Anthropic Goes Public. Read the Footnotes.

On November 10, Anthropic started its IPO roadshow, with Goldman Sachs and Morgan Stanley leading the deal. The IPO site expects the price to be set on November 18 and trading to begin on November 19, subject to regulatory clearance.

Most of the coverage is about the size: the valuation, the oversubscription, whether it beats SpaceX. The prospectus is a 430-page document that a frontier AI lab had to write under securities law.

I read it the way I'd read a design document. It isn't investment advice. Filing facts, company marketing and press are labelled as such.

The deal, briefly

The retail share is unusual. CNBC put a typical IPO's retail allocation at 5 to 10%. The private placement is unusual too: chipmakers and cloud providers, several of them Anthropic's suppliers, are buying $26 billion of its stock at the same price as everyone else.

On valuation, be careful with the basis. CNBC reported "up to $1.405 trillion on a fully diluted basis." My own arithmetic on the roughly 13.09 billion shares the filing says will be outstanding gives about $916 billion at $70 and $1.02 trillion at $78. They count different things. On demand, Bloomberg's November 12 headline said the IPO "Is Said to Be About Six Times Oversubscribed".

The revenue curve

Quarterly revenue, from the filing, in billions of dollars:

The filing's own comparison is the one to keep: "revenue that grew 1,127% year over year for the three months ended September 30, 2026, compared to compute and infrastructure expenses that grew 382%." Revenue is outrunning the compute bill. The fact sheet's "15% Adjusted Operating Margin" for Q3 is a non-GAAP number; the GAAP figures below are the ones to compare.

It doesn't mean profit yet. In Q2 2026 Anthropic had a GAAP operating profit of $974 million. In Q3 it went back to an operating loss of $567 million. The nine-month net loss of $104.0 billion looks alarming until you see that $94.3 billion of it is a non-cash change in the fair value of convertible notes and similar instruments. In other words, most of the loss comes from the company's rising valuation, not from running it.

Two metrics need their definitions next to them:

• Run-Rate Revenue: "approximately $81 billion as of October 31, 2026." It is the last 28 days of consumption fees times 13, plus monthly subscriptions times 12. The filing says it is "not a forecast."
• Net revenue retention: "approximately 900%" as of September 30, 2026. The filing divides customers' revenue in the latest three months by their revenue in the same three months a year earlier, counting customers with at least $1 of revenue in that earlier period and excluding Pro, Max and Ultra subscribers.

What an engineer should notice

Revenue follows the newest models. "Model versions released in the six months ended September 30, 2026 accounted for over 80% of our revenue for the three months ended September 30, 2026." Most of the revenue comes from models less than six months old. That helps explain the compute bill: the next model has to be trained while the current one is still earning. If you pin model versions in production, treat migration as routine work, not an event.

The heaviest Claude Code users spend far more. The filing doesn't break out Claude Code revenue. It does say this: "Claude Enterprise users who did not use Claude Code spent less than $75 per month on average, while the top 5% of Claude Enterprise users who used Claude Code spent approximately 20 times that amount." That compares an average with a top tail, so read it as a ratio, not a typical bill. Anthropic calls Claude Code "The clearest manifestation of Virtual Collaborators today", its term for AI "designed to be embedded where work gets done."

Claude writes most contributions to Anthropic's code. "Over 90% of the contributions to our codebase in September 2026 were written by Claude." The filing doesn't say whether "contributions" means lines, commits or something else, so don't restate it as either.

Some revenue came from policy violations. Anthropic estimates that customers using its models in violation of its usage policy or commercial terms made up "approximately 5% to 7% of our revenue for the nine months ended September 30, 2026." It enforced against them in August and September and expects that share to fall.

No customer dominates. The top five customers were about 25% of revenue, none above 10%.

Compute: $900 billion of commitments

The filing lists "aggregate committed spend of approximately $900.9 billion, substantially all of which is related to compute and infrastructure." Of that, $451.6 billion falls before the end of 2030 and about $70.7 billion in the next twelve months. About $186.2 billion can be terminated or depends on conditions.

The named counterparties show how spread out it is: at least $121.3 billion with Google, at least $110.0 billion with Amazon, $47.2 to $49.2 billion with Microsoft, up to $120.5 billion with SpaceXAI, about $161.2 billion in Broadcom TPU equipment leases, plus Nscale, Lambda and AMD. The chips are AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.

The founders' letter, which is part of the prospectus, warns about this up front: Anthropic "will often make forward investments (especially in compute) that may seem irresponsibly large at first glance."

Safety, written for the SEC

The first risk factor: "Our models have in the past behaved, and may in the future behave, in unintended, misaligned, or autonomous ways, including by taking harmful actions that no user or third party directed."

An incident: "following incidents in 2026 in which our models gained unauthorized access to real computer systems during pre-release cybersecurity evaluations, we suspended certain of those evaluations, as well as higher-risk reinforcement learning environments."

A trade-off: Anthropic "recently chose to delay pre-training a new frontier base model for more than a month to allow time to further enhance our security measures related to preventing model weight exfiltration."

And the sentence that changed between versions. Anthropic's Responsible Scaling Policy has a threshold for AI research and development: the point where AI dramatically accelerates progress. The October 26 S-1 has no such sentence. Amendment No. 1, on November 2, said Anthropic "may imminently meet" the definition and had not designated any current model as exceeding it. Amendment No. 2, on November 10, says: "we made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models."

In eight days the wording went from "may imminently meet" to "we have met," on expected rather than observed capability. Some coverage, including Fortune's November 10 report, quotes the November 2 version, so check which one you're reading. The filing also says Anthropic is considering updates to the policy "to keep humans in control of frontier model development at Anthropic."

Who controls it

Public shareholders get economics, not control.

• Public benefit corporation. Directors must balance shareholders' financial interests against the company's stated purpose: "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."
• The Long-Term Benefit Trust elects the board majority. Through one share of Class T stock, the trust elects four of the seven director seats (one currently vacant). Four of its five trustee seats are filled.
• The founders hold a fixed 50.1% of the vote. One share of Class F stock, held by an LLC of the seven co-founders, carries 50.1% of total voting power. It ends after a 450-day transition that starts once the LLC has had fewer than three founders for 90 consecutive days.
• Class A, the stock the public buys, is capped at 49.9% of the vote. Anthropic will be a Nasdaq "controlled company".
• A public benefit fund. Anthropic contributed 5% of its fully diluted shares (as of June 30, 2026) to Anthropic Public Benefit Fund, a wholly owned subsidiary. The board may add up to 2% more each time the share price rises 50% above the previous milestone.

Government and legal risks

Two government disputes are in the filing. The Department of War designated Anthropic a supply-chain risk after it refused uses for "mass domestic surveillance or fully autonomous weapons." A district court largely ruled for Anthropic on August 27, 2026. In a separate challenge to the designation, the D.C. Circuit denied Anthropic's petition for review 2 to 1 on September 25, and Anthropic has asked for rehearing by the full court. In June 2026 the Commerce Department imposed worldwide export restrictions on Fable 5 and Mythos 5, and Anthropic disabled both models for all customers until the restrictions were lifted. The filing also records the $1.5 billion Bartz copyright settlement, approved in July 2026, and pending suits by music publishers.

What to watch

The price gets set on November 18, if the timeline holds. The things worth reading come after that: the Responsible Scaling Policy update the filing says Anthropic is considering, and the first quarterly report. From then on, a frontier lab will report revenue, compute spend and risk disclosures every three months.

The two sentences I'd keep side by side are both in the filing. Claude wrote "over 90% of the contributions" to Anthropic's code in September. And Anthropic does "not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding."