Anthropic files S-1: an $80.7 billion first-half net loss, a $44.9 billion non-cancelable Microsoft commitment, and a BB-shaped rating expectation

The primary material is the S-1 body on SEC EDGAR, a fourteen-megabyte HTML document (anthropicpbc-sx1.htm) accepted on October 26, 2026 at 09:00:11, part of an index of 87 documents, with exhibits numbered from EX-3.1. The cover states that the offering is Class A common stock, that the company has applied to list on the Nasdaq Global Select Market under the symbol ANT, that the price range is blank and marked subject to completion, and that the risk factors section begins on page 26. These look like trivia but they bound the entire discussion: with no price range and no share count disclosed, any valuation derived from a share price is speculation.

The loss figures must be quoted exactly. The consolidated statements in the S-1 report net losses of $1.531 billion in 2023, $8.309 billion in 2024, $41.965 billion in 2025, $12.969 billion in the first half of 2025 and $80.748 billion in the first half of 2026, meaning the 2026 first-half loss is nearly twice the full-year 2025 figure. Over the same period, compute and infrastructure expense rose from $2.522 billion in 2024 to $7.326 billion in 2025, an increase of 190%. Read together, the widening loss and the compute build-out are two segments of the same curve.

Then there is the commitment. The infrastructure service arrangements section states that in November 2025 the company entered an agreement with Microsoft, amended in February and October 2026, committing to spend an aggregate of at least $44.9 billion between November 2026 and May 2033 regardless of actual usage; the commitment is non-cancelable except for Microsoft's uncured material breach or, on a per-tranche basis, for undelivered tranches. The same section discloses the agreement with Google and a Broadcom-related lease arrangement, including the ability to issue up to $42.0 billion in aggregate principal of convertible notes to Broadcom to finance obligations under that arrangement.

Regardless of actual usage is the hardest constraint in the document, because it means weak demand does not reduce the spend. Set that against customer concentration and the tension shows: the S-1 says no single customer represented more than 10% of revenue for the nine months ended September 30, 2026, and that about 79% of revenue came from outside the top five customers in the three months then ended, up from roughly 64% a year earlier. Revenue is diversifying while costs are concentrating, two very different risk shapes.

The risk factors section became news in its own right. The Verge counts a 335-page prospectus with 94 pages of risk factors and quotes the filing's language on catastrophic or existential risk. I checked the sentence in the original: the filing acknowledges that concerns have been raised within the organisation and by other researchers, regulators, policymakers and industry participants that sufficiently advanced AI systems could, if not developed, deployed and operated with appropriate safeguards, oversight and security measures, pose catastrophic or existential risks to humanity. A company writing that in a listing document is a disclosure choice it made, not one imposed from outside.

The rating thread comes from reporting and must be quoted carefully. TechCrunch's October 30 item relays the Financial Times: Blackstone, anchor investor in a recent $63 billion deal to finance Anthropic's chips, told investors in that deal it expects the AI lab to receive a high junk rating after its IPO, where previously Blackstone had said the opposite, expecting investment grade before 2027; the current expectation is the double-B range, the top tier of speculative grade. TechCrunch reached out to Anthropic and Blackstone and both declined to comment. This is a forecast, not an agency decision, and no agency has published one.

Why two letters deserve their own paragraph: a rating sets the price of borrowing. A company that can only reach speculative grade pays more on its bonds, and counterparties in secured or lease structures demand more protection. For a company committing large sums to compute, every step up in financing cost propagates through capex into training scale and, eventually, API pricing. That causal chain is my inference; the S-1 does not say it. But the sentence in the filing that the company anticipates making significant investments for the foreseeable future, including in compute and infrastructure, is the premise for the inference.

Line the S-1 up against the public narrative and a gap appears. Outside discussion of Anthropic is about model capability, safety philosophy and enterprise customers; inside the document the loudest items are three tables: losses, compute spend and non-cancelable commitments. The narrative still matters, but going public means switching from explaining what you believe to explaining what you owe. The line that the company may make forward investments that may seem irresponsibly large is the specimen of that switch: it reads simultaneously as a risk disclosure and as a self-introduction.

For developers and enterprise buyers the filing carries two practical implications. First, the longer the compute commitment, the more the model cadence is shaped by capital, which makes API pricing and quota policy look more like a cloud vendor's than a research lab's; this is not a price forecast, only a statement that the constraints have changed. Second, customer dispersion (no single customer above 10%, 79% outside the top five) gives the company an incentive to keep a broad mid-market base, which is good news for small teams on the API: large customers will not monopolise quota. Both are one step beyond the facts in the document, not conclusions it draws.

Boundaries, so headlines do not carry readers away. First, the price range, share count and valuation are blank and marked subject to completion, so any precise figure for what Anthropic is worth does not come from this document. Second, the BB rating is Blackstone's expectation as reported by the Financial Times; both companies declined to comment and no agency has published a decision. Third, some figures in the S-1, including certain revenue and Run-Rate Revenue values disclosed by date, are left blank in the body text, so this article quotes no revenue number that is not shown. Fourth, $44.9 billion is a floor on committed spend, not spend already incurred and not a cash balance. Fifth, the link I draw between financing cost and API pricing is my analysis, not company disclosure.