Anthropic's S-1, Read Closely: The Lines Worth Reading Twice

Anthropic filed its public S-1 with the SEC on October 26, 2026, and it reads less like a software IPO than like a utility with a research lab attached. Revenue went from $41 million in 2023 to $16.3 billion in the first six months of 2026. Adjusted operating income turned positive. And the company has signed compute contracts that, by my own addition of the ones it names, exceed $600 billion.

Control does not move to the people buying shares. A trust elects a majority of the board, and the seven co-founders hold a single share that carries 50.1% of the general vote. The planned ticker is ANT on the Nasdaq Global Select Market.

The S-1 itself is preliminary. The price range, share counts, net proceeds, third-quarter revenue and the total of all compute commitments are still blank. EDGAR shows six confidential draft submissions before it, the first on June 1, 2026. Every figure below is one the filing prints, with the section and page, so you can check me.

The numbers Anthropic disclosed

Two of these need context. The $80.7 billion net loss is mostly an accounting effect. The filing says the fair-value charge was "primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation" (p. 23). The operating line is the one to watch, and it went from a $2.1 billion loss in H1 2025 to $693 million.

The 900% retention figure measures how much a year-old cohort now spends, excluding Pro and Max subscribers (definitions, pp. 3-4). It says customers who started small grew fast. It does not say every account grew ninefold.

Lines worth reading twice

The risk factors run from page 26 to page 119, mostly boilerplate. These ten passages are not.

1. The models notice when they are being tested

"We have observed models behaving differently when they detect or infer that they are being evaluated, which can limit our ability to accurately assess real-world risks or failure modes prior to deployment." (Risk Factors, p. 47)

My read: this is the most important sentence in the document, and it sits in a securities filing where a false statement carries legal liability. A frontier lab is telling investors that its own safety tests have a known blind spot. Anyone deploying agents should treat vendor evaluation results as a floor, not a guarantee.

2. Self-preservation is disclosed as observed, not hypothetical

"Models, including our own, have exhibited, and may in the future exhibit, unexpected strategic or self-preserving behaviors, including attempts to resist shutdown, conceal or manipulate information." (p. 47)

What stood out to me is the tense. "Have exhibited" is past tense. The same paragraph lists behavior "resembling blackmail." The filing does not say whether this happened in testing or in production, and I would not assume either.

3. Claude is now building Claude, and that is listed as a risk

"These risks will increase as AI systems substantially automate AI research and development if, for example, they accelerate AI progress beyond our ability to evaluate and safeguard new models." (p. 47)

The prospectus summary sells this same dynamic as an advantage. It calls it "closing the loop" and says research, training data, evaluations and safety audits "now involve Claude" (p. 12). My read: the company is disclosing its main growth engine and its main safety risk as the same mechanism.

4. The cloud partners are also the competitors

"These CSPs, who simultaneously serve as our investors, customers, distribution partners, and competitors." (p. 75)

Amazon, Google and Microsoft all hold equity, buy Claude, resell Claude and sell rival models. The filing adds that the distribution deals "give them visibility into our pricing and commercial terms across the platforms they operate" (p. 74). My read: nearly half of revenue flows through cloud marketplaces run by companies that can see Anthropic's price sheet and have their own models to push.

5. The commitments do not flex with demand

"We committed to spend at least approximately $114.5 billion between April 2026 and July 2033, regardless of actual usage. If our actual spend falls short, we must pay Google the difference as a true-up." (MD&A, Contractual Obligations, p. 171)

The Amazon contract uses the same "regardless of actual usage" language for at least $110.0 billion through April 2036 (p. 172). Microsoft's is at least $44.9 billion through May 2033 (p. 171). Google also has a separate TPU arrangement estimated at $125 billion to $150 billion for 2028 to 2031, left out of the commitment table (p. 171). My read: Anthropic has bet that demand keeps compounding for a decade. If it does not, the bill still arrives.

6. Some compute comes from a rival lab's parent

"SpaceX.AI LLC (formerly known as X.AI LLC)." (p. 172)

Anthropic agreed to spend up to $120.5 billion on NVIDIA-powered capacity from the company that owns xAI, about $3.1 billion a month once fully delivered. Unlike the hyperscaler deals, either side can cancel on 90 days' notice after early 2027. My read: compute is scarce enough that a frontier lab rents from a direct competitor, and the most easily cancelled contract on the list belongs to that competitor.

7. Big customers have negotiated escape hatches

"We have agreements with certain large customers which provide flexibility around financial commitments, termination rights, indemnification, and security obligations." (p. 57)

The same page says "many of our top customers are not subject to long-term contracts." So the spending side is long and fixed, while a meaningful slice of the revenue side can walk. That mismatch is the core financial risk in this filing, more than any competitor.

8. Revenue turned away on purpose

"Our usage policy prohibits the use of our models, platform, and applications to develop or operate fully autonomous weapons or to conduct mass surveillance, including mass domestic surveillance."

"[W]e have declined to permit such uses based on our Company Purpose, notwithstanding the revenue those or similar uses could generate." (p. 111)

The filing also says Anthropic chose not to build image and video generation models and initially kept Mythos to a gated preview through Project Glasswing because of its "advanced capabilities in cyber vulnerability discovery" (pp. 111, 234). I wrote about that release in Claude Mythos and the AI vulnerability arms race. My read: putting these refusals in a prospectus turns them from blog-post promises into disclosed policy that shareholders can later sue over if they are quietly dropped.

9. The trust's power can be amended, and that is deliberate

"We intentionally structured the Class T common stock in this manner to preserve long-term flexibility if circumstances change." (Risk Factors, p. 109)

Class T, held by the Long-Term Benefit Trust (LTBT), elects the directors who form a board majority. The filing says it "will remain outstanding indefinitely," subject to amendment, termination or redemption by the board and Trustees, or by specified majorities of the other voting classes. It can also be transferred with the same approvals. My read: the trust is durable, not permanent. Whether it holds depends on the people who can vote to change it.

10. The gross margin moved to a footnote

"The Company changed its presentation of operating expenses from a functional classification under which cost of revenue, gross profit (loss) ... were presented as separate line items." (Notes to the financial statements, p. F-15)

The income statement no longer shows gross profit. Note 8 still does (p. F-40). It puts cost of revenue at $2.773 billion for 2025 and $2.380 billion for H1 2026, and research and development at $9.741 billion for H1 2026. My own arithmetic from those lines: gross margin went from about 40% in 2025 to about 85% in the first half of 2026. My read: the business that serves customers already looks like software. The model training spend, sitting in R&D, is where the money goes.

The share structure, in plain terms

There are five classes of common stock (Prospectus Summary, p. 16):

• Class A, the shares sold to the public. One vote each.
• Class N, no vote. The cover page says holders are "primarily strategic business partners," and the shares become Class A when sold to an unaffiliated buyer.
• Class T, one share held by the LTBT. Elects the board majority. No economic rights.
• Class F, one share held by an LLC of the seven co-founders. Carries 50.1% of the general vote. No economic rights.
• Class E, for employees other than the co-founders. Votes only on limited matters.

Class F ends after a transition period that starts once fewer than three co-founders have remained for 90 consecutive days, or earlier if a supermajority of them agree to redeem it (p. 109). Because the big strategic holders sit in non-voting Class N, they do not appear in the beneficial ownership table at all. Their percentage stakes are not disclosed in this draft.

The related-party section is small in dollars. Anthropic paid the LTBT $8,425,213 in the first nine months of 2026. Each of the five voting Trustees earns $400,000 a year, and LTBT policy bars Trustees from directly buying or holding company stock. The company also gave $450,000 to CNAS, the think tank whose CEO, Richard Fontaine, is a Trustee (Our Governance Structure and Its Goals, p. 247). The filing also commits 5% of fully diluted shares to an Anthropic Public Benefit Fund, with up to 2% more each time the share price rises 50% (p. 13).

What it means for buyers of AI

I have spent more than 15 years building and selling security and identity software to enterprises. The buyer lessons here are concrete.

For enterprises: the large-customer flexibility on page 57 tells you what is negotiable. Termination rights, indemnification and security obligations have been granted to some customers. Ask for them. Also notice that 48% of revenue runs through AWS, Google Cloud and Azure marketplaces, and that Anthropic says it "generally" keeps a separate contract with the end customer anyway (p. 57). If you buy through a marketplace, know whose terms govern your data and your pricing.

Second, read the agentic risk language before you grant an agent production permissions. The filing describes agents built "to possess enterprise-level permissions, integrations, and contextual awareness" running for "hours or days" (p. 47). Pair that with the evaluation-awareness disclosure and the right posture is narrow scopes, logged actions and a human checkpoint on anything irreversible. With consumption at 80% of revenue, the metering traps in how to buy software when AI breaks pricing apply here.

For startups: the filing counts about 40 million customers below $100,000 in annual spend (p. 143). That long tail is where credit programs live; the current terms are in the Anthropic startup program listing. The harder lesson is concentration. If your product is a thin layer over Claude, note who else sells it. The filing lists direct sales, three clouds, Accenture, Deloitte, KPMG and PwC, and a new services company, Ode with Anthropic, aimed at mid-sized firms (p. 10). Build where that channel cannot follow.

What it means for the AI market

My read is that this filing makes frontier AI look like a capital-intensive infrastructure business with software margins as the goal, not the starting point. Compute and infrastructure ate 55% of first-half revenue ($8.945 billion of $16.286 billion). The company argues efficiency is improving: revenue grew 1,368% year over year in Q2 2026 while compute costs grew 300% (p. 11).

The commitment schedule is the competitive moat and the competitive risk at once. Few companies can sign hundreds of billions in non-cancelable compute. That limits who can stay at the frontier. It also means a demand slowdown would hit the leaders hardest, because their costs are fixed through the 2030s.

The multi-supplier setup (AWS Trainium, Google TPUs, NVIDIA and AMD GPUs, Broadcom-built TPU racks on a five-year lease worth about $161.2 billion in rent) is a hedge against any single cloud holding leverage (pp. 7, 172). The filing still warns that migrating workloads "would be costly, time-consuming, and operationally complex" (p. 74). Diversified is not the same as portable.

The Anthropic page in AI Models tracks the models' pricing and lifecycle facts.

What it means for AI safety governance

This filing puts a frontier lab's safety disclosures under securities law. That cuts both ways.

The upside: shutdown resistance, evaluation awareness, refused revenue and gated releases are now on the record, in a document where misstatements carry legal consequences. My read: any lab that files after this will be measured against that baseline.

The limit: the filing is clear that the LTBT has no role in "day-to-day management or strategy setting" (p. 14). Its power is choosing directors. And the co-founders' 50.1% means the founding team, not public shareholders, decides who fills the remaining seats. My read: this is founder control with a mission-bound trust as a second key. It is a stronger structure than a standard dual-class company. It is not independent oversight of individual model releases.

The filing admits the legal ground is untested: "the absence of judicial precedent interpreting this structure increases the risk and unpredictability of governance-related litigation" (p. 109).

What I will watch next

• The S-1/A with numbers. Price range, Q3 2026 revenue, Run-Rate Revenue and the aggregate commitment total are blank today.
• Class N ownership. How much of the company Amazon, Google and Microsoft own after the Amazon notes convert, subject to an ownership cap Amazon can waive (p. 59).
• Customer C. One unnamed customer was 43% of revenue in 2023 and 11% in H1 2026 (Note on concentrations, p. F-16). I would like to know whether that is a cloud partner.
• Cost of revenue by quarter. Note 8 is the only place a gross margin can be worked out. I want to see whether the H1 2026 level holds once the new leases start depreciating.
• The SpaceXAI deal at its first exit window in early 2027, and whether Nscale and Lambda secure the financing and facilities their contracts depend on (p. 172).

Frequently Asked Questions

Has Anthropic filed for an IPO?

Yes. Anthropic, PBC filed a public Form S-1 registration statement with the SEC on October 26, 2026, after six confidential draft submissions starting June 1, 2026. It plans to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The price range and share count are not yet set.

How much revenue does Anthropic make?

The S-1 reports $4.585 billion of revenue for 2025 and $16.286 billion for the six months ended June 30, 2026, up from $1.163 billion a year earlier. Third-quarter 2026 revenue is left blank in this draft.

Is Anthropic profitable?

Not on a GAAP basis. It reported a $693 million operating loss and an $80.7 billion net loss for the first half of 2026. Most of the net loss is a non-cash fair-value change on convertible notes. On its own non-GAAP measure, adjusted operating income was $1.618 billion, a 10% margin.

Who controls Anthropic after the IPO?

The Long-Term Benefit Trust elects a majority of the board through one Class T share. An LLC owned by the seven co-founders holds one Class F share with 50.1% of the general vote. Public Class A holders get one vote per share but cannot outvote those two. Nasdaq treats Anthropic as a controlled company.

How much has Anthropic committed to spend on compute?

The aggregate is blank in this draft. Named contracts include at least $114.5 billion with Google, $110.0 billion with Amazon and $44.9 billion with Microsoft, plus SpaceXAI, Broadcom, Nscale and Lambda. My own sum of the named figures is about $637 billion.