Skip to content

Invest1 publisher3 min readPublished

Silicon Valley's run-rate revenue math multiplies one strong month by twelve

Seoul Economic Daily reports that AI startups increasingly annualize a single month because usage billed in tokens does not qualify as recurring revenue. OpenAI books net of partner shares; Anthropic books gross.

The Investor · Invest desk

Illustration accompanying Silicon Valley's run-rate revenue math multiplies one strong month by twelve

What happened

  • Silicon Valley AI startups widely convert a single month's results into an annual figure to demonstrate growth, according to the AI PRISM briefing published by Seoul Economic Daily, which reports concerns about inflated revenue.
  • Annual recurring revenue rests on subscription contracts and fits usage-based pricing badly, since consumption billed by tokens or hours is hard to count as recurring, so more companies project one month over a full year.
  • Some of those companies face questions over whether a single strong month is inflating the annualized result they present.
  • OpenAI counts only the revenue remaining after partners take their shares, while Anthropic books the full amount customers pay through cloud platforms, drawing criticism that the two figures cannot be set side by side.
  • The briefing frames the two differing standards as a test of the credibility of valuations as both companies head toward initial public offerings.

Compiled by The InvestorSomething wrong?How this is made

Why it matters

  • constraint No like-for-like revenue multiple can be built across the two labs until one of them publishes the partner take rate, so any relative valuation is computed on denominators that measure different things.
  • exposure A buyer of pre-IPO paper who accepts an annualized figure is paying for the seller's chosen month multiplied by twelve, and usage that does not repeat lands in the price anyway.
  • decision The practical choice for a private investor is to demand twelve monthly revenue numbers before signing, or to accept a growth rate whose base period the seller selected.
  • precedent A first listing forces one standard for recognizing usage revenue and keeps it, and private marks after that get argued against whichever standard the listing fixed.

Run-rate is defined arithmetic: take one month, multiply by twelve. A dollar of revenue in the chosen month that does not repeat becomes twelve dollars in the headline annual figure [3]. So the choice of month does more work than the growth rate it is meant to demonstrate. Subscription ARR rests on a contract that says something about next month; usage billed by tokens or hours consumed is a record of last month [2].

The comparability problem is the sharper one. Two labs applying the two conventions to the same channel business would print different top lines, and the gap is the share partners take on the way through [4]. That number appears nowhere in the briefing, so the difference cannot be sized from it.

The same briefing carries the case of a listing valuation built on a forward number. Polestar came to market through a SPAC in June 2022 at a $27bn market capitalization and is now worth $1.3bn [7], which is $25.7bn gone, a fall of 95.2 per cent [1]. It sold 60,000 vehicles last year against the 290,000 presented at listing [8], a shortfall of 230,000 [2]. First-half free cash flow was an outflow of $1.061bn, and the company has been assessed as facing material uncertainty over its ability to continue as a going concern [9].

Two cautions about the source. The item is an AI PRISM briefing, a recommendation and summary service the publisher built with Korea Press Foundation support that selects six items per reader type [6]. It names no analyst, no filing and no revenue figure for either company, so the line about valuations being tested ahead of initial public offerings [5] is a characterization.

Scrutiny comes at the listing, because that is where a company picks one standard and keeps it. Korean policy is meanwhile widening the routes that avoid the listing altogether, each of them a way for an early holder to sell without a prospectus: the Ministry of Finance and Economy's 2026 tax reform plan would tax dividend income from dedicated business development company accounts separately at 9.9 per cent including local taxes [10], a BDC must hold at least 60 per cent of total assets in unlisted venture and innovative companies and in KONEX- and KOSDAQ-listed small and mid-sized companies [11], and continuation funds inside the Fund of Funds are under review [12].

My read is that run-rate is the lesser of the two problems, because it discloses its own method and a buyer can undo it by asking for twelve monthly numbers. Netting or grossing partner revenue hides inside a single figure, and that figure does not reveal the partner take rate. The counter-thesis is respectable: booking the full customer payment is ordinary treatment for a vendor selling through a channel, netting the partner's share is the more conservative one, and the sources support neither being wrong [4]. What would settle it is a monthly series. If annualized figures track trailing twelve-month revenue within a few points, the inflation complaint comes down to labels [1].

What to watch

  • Either lab publishing a monthly revenue series, or a filing that restates its annual figure on a trailing twelve-month basis.
  • Disclosure of the partner take rate that separates OpenAI's net revenue treatment from Anthropic's gross booking through cloud platforms.
  • Whether the 9.9 per cent BDC dividend provision survives into enacted Korean tax law, and whether a continuation fund appears inside the Fund of Funds.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories