Anthropic S-1 teardown (filed 2026-10-26)
Anthropic publicly filed its S-1 on EDGAR on 2026-10-26 (CIK 2130357; it had filed a confidential draft in June [report]). Class A shares, Nasdaq Global Select, ticker "ANT". Goldman Sachs is the left lead, then Morgan Stanley, plus 20 other banks. The price range, share count, raise size and valuation are all blank in this filing. Every number below is tagged: [S-1] = read from the filing, [report] = press, [Ray] = my arithmetic or inference.
Why this is in the vault
This is the primary source on the vendor at the center of the founder's main bet (phData's Claude practice and his Anthropic cert path). It also states, in the issuer's own words, the "production gap" thesis that our credibility lane is built on.
⚠️ Structural Bias
A prospectus is a selling document written by the issuer and its underwriters. The non-GAAP metrics and the "Strengths of Our Financial Model" framing are chosen by Anthropic. The risk factors (more than 90 pages, per CNBC) are lawyer-driven and list every risk, so they overstate some risks and hide which ones matter most. Run-Rate Revenue annualizes a trailing 28-day window (S-1 definitions). Q3 revenue is not in the filing.
The numbers that matter
| Metric | Value | Source |
|---|---|---|
| Revenue, FY2023 / FY2024 / FY2025 | $41M / $386M / $4.585B | [S-1] summary financials |
| Revenue, H1 2025 / H1 2026 | $1.163B / $16.286B (about 14x) | [S-1] |
| GAAP operating loss, H1 2026 | -$693M (margin -4%); Adjusted operating margin +10% | [S-1] non-GAAP table |
| Net loss, H1 2026 | -$80.7B, of which -$72.4B is a non-cash fair-value markup of convertible notes and SAFEs | [S-1] |
| Operating cash flow, H1 2026 | +$3.9B (vs -$1.5B in H1 2025) | [S-1] cash flows |
| Compute and infrastructure as % of revenue | 201% in H1 2025, 55% in H1 2026 | [S-1] MD&A |
| Cash + short-term investments, 2026-06-30 | $131.7B; $141.1B of capital raised to date | [S-1] |
| Revenue mix, H1 2026 | 80% consumption (API / committed spend), 20% subscription | [S-1] MD&A |
| Revenue through CSP marketplaces | 47% (FY2025), 48% (H1 2026) | [S-1] financial statement notes |
| Concentration | Top 5 customers about 25% of 9M-2026 revenue; none above 10% | [S-1] risk factors |
| Customers above $100k TTM | about 1,500 (end of 2025) to more than 11,900 (2026-09-30); about 250 above $10M | [S-1] |
| Net Revenue Retention | about 900% (2026-09-30) | [S-1] |
| Revenue from usage-policy violators | an estimated 5-7% of 9M-2026 revenue | [S-1] MD&A |
| Headcount | about 5,300; TTM revenue per employee above $6.6M | [S-1] Business |
| Named compute commitments | Google $114.5B, Amazon $110.0B, Microsoft $44.9B, SpaceXAI up to $120.5B, Nscale up to $44.6B, Lambda $41.3B, Broadcom leases $161.2B. Excluded from the table: Google TPU purchases estimated at $125-150B, AMD above $20B | [S-1] Contractual Obligations |
| Sum of named commitments | about $637B before the excluded TPU and AMD amounts. This mixes firm and "up to" amounts | [Ray] |
| Valuation / raise | Last private round $965B; speculation of about a $100B raise at about $2T; trading possible before Thanksgiving | [report] Axios; earlier vault notes |
Governance in one paragraph
Anthropic is a public benefit corporation (PBC) with five share classes [S-1 cover]:
- Class A (public): one vote per share.
- Class N (the filing says "primarily strategic business partners"; Amazon and Google confirmed, others [Ray] inference): non-voting.
- Class T: held by the Long-Term Benefit Trust, which elects the board majority.
- Class F: one share held by the co-founders' Founder LLC, carrying a fixed 50.1% of general voting power.
- Class E (employees): votes on limited matters only.
It will be a Nasdaq "controlled company". Separately, 5% of fully diluted shares went to an Anthropic Public Benefit Fund. The board may issue up to another 2% each time the share price rises 50% above the prior milestone [S-1]. The design means public holders get economic exposure with almost no control, plus a dilution mechanism that kicks in when the stock does well.
What we make of it
- The operating business turned the corner in H1 2026. GAAP operating loss was nearly breakeven and operating cash flow was positive [S-1]. The $80.7B "loss" headline is almost all non-cash note markups [S-1]. The revenue story is better than the headlines suggest.
- The balance sheet risk is forward commitments, not burn. Roughly $637B+ of named compute deals runs out to 2036-2045 [Ray, from S-1]. That is about 20x annualized H1 revenue [Ray]. The bet is that revenue keeps compounding into a fixed cost base.
- The enterprise channel is the cloud providers. About half of revenue flows through CSP marketplaces [S-1], so AWS Marketplace and Bedrock are the de facto sales channel.
- The S-1 says outright that deployment is the bottleneck. Fortune 500 spend on Claude is under 0.1% of their expenses [S-1]. The Frontier Academy exists because of "the shortage of engineers who can take AI from an initial use case into production" [S-1].
Exciting
- About 14x year-over-year H1 revenue growth, with compute dropping from 201% to 55% of revenue [S-1].
- The ecosystem is written into the prospectus. It names SIs (Accenture, Deloitte, KPMG, PwC) and data platforms (Snowflake, Salesforce). The Claude Partner Network committed $100M in 2026 and has 150,000+ certified practitioners across 4,000+ organizations [S-1].
- The Claude Frontier Academy ($100M, October 2026) aims to train 10,000 engineers by end-2027, "beginning with cohorts from leading consulting firms" [S-1].
- Cowork adoption is outpacing Claude Code's adoption at the same stage [S-1].
Concerning
- Control: Founder LLC holds 50.1% of votes and the LTBT elects the board majority. Public Class A holders cannot influence a change of control [S-1 risk factors]. On top of that, Public Benefit Issuances dilute holders after each 50% price rise [S-1].
- Compute commitments: the Google and Amazon deals are take-or-pay with true-ups [S-1]. The Nscale and Lambda capacity depends on those neoclouds getting financing or leases [S-1]. That is happening in a month when CoreWeave and Nebius fell more than 30% on data-center credit fears [report, CNBC].
- Revenue quality: 5-7% of revenue comes from customers later found to be violating usage policy [S-1]. Over 80% of Q3 revenue came from models released in the prior six months [S-1], so revenue depends on a constant release cadence. Run-Rate Revenue is a 28-day annualization (see [[2026-10-04-mostlymetrics-annualized-revenue-crimes]]).
- Channel dependency: Amazon, Google and Microsoft are compute suppliers, sales channels, investors and competitors all at once [S-1]. The top five customers are about 25% of revenue [S-1].
- Valuation: about $2T [report] against roughly $33B of annualized H1 revenue is about 60x [Ray]. The comp that matters is SpaceX: it raised about $86B at $1.77T and has traded below its $135 IPO price at points since [report, Axios].
Mapping against Ray Data Co
(a) phData and the founder's career: strong tailwind, but the cert is now table stakes.
- 150,000 people hold Anthropic certifications [S-1]. His CCA-F (850) is an entry ticket, not a differentiator.
- The scarce asset is the Services Track's other legs: production joint customers and public stories ([[2026-06-11-phdata-anthropic-partner-tier-cert-leverage]]).
- Concrete moves:
- Get phData into a Frontier Academy consulting cohort. The S-1 describes "each participant leading a Claude deployment at their own organization," which is close to his DSA/TAL seat.
- Keep the Architect Pro (12-22) cert.
- Pitch Snowflake + Claude as a joint motion, since the S-1 names Snowflake.
- Watch Ode with Anthropic [S-1]. It is an Anthropic-backed, PE-partnered services company for mid-sized firms, so it is a potential competitor to consultancies in phData's mid-market [Ray inference].
(b) Credibility.
- The S-1 is third-party-grade evidence for The Denominator thesis ([[2026-10-02-site-spec]]). Virtual Collaborators "carry the access rights, system connections, and context that their organizations grant them" [S-1]. Production engineering is the stated constraint.
- Use it as evidence inside an original re-frame, not as the topic. A "what the S-1 implies for your data team" piece is in-lane, buyer-facing and contains no client detail.
- Do not mention unannounced phData-Anthropic arrangements publicly.
(c) Investing: considerations, not a recommendation.
- Getting IPO shares: I found no directed-share program in the S-1 (a text search, so absence is not proven). A European retail tranche exists [S-1]. US retail allocations at mega-IPOs are usually small and broker-gated [Ray].
- Lockups: the 180-day lockup has staggered early releases (day 1, 27, 48, 62, 76; more tied to the FY26 and Q1-2027 earnings releases) plus an extended 365-day tranche [S-1 p.308]. That means supply overhang arrives in steps.
- Indirect exposure: AMZN (Amazon's $6.2B of notes convert to Class N; $110B compute deal), GOOGL (14.99% ownership cap; $10B concurrent placement option; TPU deals), MSFT, AVGO (Broadcom), AMD, NVDA (through SpaceXAI, Nscale and Lambda), SpaceX [S-1].
- For the memory capital-cycle thesis ([[2026-05-18-memory-cycle-v1.1]]): the commitments ramp from 2027 and run to 2036 (Amazon) with obligations through 2045, which supports the Phase 2 read. The weak points are the neocloud financing contingencies [Ray].
- What would need to be true to buy at about $2T: revenue keeps compounding without the release treadmill slowing, gross economics hold as commitments come online, and the governance discount is already in the price.
- This stays paper-only; no trades proposed.
Positioning verdict: well positioned on the right vendor, the right partner type, and two banked certs. The gaps:
- No named production deployment story he can publish.
- No Frontier Academy seat yet.
- The Denominator is not live yet ([[2026-09-29-nov30-credibility-targets]]).
Related
- [[2026-06-03-alphasignal-anthropic-ipo-filing]]
- [[2026-09-30-innermost-loop-devday-dots-anthropic-518b-si-accord]]
- [[2026-10-04-mostlymetrics-annualized-revenue-crimes]]
- [[2026-06-11-phdata-anthropic-partner-tier-cert-leverage]]
- [[2026-06-12-2026-2027-ipo-pipeline-lead-magnet]]
- [[2026-05-18-memory-cycle-v1.1]]
Addendum: Frontier Academy detail (founder follow-up, 2026-10-26 10:5x ET)
- S-1 (Business, about p.194): launched October 2026, $100M commitment, aims to train 10,000 engineers by end-2027, "beginning with cohorts from leading consulting firms and enterprises"; each participant leads "a Claude deployment at their own organization" [S-1].
- Press [report]: announced 2026-10-02. Participants are called "Frontier Deployed Engineers." Entry is by company nomination with a named Claude project. Format is in-person days on a simulated deployment, then a 12-week residency. Cohorts run in SF, NYC and London. First credentials are expected in early 2027. First-cohort firms: Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Commonwealth Bank of Australia, Novo Nordisk. Sources: mixed-news.com, techradar.com, resultsense.com (2026-10-05), edtechinnovationhub.com.
- [Ray] No boutique data/AI firm is named, so there is an opening for a Snowflake-native nominee. The founder's next step is internal: find out who owns the Anthropic relationship at phData. Offered to draft the nomination note.