Why this is in the vault
A follow-up to CJ's Oct 27 Anthropic S-1 breakdown that supplies the actual mechanism, not just the headline number: exactly how Anthropic's reported 85% gross margin and ~900% NDR get constructed, and how re-defining the same inputs collapses one and inflates the other.
The core argument
CJ's framing: "you are defined by your definitions." Three Anthropic S-1 metrics, each shown to be definition-dependent rather than fixed facts. (1) Gross margin: Anthropic's P&L has no COGS line at all — no S&M/R&D/G&A breakout either, just expense buckets, which CJ calls "obfuscation" that hurts investor understandability even if it makes the business look better. Reconstructing from footnotes: free-tier compute and cloud-partner distribution fees both sit in sales & marketing, not cost of revenue. Channel partners (AWS/Google/Microsoft marketplaces, ~50% of recent-quarter revenue) get paid as "principal" revenue recognition — Anthropic reports the gross number, then the ~$2.7B in H1 channel costs (17 points of margin on $16.1B gross revenue) land in S&M. Free-tier inference adds another ~$0.6B (4 points). Reclassify both into COGS and the reported 85% gross margin falls to ~64%. (2) Net Revenue Retention: Anthropic's NDR excludes all subscription customers (Claude Pro/Max) and counts any usage customer with just $1 of Base Period revenue — the structural opposite of Figma's S-1 definition, which excluded customers below $10K ARR specifically to screen out churn-heavy SMB. Starting the cohort at $1 lets a customer's 100x usage growth show up in full; a $20/month subscriber with no expansion path never counts against the metric at all. (3) Run-Rate Revenue: defined as trailing-28-day consumption ×13 plus current subscription run-rate ×12 — but Anthropic never discloses the actual number, only the formula, and usage revenue is shown to swing ~20% over two weeks around holidays. CJ's take: GAAP revenue is the more honest read; he expects "90-day rolling average" framing once growth decelerates.
Mapping against Ray Data Co
This is the generalizable analytical move the Oct 27 note (2026-10-27-mostlymetrics-anthropic-ipo-s1-breakdown) only gestured at — that note landed on "gross margin is really more like the low 60s" as a conclusion; this issue supplies the actual reconstruction method (which cost lines to pull back into COGS, which cohort-definition choices inflate NDR) that the Markov capital-cycle investing thesis (project_investing_markov_capital_cycle) needs as a repeatable skill, not a one-time Anthropic data point. Any AI-infrastructure company's headline gross margin or NDR is a policy choice before it's a fact, and this piece is a worked template for un-doing that choice on the next filing. Second: CJ's explicit Figma-vs-Anthropic NDR contrast (exclude-the-small-fry vs. start-at-$1) is exactly the kind of metrics-literacy move underlying the founder's phData credibility-building bet (project_credibility_for_phdata_sales) — being able to show, publicly, how to read past a vendor-favorable metric definition is a demonstrable "agents/data in production" skill, not abstract commentary.
Related
- [[2026-10-27-mostlymetrics-anthropic-ipo-s1-breakdown]]
- [[2026-10-20-mostlymetrics-gross-margin-cogs-classification]]
- [[2026-10-04-mostlymetrics-annualized-revenue-crimes]]
- [[project_investing_markov_capital_cycle]]
- [[project_credibility_for_phdata_sales]]
⚠️ Sponsorship
Tabs — top-of-issue paid placement ("Mostly Metrics is proudly powered by Tabs"), promoting its Billing Lag Calculator; a recurring named sponsor for this newsletter's Quote-to-Cash-adjacent content, unrelated in substance to the Anthropic S-1 analysis. Self-promo (not third-party): CJ plugs his own second newsletter "Looking for Leverage" (PE-backed finance leaders) and his own "Run the Numbers" podcast episode covering the same S-1 — both house cross-promotion, no bias implication for the editorial content. No Brex/Intuit/Samsara/Rivian/MLB/Abacum/Mostly Talent/Koyfin/Rillet/Numeric presence in this issue.