06-reference

mostlymetrics anthropic ipo s1 breakdown

2026-10-27·reference·source: Mostly Metrics·by CJ Gustafson
ipoanthropics1-filingrevenue-per-employeenet-revenue-retentionai-economics

Why this is in the vault

Ground-truth financial metrics pulled straight from Anthropic's own S-1 (900% net revenue retention, $6.6M revenue-per-employee, real vs. headline gross margin) — and the fact that this is the third sender to cover the same filing within 48 hours (Every's "Is Anthropic Worth $2 Trillion?" yesterday, Stratechery's cost-of-compute angle today) is itself a signal that this filing is the highest-value primary-source data point of the year for anyone pricing Anthropic's enterprise trajectory.

The core argument

Anthropic filed its S-1 to list on Nasdaq (ticker ANT), with bankers reportedly pitching a ~$2 trillion valuation — roughly double the $965B Series H from five months earlier. CJ's numbers: $19.7B TTM revenue through June 2026, $16.3B in H1 2026 (up 1,300% y/y), split 80% consumption/"pay-for-intelligence" and 20% subscription. Over 11,900 customers pay $100K+/year, ~1,900 pay $1M+, and NDR runs ~900% (vs. Snowflake's 158% and GitLab's 152% at their own filings). Headline gross margin is ~85%, but CJ walks through the footnotes to show free-tier compute and cloud-partner fees are buried in sales-and-marketing, not cost of revenue — load those back in and the real number is "more like the low 60s." Revenue per employee is $6.6M (up from $1.4M a year ago, on ~5,300 employees), against CJ's $450K public-company rule of thumb. The business was profitable for exactly one quarter ($974M operating income in Q2, an 8% margin) but posted a $693M operating loss for H1 overall, plus an $80.7B GAAP net loss driven almost entirely by non-cash convertible-note markups. Five named red flags: cascading/interconnected risk factors, Taiwan/TSMC chip concentration, expensive non-cancelable compute commitments, ongoing copyright exposure (the Bartz settlement, book-scanning details), and a governance structure where founders hold 50.1% of the vote via a single Class F share. On distribution: Anthropic sells through all three major clouds (48% of H1 revenue ran through cloud marketplaces, up from 11% in 2023), plus a consulting channel (Accenture, Deloitte, KPMG, PwC) and embedded distribution through Salesforce and Snowflake. CJ's own framing question: can Claude be "the most impressive technology to ever exist" and still get commoditized and margin-compressed like commercial aviation.

Mapping against Ray Data Co

The filing names Snowflake explicitly as an embedded-Claude distribution partner ("Salesforce and Snowflake embed it in their platforms") — that's the exact seam the founder's phData cert-escalator bet (project_phdata_cert_escalator_path) sits on: Snowflake certs plus Anthropic certs as two order-agnostic escalators, priced as a bet that Snowflake-Anthropic integration deepens rather than one platform absorbing the other. Anthropic's own S-1 confirming that channel as material (not just a feature-page claim) is a stronger confirmation of that thesis than anything phData-internal could provide. Second, the gross-margin footnote-vs-headline gap (85% stated, ~60s% real once free-tier compute and cloud-partner cuts are loaded back in) is a direct, numeric gut-check for the Markov capital-cycle investing thesis (project_investing_markov_capital_cycle): it's primary-source evidence that even the frontier lab with the best unit economics in the dataset is still absorbing real compute cost through indirect accounting lines, which is the same obscuring mechanism the thesis has to price into any neocloud or chip-cycle read. Third, NDR of ~900% paired with CJ's own observation that Claude is simultaneously "a commodity product with little to no switching costs" is the sharpest version yet of the question this series keeps raising for any phData-adjacent positioning: enterprise AI spend is growing explosively and is not sticky by product lock-in, which argues for services/integration (phData's actual lane) capturing durable value rather than the model layer itself.

Related

⚠️ Sponsorship

Numeric — top-of-issue paid placement ("Mostly metrics is proudly powered by Numeric"), promoting Alec Schon's finance-engineer newsletter (weekly Claude-skill downloads). Known rotating-pool/affiliate sponsor for this newsletter, unrelated in substance to the Anthropic S-1 analysis — no bias implication for the editorial content. No Brex/Intuit/Samsara/Rivian/MLB/Abacum/Mostly Talent/Koyfin/Rillet/Tabs presence in this issue.