"There are no new crimes, just new ways to calculate revenue" — CJ Gustafson (Mostly Metrics)

Email subject: "There are no new crimes, just new ways to calculate..." Snippet: "Benchmarks for Operators 10/4/26."

Why this is in the vault

CJ's reconciliation of Anthropic's leaked FY25 financials against the 28-day-annualized-revenue trick (Higgsfield founder Alex Mashrabov's admitted "same way OpenAI and Anthropic do" methodology, surfaced in a Harry Stebbings interview) is the clearest public unit-economics breakdown available for the Anthropic IPO the vault is already tracking, plus a reusable red-flag test for spotting inflated AI-company revenue claims generally.

Mapping against Ray Data Co

This is a direct diligence input for [[project_investing_markov_capital_cycle]]: CJ pins Anthropic's IPO to "late November" and reconciles the leaked FY25 numbers ($4.6B actual revenue, net loss $42B of which $34B is a noncash convertible-instrument revaluation, leaving an operating loss of roughly $8B — i.e., ~$12.6B total cost to generate $4.6B of topline) against the ~$65B annualized run-rate figure that's been driving the OpenAI-vs-Anthropic revenue-race headlines this vault already tracks ([[2026-10-03-innermost-loop-argon-cpa-anthropic-ipo-autowarcom]], [[2026-09-30-innermost-loop-devday-dots-anthropic-518b-si-accord]]). The 28-day-times-13 annualization mechanic he exposes is a reusable underwriting heuristic: any AI-company revenue claim built on a 28-90 day window is closer to theoretical capacity than realized revenue, and that distinction matters directly when underwriting an Anthropic IPO entry once it prices. Secondarily, the operating-loss-vs-revenue math is a concrete frontier-lab unit-economics data point for [[project_credibility_for_phdata_sales]] — exactly the kind of agents-in-production cost structure a phData pitch references.

The core argument

• Anthropic's leaked FY25 numbers: $42B net loss, $4.6B revenue, plans to spend ~$500B on compute. But $34B of the loss is a noncash revaluation of convertible financing (dilution risk to shareholders, not a cash operating cost) — the operating loss from running the business is closer to $8B against $4.6B revenue, i.e. ~$12.6B total cost to generate $4.6B of topline.
• Separately, Higgsfield's CEO told Harry Stebbings they calculate annual revenue "the same way OpenAI and Anthropic do": take the last 28 days of revenue and multiply by 13. CJ cross-checked this against Anthropic's own numbers — $4.6B actual 2025 revenue vs. the ~$9B annualized figure being quoted at the same year-end — and the gap checks out.
• The 28-day-annualized number is an expression of theoretical capacity, not realized revenue. It flatters fast-growing companies (every incremental dollar from a good month gets "13 little friends") and will equally punish them on the way down (a bad month annualizes just as efficiently).
• CJ proposes adding "28-day annualized revenue" as a new rung to his Lead Edge revenue-quality hierarchy — one level above CARR in questionable-but-not-dishonest territory, since it's at least live revenue rather than ramped three-year contract value.
• Calls the Anthropic 2026 print (not yet reported) the real test of whether unit economics are improving, and ties the open question to "a late November IPO."

Issue contents

Weekly Valuation and Efficiency Metrics benchmark block (Koyfin-sourced, recurring weekly feature):

• Revenue Multiples — EV/NTM-revenue comps across 9 sector indexes (~131 public companies total): Security & Identity, Data & AI Infrastructure, Dev Tools & Observability, Horizontal SaaS & Back Office, GTM/MarTech, Vertical SaaS, Take-Rate Platforms, Payments & Money Movement, Consumer Fintech/Lending/Crypto.
• Efficiency — restated formulas for CAC Payback Period, Revenue per Employee (~$450k/employee rule of thumb at scale), and Rule of 40, plus a note that fewer companies now disclose exact Net Dollar Retention.
• OPEX — GAAP cost-bucket breakdown (COGS, S&M, R&D, G&A) framed around operating leverage.
• Reader Happy Hour event promo (London 10/15, Boston 10/20) — no RDCO relevance.

⚠️ Sponsorship

• Brex (masthead, "proudly powered by" block): full narrative ad for corporate cards + AI-native close/reconciliation workflows, CTA to brex.com/grow?partnerId=metrics. Confirmed recurring third-party sponsor in the known pool (first logged 2026-05-11).
• Koyfin ("our data partner," affiliate link ?via=metrics on the weekly benchmarks CTA): recurring affiliate relationship, confirmed standing since 2026-08-02. CJ has an ongoing commercial incentive to keep citing Koyfin as the source for this weekly data block — treat the benchmark charts as sourced from a paying partner, not neutral tooling.
• Tabs did not recur in this issue. The 2026-10-01 "Quote-to-Cash Part 1" issue introduced Tabs as a new 3-part series sponsor; this 10/4 issue runs Brex as masthead sponsor instead, consistent with the unknown-until-scanned posture for this newsletter. Watch whether Tabs resumes in Quote-to-Cash Parts 2-3.
• No Mostly Talent (CJ's own recruiting arm) self-promo observed in this issue.

Related

• [[2026-05-03-mostlymetrics-revenue-hierarchy-is-it-cake]] — the original six-level Lead Edge revenue hierarchy CJ references here and proposes extending with a new 28-day-annualized-revenue rung.
• [[2026-06-11-mostly-metrics-complete-guide-to-arr]] — CJ's fuller taxonomy of CARR-as-ARR abuse, which this issue calls "way less criminal than" the 28-day annualization trick.
• [[2026-10-03-innermost-loop-argon-cpa-anthropic-ipo-autowarcom]] — same-week coverage of Anthropic targeting an IPO at up to $2T; this note's unit-economics reconciliation feeds that thread directly.
• [[project_investing_markov_capital_cycle]]
• [[project_credibility_for_phdata_sales]]