How do two violently contradictory numbers live in the same half-year report? $16.3 billion and $80.7 billion. On October 26, 2026, Anthropic, PBC formally filed its Form S-1 registration statement with the U.S. Securities and Exchange Commission, moving its IPO into the registration stage: for the six months ended June 30, 2026, revenue reached $16.3 billion, roughly 13 times the $1.2 billion recorded in the same period last year [Secondary/Computed]; the net loss over the same span was $80.7 billion [Secondary]. Between those two figures sits the single most misread footnote in the prospectus: about $72 billion of that loss comes from non-cash accounting adjustments tied to convertible notes and other financing instruments - it does not mean the company burned $80.7 billion in actual cash. Reading that ledger correctly is the first gate for deciding whether this AI company deserves your attention.
A note on scope: this piece anchors on Anthropic's official announcement and the S-1 cover page [Official]; financial figures come mainly from TechCrunch's two October 26 reports on the filing [Secondary]; some data points trace to WSJ and Reuters coverage of the draft prospectus leaked around October 20 [Secondary, draft basis]. Calibers are tagged inline throughout; unless noted, financials cover the six months ended 2026-06-30.
1. The Filing Basics: Ticker ANT, Terms Undetermined
The official language is deliberately restrained: Anthropic "intends to list its Class A common stock on the Nasdaq Global Select Market under the symbol 'ANT'" [Official]. One detail worth recording: several financial relay pieces rendered the ticker as "ANTP," which contradicts the official filing - the S-1 cover says ANT. When citing secondary coverage, that one-letter difference deserves a second look.
Just as undetermined: the number of shares, the price range, and therefore the total raise - the official text reads "have not yet been determined" [Official]. There is no listing date either. The registration statement has been filed but has not yet become effective; no shares can be sold until it does, and the announcement itself does not constitute an offer to sell. In other words, every "when will it list" claim circulating right now is media speculation, not a company commitment.
The underwriting syndicate, however, is already on display. The S-1 cover page lists 22 banks: Goldman Sachs and Morgan Stanley lead, followed by J.P. Morgan, Citigroup, Barclays, Wells Fargo, BofA, Deutsche Bank, UBS and the rest of Wall Street's mainstream institutions [Official]. Counsel on the company side is Wilson Sonsini, with Freshfields also listed [Official]. The sheer size of that syndicate is itself a signal: under the draft-stage framework of "up to roughly $100 billion raised" reported in August [Secondary], this would rank among the largest IPOs in history - a deal half of Wall Street has to catch.
2. The Ledger: Three Loss Figures You Must Not Mix
The most professional - and most easily botched - part of this prospectus is that "loss" comes in three different calibers.
First, the operating caliber: the operating loss for H1 2026 was $693 million, narrowing sharply from $2.1 billion in H1 2025 [Secondary]. Read alone, this is the healthiest box in the ledger: the core business is converging toward break-even.
Second, the net loss caliber: a net loss of $80.7 billion for the half, of which roughly $72 billion stems from non-cash accounting adjustments on convertible notes and other financing instruments [Secondary; draft-basis reporting matches]. These adjustments arise from fair-value swings in convertible instruments - they are book entries, not cash out the door during the period. So "Anthropic lost $80 billion in six months" is a scary but inaccurate headline; the real operating bleed is far smaller. The comparatives: a net loss of about $13 billion in H1 2025 [Secondary, draft basis]; $42 billion for full-year 2025 per Reuters' September exclusive [Secondary]; full-year 2025 revenue of $4.6 billion [Secondary].
Third, the non-GAAP caliber: per the FT, Q3 alone brought $16.8 billion in revenue - a single quarter exceeding the entire first half - and a second consecutive quarter of "adjusted profitability" [Secondary]. That is a non-GAAP, adjusted figure and cannot be compared directly against the $80.7 billion GAAP net loss; any article mixing them without saying so is misleading you.
Pieced together, the three calibers form one picture: operations are narrowing, accounting noise is amplified by financing instruments, and the adjusted view already shows flickers of profit. But the prospectus pours its own cold water, stating that the company "may not achieve or sustain profitability in the future" [Secondary, quoting the S-1]. The most honest sentence in a registration statement usually hides in the risk factors.
3. The Compute Bill: $9B in Six Months, $450B Committed
The other side of 13x revenue growth is ferocious cost-side transfusion. Compute and infrastructure spending approached $9 billion in H1 2026, exceeding the $7.3 billion spent across all of 2025 [Secondary, draft basis] - half a year burned more than all of last year.
The forward-looking commitments are fiercer still: the S-1 discloses roughly $450 billion in signed future compute commitments over the coming decade - Google Cloud $111 billion, AWS $110 billion, Microsoft $31 billion - rising to as much as $750 billion if pending deals close [Secondary, draft basis]. One source conflict to flag: some Chinese relay reports put signed commitments "above $600 billion," which differs from the WSJ's caliber of "roughly $450 billion signed, up to $750 billion including pending deals." This article follows the WSJ caliber and presents both for readers to weigh.
Read the $450 billion backwards: these are purchase obligations spread over ten years, not money already spent, but they form a rigid, debt-like constraint on the future. If revenue does not climb at the slope the prospectus projects, the moat becomes a shackle. Compute suppliers are simultaneously Anthropic's shareholders and its channels - and after this filing, that "investment-for-compute" entanglement is formally a public-market risk disclosure rather than an open secret.
4. The Stickiness Story: What the Prospectus Most Wants You to Believe
If the financials are the ledger, the strategy chapter is the narrative. TechCrunch's second filing-day story dissects the S-1's enterprise-stickiness playbook, the most differentiated material in this filing.
Customer layering first (as of 2026-09-30): more than 11,900 customers spend at least $100,000 a year; over 1,900 of them spend $1 million or more; roughly 250 spend $10 million or more [Secondary]. Retention next: net revenue retention of roughly 900% - a cohort that spent one dollar last year spends nine this year [Secondary]. That number is rare by any enterprise-software standard; even AWS in its golden years never printed anything like it.
The mechanism is laid open in the filing: among customers spending $100,000 or more, more than half now use both the Claude platform and first-party applications (Claude Code, Cowork), up from about 30% at the end of last year [Secondary]. The pay-amplification effect is starker still: in Q3, the top 5% of heavy Claude Code enterprise users spent about 20 times more per month than ordinary enterprise users who do not use coding tools [Secondary]. The deeper the tooling habit, the faster the invoice grows - developer habit converted into a revenue lever.
The S-1 frames this as a three-layer flywheel - models, developer platform, applications - in which "an advance at any layer compounds across all three layers" [Secondary, quoting the S-1], explicitly benchmarked against the cross-product ecosystem plays of AWS and Uber. Two execution mechanisms follow: an internal Applied AI team embeds Claude into customers' existing technology stacks, and a partnership with private-equity firms including Blackstone ("Ode with Anthropic") places forward-deployed engineers inside enterprises to build custom tooling [Secondary]. On channels: AWS, Google Cloud and Azure act as joint go-to-market partners, but Anthropic directly holds the customer contracts and pays the cloud providers - by its own account, to avoid channel conflict and keep pricing power [Secondary].
Yet the prospectus simultaneously discloses a structural risk: cloud partners increasingly view Anthropic as a competitor and may favor their own or their portfolio companies' rival models, potentially constraining Anthropic's access to compute [Secondary]. Channel and adversary in one body - the most contorted joint of this business model, and the company is not hiding it.
The most ambitious line in the strategy chapter: the company intends to "reinvent our business," launching "new classes of products whose aim is to dwarf the economic value of our existing ones," while devoting "less effort" to "entrenching our current products and business models" [Secondary, quoting the S-1]. In plain terms: we may disrupt our own products at any moment, so do not value us with a traditional company's moat logic.
5. Governance: Those Who Pay Do Not Decide
The most controversial section is governance. Seven co-founders - Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah - through a "Founder LLC" will collectively control 50.1% of voting power after listing [Secondary], officially justified as protecting the AI-safety mission. The share structure is multi-class: Class A carries one vote per share (the public offering), Class T can elect a majority of directors, Class F holds 50.1% of voting power and is non-transferable, and Class N carries no votes [Official, S-1 cover]. A Long-Term Benefit Trust, established in 2023, holds the right to appoint a majority of the board [Secondary].
TechCrunch's verdict is blunt (media commentary, not this article's position): economic stakes and voting power are badly mismatched, and public shareholders will hold virtually no real control over a company that itself acknowledges its technology may pose "catastrophic or existential risks" [Secondary]. Buyers of ANT acquire an economic claim, not a voice - a premise to settle before placing any order.
6. Cooling Valuation and Three Risks
The $2 trillion outlook from August has met a real-world correction: Reuters reports that at the investor day, after seeing the latest financial disclosures, some investors considered marking valuation expectations down to about $1.5 trillion, versus the roughly $2 trillion previously attached to a November offering [Secondary, draft basis; all "reportedly"]. From pitching the pie to auditing the books, $500 billion of expectation evaporated in four months.
One more precision point: the "$30 trillion TAM" that circulated during the draft stage was not repeated in either TechCrunch report on the formal filing, and we could not directly verify the TAM language in the S-1 body - treat the S-1 text as final and do not mistake a draft-stage pitch for a prospectus reaffirmation.
At least three layers of risk, all disclosed in its own filing. First, the timetable: officially only "proposed," with no listing date; "within weeks" is a WSJ podcast claim and "a November share sale" is a Reuters one - both draft-period reporting [Reported]. Second, customer concentration: about one-fifth of revenue comes from the two largest customers, with the S-1 warning that "any adverse change in demand from our largest customers" could adversely affect its business and prospects [Secondary, quoting the S-1]. Third, the cloud-partner antagonization described above. None of this is bearish spin; it is the company's own due disclosure.
On division of labor: this article pairs with our August 26 piece on the $2 trillion valuation and the $30 trillion TAM pitch - that one, built on WSJ draft reporting, handled the vision; its projected "September prospectus disclosure" ultimately slipped to the actual filing on October 26. This one, built on the formal filing, handles the S-1 ledger and strategy teardown. The same batch also includes an open-source teardown of Tencent's Hy-Nomas interactive world-generation model and an SOP on the LibTV and WorkBuddy video factory, covering the open-source and production-tooling threads respectively; see related reading below.
One-sentence closer: when $72 billion of accounting adjustments decide an $80.7 billion headline loss, and $450 billion of compute commitments lock in the decade's ceiling, the real question this prospectus answers is not "how much can Anthropic earn," but "will Wall Street front ten years of compute bills for a business that cannot yet fund itself."
Sources
- [Official] Anthropic announcement, "Anthropic, PBC files registration statement on Form S-1" (2026-10-26): https://www.anthropic.com/news/s1-sec
- [Official] SEC EDGAR, Anthropic PBC Form S-1 cover page (2026-10-26): https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm
- [Secondary] TechCrunch (Rebecca Bellan, 2026-10-26), "Anthropic officially files for its public debut"
- [Secondary] TechCrunch (Rebecca Bellan, 2026-10-26), "Anthropic's IPO filing reveals its strategy for keeping enterprise customers hooked"
- [Secondary] WSJ and Reuters reporting on the draft prospectus (2026-10-20/21, as relayed by Yahoo Finance and Reference News)
- [Secondary] NetEase and Communication World relaying WSJ and FT (2026-10-21): https://www.163.com/dy/article/L9OLPFMU051288FS.html
Related reading
- Anthropic's $2 Trillion IPO Run Starts With a $30 Trillion Pitch to Wall Street (the August piece drew the pie; this one audits the books - the gap between $2 trillion and $1.5 trillion lands between the two)
- Tencent Hy-Nomas: An Interactive World-Generation Model, Open-Sourced and Reviewed (same batch, open source)
- The LibTV x WorkBuddy Video Factory SOP (same batch, SOP)