--- title: "Anthropic's reported $2 trillion IPO rests on customers like Harvey, whose gross margin fell below minus 50%" description: "Harvey's books show why the two customers carrying a fifth of Anthropic's first-half revenue have an exit: post-train open weights, keep Claude for the hard 20%. The prospectus must price both." publisher: "The Inference" section: "Deals" published: 2026-10-24T17:11:02.246Z modified: 2026-10-24T17:11:02.246Z canonical: https://theinference.org/article/anthropic-s-reported-2-trillion-ipo-rests-on-customers-like-harvey-whose-gross-margin-fell-below-minus-50 language: en keywords: "Anthropic, IPO, Open Weights, AI Startups, Harvey" --- # Anthropic's reported $2 trillion IPO rests on customers like Harvey, whose gross margin fell below minus 50% > Harvey's books show why the two customers carrying a fifth of Anthropic's first-half revenue have an exit: post-train open weights, keep Claude for the hard 20%. The prospectus must price both. [Anthropic](https://theinference.org/markets/companies/anthropic) is days away from the filing that prices it at a reported $2 trillion, and the draft numbers leaking through the Wall Street Journal carry a warning the growth headline hides: [two customers were 20% of first-half revenue](https://www.calcalistech.com/ctechnews/article/xe3rgeeey), roughly $3.26 billion [2][4]. Then [this week's Bloomberg report on Harvey](https://es-us.noticias.yahoo.com/harvey-startups-usan-modelos-abiertos-061442802.html) showed what happens on the other side of that ledger. [Harvey](https://theinference.org/markets/companies/harvey), the legal-AI company valued at $15.6 billion, saw gross margin fall from about 50% at the start of 2026 to below minus 50% by June, after a March update drove token consumption up more than 20 times from January [1][3]. Margins turned positive again only after Harvey stopped renting the hardest-working portion of its stack [1][3]. ## The customer with the 20x token bill builds its own model In August Harvey launched Tenet, its first model post-trained on the open-weight Kimi K3 from China's Moonshot AI, with Fireworks supporting [1]. Harvey says Tenet completed nearly twice the reserved tasks of its LAB benchmark versus stock Kimi K3, and that open weights cut the cost per token while training aimed to cut tokens per task [1]. [Anthropic's own investor-forum slide](https://thenextweb.com/news/model-costs-startups-open-weights) acknowledged the pattern while conceding Harvey still needs Claude Opus for its hardest tasks [1][3]. *Chart: **One quarter now beats Anthropic's entire first half** Bars showing Anthropic revenue by period: $1.1B in H1 2025, $16.3B in H1 2026, $16.8B in Q3 2026 alone, with the Q3 bar marked as an estimate.* *Reported revenue in US$ billions. H1 2025 and H1 2026 figures come from the draft S-1 reported by WSJ; Q3 2026 is an unaudited company claim from an investor day. Sources: 2, 4. [2][4]* ## Four more verticals make the same call [Decagon](https://theinference.org/markets/companies/decagon) now routes 80% of customer queries to models it controls [1][3]. [Abridge](https://theinference.org/markets/companies/abridge) is building a clinical model on [NVIDIA](https://theinference.org/markets/companies/nvidia)'s open Nemotron family. Ramp raised $750 million in June and is weighing training its own models for the first time; co-CEO Karim Atiyeh said it "made absolutely no sense a year ago" and now does [1][3]. [Rogo](https://theinference.org/markets/companies/rogo) is studying the move. ## What the $2 trillion is actually priced on The math cuts both ways. Anthropic's first-half compute spend fell to about 55 cents per revenue dollar from over $2 a year earlier, and Q3 revenue of $16.8 billion annualizes past $60 billion [2][4]. But [a draft S-1 lists roughly $450 billion in noncancelable compute commitments](https://memeburn.com/anthropic-s-1-why-the-80-7b-loss-is-mostly-non-cash/), and the growth story assumes big spenders keep renting [4]. Harvey's books, Decagon's routing and [Replit's move, which we covered on October 21](https://theinference.org/article/replit-s-exit-from-anthropic-s-top-five-lands-the-same-week-anthropic-asks-the-market-for-750-billion), all say the highest-volume tasks are the first to leave the API [4]. The audited number that decides the listing, the public S-1, must appear by about October 25 to hold a November 9 roadshow [4]. That filing lands this weekend; the customer-exit footnote is the page to read. Deep dive below to learn more. ## Takeaway Anthropic's reported $2 trillion listing assumes its biggest customers keep renting Claude. Harvey's own books show why they may not: post-train open weights, keep Claude for the hardest 20%. The public S-1, due by about October 25, is where that customer-exit risk gets priced. ## Sources 1. [Yahoo Noticias (es), "Harvey y startups usan modelos abiertos para reducir costos", 28 September 2026](https://es-us.noticias.yahoo.com/harvey-startups-usan-modelos-abiertos-061442802.html) 2. [Calcalist/CTech, "Anthropic's $2 trillion IPO question: How much growth can investors buy?", 24 October 2026](https://www.calcalistech.com/ctechnews/article/xe3rgeeey) 3. [The Next Web, "AI model costs are pushing startups towards cheaper open weight" (citing Bloomberg), 21 September 2026](https://thenextweb.com/news/model-costs-startups-open-weights) 4. [Memeburn, "Anthropic S-1: Why the $80.7B Loss Is Mostly Non-Cash" (citing WSJ draft S-1), 22 October 2026](https://memeburn.com/anthropic-s-1-why-the-80-7b-loss-is-mostly-non-cash/) --- Anthropic's reported $2 trillion IPO rests on customers like Harvey, whose gross margin fell below minus 50% — The Inference. Canonical: https://theinference.org/article/anthropic-s-reported-2-trillion-ipo-rests-on-customers-like-harvey-whose-gross-margin-fell-below-minus-50