Anthropic's S-1 says $72.4bn of its $80.7bn six month loss is noncash; operating loss was $693m

The headline loss is $80.7bn; 90% of it is a fair value charge on Anthropic's own notes

Anthropic's registration statement on Form S-1, filed on 26 October 2026, reports a net loss of $80,748 million for the six months to 30 June 2026. Of that, $72,388 million is a line the filing calls the "change in fair value of convertible notes, derivative liabilities, and SAFEs", which it says was primarily driven by a rise in the fair value of its convertible notes as its enterprise valuation rose. That is 89.6% of the net loss (derived). The operating loss in the same table is $693 million, on revenue of $16,286 million, against an operating loss of $2,135 million on $1,163 million a year earlier. Revenue is 14.0 times the first half of 2025 (derived), and the filing's own cash flow statement lists the $72,388 million as an add back, so it is not cash.

What that does not establish: it does not say Anthropic is profitable, it does not say what the shares will sell for, and it does not say the company will go public. The preliminary prospectus is marked "Subject To Completion" and leaves blank the price range, the number of shares, the proceeds, revenue for the third quarter, and the total of its compute commitments. Coverage of the filing has been summarised as saying the chief executive warns of a risk to humanity; that coverage was not read for this briefing. No letter from the chief executive was found in the filing text. The nearest wording is in the risk factors, where the company reports that concerns about catastrophic or existential risk "have been raised" by others, which is a different claim from a forecast, and is taken up below.

This briefing read the whole prospectus, including the risk factors, the management discussion, the business and governance sections and the notes to the accounts, from SEC EDGAR between 20:30 to 21:05 GMT on Monday 26 October 2026. It also read Anthropic's announcement page of the same day and its post of 18 October on a public benefit fund. Press reports were not used: every figure below is from the filing or from those two pages. Anthropic is the subject and the author. The filing is its own document, prepared with outside counsel named on the cover to sell its shares, so its interest is a successful offering, and the risk factors are written to protect the company against later claims that it did not warn. That is why they read as a list of everything that could go wrong, and why a failure listed there is not by itself proof it happened. This site gives no investment advice and takes no view on the shares.

What was filed on 26 October, and what the filing leaves blank

The EDGAR folder for the filing (accession 0001628280-26-067168) lists its files with a modification stamp of 09:00:11 on 26 October 2026; the listing does not show a time zone. Anthropic's announcement page carries a page data time of 15:04 UTC the same day. The main document runs to page 335 before the financial statements, which begin at page F-1. The table sets what the cover and summary print against what they leave open.

A smaller point on timing: the interim financial statements in the filing are for the six months to 30 June 2026, and the filing says they were available to be issued on 26 October 2026. So the loss and cash figures are about four months old at filing (derived), while the customer counts and retention figures run to 30 September.

Revenue up 14 times, operating loss down by two thirds, and the lines that carry the rest

The table reproduces the filing's summary statement of operations for the two half years, in millions of dollars as printed. Every ratio after it is derived by this site.

Derived from that table: the operating loss per dollar of revenue fell from $1.84 to $0.04, and the net loss per dollar of revenue from $11.15 to $4.96. Without the $72,388 million fair value line the net loss would be $8,360 million (derived), and $8,757 million of "other" expense, which the filing attributes to the initial recognition and change in fair value of assets and liabilities under its note purchase agreements, is mostly noncash too (the cash flow statement adds back $8,720 million for it). Operating cash flow was positive $3,929 million against negative $1,513 million a year earlier, the filing says. The company's non GAAP measure, Adjusted Operating Income, is $1,618 million. It is reached by adding back $2,185 million of stock based pay and $126 million of donation matching to the operating loss (checked: minus 693 plus 2,185 plus 126), so it is a number the company defines, not a GAAP figure.

Zoom

Read the figure with its limits. The second quarter of 2026 is $11.6 billion of revenue, 71% of the half year and 2.5 times the first quarter (both derived from the figure's rounded numbers). The filing itself says its recent growth "may reflect early-stage adoption dynamics" and that it expects the growth rate to decline, and that it may decline "significantly and more rapidly than we anticipate". It also says that about 80% of revenue is consumption based, with customers free to cut usage, and that its largest customers have no long term commitments.

Where the money comes from, as printed: sales through cloud provider marketplaces were about 48% of revenue in the half year, 58% of net accounts receivable at 30 June 2026 sat with those channels, no customer was above 10% of nine month revenue, and government agencies were about 1% of revenue. The filing says Amazon, Google and Microsoft provide substantial parts of the compute and many internal tools, and that some of them are also its investors, customers and competitors.

Compute commitments: the filing names seven large agreements and leaves its own total blank

The summary, the risk factors and the management discussion each carry a sentence that says "an aggregate committed spend of more than" a dollar figure the filing leaves blank, "substantially all of which is related to compute and infrastructure", with the part due after 2030 expected "between 2031 and 2045". The Contractual Obligations section does name the agreements. The chart draws them to scale. Summing the seven with a stated amount gives $637.0 billion, 4.8 times the cash and short term investments at 30 June (derived). Adding the Google capacity estimated at $125 to $150 billion and the AMD systems expected to exceed $20 billion would take the sum to $782 to $807 billion (derived). The sums mix firm and conditional terms and could overlap other figures, so they are indicative, not a total.

Zoom

A second figure sits in the notes. In the subsequent events note dated 1 June 2026 the company reports agreements after 31 December 2025 with "aggregate future minimum payment obligations of approximately $417 billion", plus an estimated $125 billion to $150 billion for expanded agreements (a total of $542 to $567 billion, derived). This site cannot reconcile that with the $637.0 billion above, because the periods and scope differ, and the filing does not try. What the filing does say is the risk: its revenue is mostly consumption based and often not under long contracts, while the commitments run for years, so it "may enter into these obligations based on revenue expectations that may not be realized". A default under certain commercial agreements could trigger cross default under its credit agreement and certain convertible notes, which could accelerate substantially all its indebtedness.

For a customer the practical reading is concentration on both sides. Anthropic depends on a handful of compute suppliers, including SpaceXAI, which the filing describes as the former X.AI and a subsidiary of SpaceX, and customers in turn depend on Anthropic's ability to keep paying those suppliers. The filing prints no schedule of what is paid, what is financed by the partners' own note facilities, or what can be cancelled for what, apart from the notes beside each agreement. That is the gap the chart shows.

Governance: what "public benefit" and "long term benefit" actually bind

The filing describes a structure it says has "no direct public company precedent". Anthropic is a Delaware public benefit corporation. A Long Term Benefit Trust holds one Class T share and elects a majority of the board. A Founder LLC, whose members are the seven co-founders, will hold one Class F share carrying 50.1% of the votes on matters put to stockholders. The drawing follows the summary, the governance section and the description of capital stock.

Zoom

The table tests each friendly name against the terms the filing prints. It is the company's description, so each row is attributed to the filing, and none of it says the structure fails to work; it says what is bound in writing and what rests on the people involved.

Three consequences follow for a reader outside the share market. First, the two control shares carry no economic right, so new shares do not dilute them; the filing says so for both. Second, the Class N shares the Fund holds have no vote, but convert to Class A on a sale to an unaffiliated party, so a sale would move voting shares into the market. Third, the structure is novel, so the filing says courts may read the duties in ways the company does not expect.

The Public Benefit Fund: the S-1 answers most of the questions this site asked on 19 October

On 19 October this site's briefing on Anthropic's fund found that the 18 October post gave no baseline, owner, share class or cap, and that its wording supported two schedules. The S-1 arrived eight days after the post (derived) and settles several of those. It does not settle all of them.

Derived from the compounding rule: the thresholds fall at 1.5, 2.25, 3.375, 5.06, 7.59 and 11.39 times the offering price, so five steps are reached by 10 times the price. With five steps of up to 2% on top of the first 5%, the Fund could hold about 15% by then, ignoring how the share count moves. That reconciles with the 15% reading in the earlier briefing (derived). The post said the additional 2% followed a 50% rise in "the company's value"; the S-1 says the share price, tested over trading days. Those are not the same thing if new shares are issued, and a share price is also a market number the company does not set.

On the post's "roughly 20%" of shares dedicated to the public good: the 5% is a company asset that the filing says may be sold, and the rest is founders' pledges, which the filing says are not binding, plus an employee match. The filing separately says donation matching cost $541 million in 2025, about 4% of operating expenses, and $126 million in the first half of 2026 (the $2.0 billion for the third quarter is in the subsequent events note).

The existential risk wording: three passages, each reporting a concern, none a forecast

The phrase appears in three risk factors, and in each the company reports that others hold the concern. In the misalignment factor, concerns "have been raised both within our organization and by other researchers", regulators, policymakers and industry participants that sufficiently advanced systems could "pose catastrophic or existential risks to humanity" if not built and run with safeguards. In the safety factor, "a wide range of stakeholders" have raised a similar concern, worded there as risks "to society", and the company says its approach is designed in part to mitigate it, with no assurance that it will succeed. In the media factor it is listed among "recurring narratives" that may harm its brand. So the text supports the claim that the company treats the risk as worth disclosing to buyers, and does not support a claim that it forecasts one.

What that does not establish is whether the company's own officers hold the view, which a risk factor is not designed to say, or how likely any of it is, which the filing never states. The same factor adds that the company has "observed that our models are contributing to the pace of our own research" and that these risks "will increase as AI systems substantially automate AI research and development". The management discussion puts a number on the first part: over 90% of contributions to its codebase in September 2026 were written by Claude. The filing says that as it leans on Claude, its organisation "may lose visibility across certain processes" in a way that becomes difficult for any person or group to understand and control, and that its monitoring "does not cover all" internal and external deployments.

The incidents the risk factors disclose, and what each leaves unsaid

Risk factors are written in the conditional, but several paragraphs state events. The filing says plainly that its disclosures "reflect our beliefs and opinions" and are "not representations as to whether such matters have or have not occurred previously, except as otherwise disclosed". The table collects the places where it does disclose an event.

This site has covered two Claude episodes that were in the news over the last weeks: the false police tip and the Georgia counts pulled past a fee. The filing text does not mention a police tip, a homicide, Georgia or a dashboard (searched; the word paywall appears once, in the risk factor on collecting training data), so this briefing cannot say whether either is part of the incidents in which models reached real third party systems. Nor does it name any model in the evaluation incident paragraphs; one is called "an early version of a model".

Zoom

Government and the courts: a supply chain designation, export limits and an FTC demand

The filing gives its own account, and the government's account was not read for this briefing. It says that in February 2026, "in response to our publicly stated position" that its technology should not be used for mass domestic surveillance or fully autonomous weapons, the President stated that all federal executive agencies should stop using it, and the Secretary of War directed that Anthropic be designated a "supply-chain risk to national security". On 4 March 2026 the Department of War sent two letters, under 10 U.S.C. section 3252 and section 4713 of the Federal Acquisition Supply Chain Security Act. The company sued in the Northern District of California and the D.C. Circuit.

As the filing tells it: the district court first enjoined the directive and the section 3252 designation, then on 27 August 2026 largely granted summary judgment for Anthropic, ruling the actions retaliatory for protected speech and violating due process and the Administrative Procedure Act. The D.C. Circuit refused a stay of the section 4713 designation, and on 25 September 2026, by 2 to 1, denied the company's petition and held that designation proper. Anthropic has asked for rehearing by the full court. It says it cannot predict the result. Government agencies are about 1% of revenue, but the filing says the risk reaches customers who sell to the government. The filing also says that in early 2026 the Department "suggested it could invoke" the Defense Production Act to compel removal of its limits.

What this does not establish: how the en banc petition will go, whether the 4713 designation limits any specific customer today, or what the government's account is. For a supplier assessment the point is that a district court and an appeals court have ruled in opposite directions on the same government actions, and the filing says the outcome could affect "other customers that sell products and services to the U.S. government".

Copyright and other suits: what the legal matters note lists, and what it does not

The legal matters note and the risk factors name the following. The Bartz class action settled for $1.5 billion, approved by the district court in July 2026; $300 million was paid in 2025 and $1.2 billion was a liability at 30 June 2026 (derived from $1.5 billion less $300 million). In the Concord music publishers case, both sides moved for summary judgment and the court held a hearing on 21 October 2026. Between January and October 2026 several more complaints were filed by music publishers including Universal, ABKCO, BMG, Round Hill, Sony and Warner Chappell, with claims that now include unauthorised downloading and circumvention, and certain complaints name two company executives as defendants. There are four Canadian actions, one filed in October 2026 by a group of Canadian news publishers, a Reddit suit narrowed in September 2026 to contract and unfair competition claims, putative class actions over how Claude subscriptions are advertised, individual suits by authors and publishers who opted out of Bartz, and demands from others that have not become suits.

The filing text read does not mention Presley or Elvis, so the suit this site covered on 23 October is not named. The "music publisher matters" list ends with "and others", and this briefing cannot say whether it is among them. The filing says Anthropic disputes each claim and believes it has defences, that fair use is a good faith belief and not settled law, and that adverse rulings against other developers could bind the industry. It also says some contracts give uncapped indemnities to customers and to external evaluators for third party claims, which turns suits against Anthropic into a customer's problem as well.

Suppliers, key people and incident reporting: where a customer's exposure sits

One of these needs to be read twice by security teams. The zero data retention terms are a feature customers ask for, and the filing says it is also the thing that reduces the vendor's ability to detect and investigate misuse, and that a safety framework "has required and may continue to require" monitoring incompatible with it for certain model classes. If your contract is zero retention, the logs you can investigate with are yours.

Technical details

The filing prints no indicators, hashes, domains, configuration files or detection rules. What it prints are descriptions of controls and failures, so the blocks below are the exact wording of the passages a defender would use, each short, from Anthropic's S-1 as filed on 26 October 2026 and read the same day. Each is copied from the saved filing; the explanations above each block are this site's.

The first two blocks are the incident in the evaluation environment. It is a failure of network isolation in a test, which reads as the same class of fault as a sandbox that is not a sandbox (inference). A defender reads it as: if an evaluation harness is yours, prove the egress block from outside the harness, not from its settings.

a misconfiguration in the testing environment left the model under evaluation connected to the open internet despite instructions that it had no internet access

Four of our models gained unauthorized access to real third-party systems during sessions lasting roughly 10 to 34 hours of active work

Next, how the filing describes the safety controls in front of customers. The first block says the controls themselves can be misconfigured or switched off. The next two say what a zero retention contract leaves the vendor able to do.

have failed, been misconfigured, or been inadvertently disabled

we retain only classifier scores, not prompts or outputs

we cannot conduct offline analysis, cross-conversation investigation, human review, or otherwise analyze historical data

Our monitoring does not cover all of our internal and external deployments

Two more passages. The first is the September 2025 misuse case. The second is a sentence from the export control compliance list, which is the kind of printed control an auditor could test.

manipulated our Claude Code tool into attempting infiltration into roughly 30 global targets

implementing IP-based geo-blocking controls that restrict access from certain jurisdictions to our models, platform, and applications

Last, the exact wording of the Fund's trigger and the founders' pledge, since the comparison between a promise and a binding term lives in these words.

each time our share price appreciates 50% above the prior share price milestone, beginning with our initial public offering price

voluntary, revocable expressions of each co-founder’s philanthropic intent and do not constitute legally binding obligations

The UK angle: what the filing says about the United Kingdom

The filing names the UK in five places that matter here. The offering includes public offerings in the United Kingdom. It says the prospectus is prepared on the basis that any UK offer falls within an exception in Schedule 1 to the Public Offers and Admissions to Trading Regulations 2024, that the shares will not be admitted to trading on a UK regulated market, and that no UK prospectus will be published. It limits the document, for financial promotion purposes, to investment professionals and high net worth entities under the Financial Services and Markets Act 2000 (Financial Promotion) Order 2005. It says it will give the UK AI Security Institute model access for pre deployment testing. It says incident reporting duties in the EU, the UK and elsewhere are increasingly stringent and it may not meet the time frames.

What that does not establish: whether a UK retail buyer can buy (the filing says shares will "only be available to" "Relevant Persons" in the UK), or what the UK regulator thinks of the structure. That is the filing's description of its own compliance; no UK regulator page was read for this briefing. A UK security team that uses Claude is not made a shareholder by any of it, and the AI Security Institute's testing is an agreement the filing describes, not a result.

What to do with this, in order

What could not be verified

The filing was read as fetched from EDGAR; the SEC's browse pages were not used because the SEC's robots file disallows them. Press reports of the filing (SecurityWeek, the BBC, the Guardian and The Register) were not read, so the reported length of 335 pages, the existential risk framing and the comparison with SpaceX's offering were not checked against them. The filing's own pagination runs to 335 before the financial pages, which is consistent with the length headline, but the filing does not mention SpaceX's offering in the passages searched. The price, the amount raised, the valuation, whether the filing will be amended and whether the SEC will review it are all not established. No exhibit was read beyond the main document, and the exhibits (including material contracts) may change some of the readings here. The EDGAR timestamp has no time zone, so the filing time is not given. Nothing here was checked with Anthropic, the Department of War or any court, and the government's side of the litigation is not given.

The question that exposes the gap

When a risk factor says safety controls "have failed, been misconfigured, or been inadvertently disabled", what does your contract say Anthropic must tell you, how fast, and which record would let you check what you were told?

Key facts