that settle it are which month, what share is contracted rather than consumed, what last year's cohort renewed at, and what the margin is after inference costs. Linear's Karri Saarinen, whose company crossed $100 million in ARR selling multiyear contracts and has operated since 2019, says the operating history is what lets him forecast actual ARR, and that for newer AI startups it has gotten really murky ^[13][14]. Until usage-based businesses adopt a disclosure convention that survives an audit, the definition belongs to whoever writes it into the data room request.