Thirty-one times revenue is what the roadshow has to defend. A $2 trillion ask against annualised revenue above $65bn at the end of July works out at about 30.8 times sales [1][3][4], where the $965bn private mark was about 14.8 [2]. So the pitch is that public money pays roughly double what the last private round paid, or rather, the more interesting version of the same pitch, that it pays the same multiple and the revenue line closes the distance: at 14.8 times, $2 trillion needs about $135bn of annualised revenue, a little over twice the July rate [3]. July was already running some 40% above the annualised second quarter [8], so that is a year of compounding rather than a fantasy.
Where the revenue comes from matters more than the slope. Second-quarter revenue of $11.6bn, reported by the Wall Street Journal, was 1.73 times OpenAI's $6.7bn [10][5], while Ramp's July reading has 43.5% of US companies buying Anthropic subscriptions or tokens against 39.7% for OpenAI [12], a customer base only about 1.10 times as wide [5]. The difference is spend per customer, and list price is a candidate explanation: Fable 5 is priced at $10 per million input tokens and $50 per million output against $4 and $20 for OpenAI's GPT-5.6 Sol, 2.5 times on both legs [14]. By Ramp's count, Fable's first-month share of enterprise model spend ran 1.9 times its share of tokens [4]. Buyers pay the higher price anyway.
The cost side is where the disclosure gap bites. The revolving facility Reuters reports Anthropic is trying to close equals 23% of the July run rate [6][7][4], and the source material does not say what a draw would fund. Gartner's forecast that cost per agentic workflow rises more than fivefold through 2028 even as unit token economics improve [15] is the reason a 31-times multiple needs a gross margin line beside it, and margins, cash burn, model costs and customer concentration are exactly what nobody outside the company has seen [5].
The $2 trillion figure itself is sourced to unnamed investors and backers, not to Anthropic [3][9], the company and its banks declined to comment [8], and the same people describing the calendar cautioned that it could move [16]. Two cheaper outcomes are available: file into a range nearer the multiple OpenAI's own mark implies [7] and let the aftermarket argue, or miss late September and the pre-midterm window with it [2]. The multiple survives only if the filing shows cost per token falling faster than Gartner's workflow cost climbs; a gross margin in the high fifties with no customer above single digits would settle it, and a redacted concentration table would settle it the other way.
