Reading the Anthropic S-1

On October 26, 2026 Anthropic, PBC filed its registration statement with the SEC to list Class A shares on the Nasdaq under the ticker ANT. Underneath the 22 banks on the cover is the first audited view of a frontier AI lab, and it is a very unusual company. In one half-year its revenue grew fourteen-fold, it booked its first operating profit, and it reported an $80.7bn net loss. It has signed around $640bn of named compute contracts on a revenue base that is mostly billed month to month. It also has five classes of stock, and none of them lets a public shareholder control anything. This piece goes through the filing section by section, separates the accounting from the economics, and ends with a simple model of what a $1.5–2tn valuation has to assume. Every number comes from the prospectus unless it is marked as press reporting.

What the document is, and what it leaves blank

An S-1 is the long-form disclosure a company must publish before it can sell shares to the public. Anthropic filed confidentially on June 1, 2026, went through five confidential amendments, and made the filing public on October 26. Goldman Sachs, Morgan Stanley, J.P. Morgan and Citigroup lead a syndicate of 22 banks. The preliminary prospectus leaves every offering term blank: the price range, the share count, the proceeds, Q3 2026 revenue and the total of the compute commitments. Those figures arrive in an amended S-1/A just before the roadshow. Press reports put the target valuation at $1.5–2tn and the first trade around mid-November. Neither figure is in the document, so treat both as reporting.

The financial statements cover 2023–2025 (audited) and the first half of 2026 (unaudited). Operating metrics run to September 30, 2026. One technical detail matters for every per-share number: on September 17, 2026 the company did a 10-for-1 split of its Class A and Class N shares. The Series H round that closed in May–June 2026 at $589.01 per preferred share is therefore ≈$58.90 per post-split share.

1. The growth curve

Annual revenue went from $41m (2023) to $386m (2024) to $4.6bn (2025). The quarterly table below shows the shape of it: every quarter is larger than the last, and the last two are steep. Q2 2026 revenue of $11.6bn was 2.4× the previous quarter and 14.7× the same quarter a year earlier (+1,368%, the filing's own figure). Annualised, Q2 is a $46bn run-rate. Trailing twelve months to June were $19.7bn.

The chart also shows how the company turned profitable. Compute expense grew too (from $0.9bn to $5.6bn a quarter), but more slowly than revenue. Compute fell from 251% of revenue to 48% in six quarters, and Q2 2026 produced the company's first operating profit, $974m (8% margin; 19% on the company's “Adjusted” basis, which adds back stock compensation).

Three operating facts sit behind the curve:

• Customers. Customers paying $100k+ a year went from ~20 (end 2023) to ~200 (2024) to ~1,500 (2025) to 11,900+ by September 2026. 1,900+ pay over $1m and ~250 pay over $10m. More than half of the Fortune 500 is in the $100k+ group. Below them is a long tail of about 40 million smaller customers.
• Expansion. Net revenue retention is “approximately 900%”: a cohort of customers spends nine times what it spent a year ago. Strong software companies report 110–130%. It describes a market that is still filling up, not a steady state, and the filing says so: it expects the growth rate to decline, “and it may decline significantly.”
• Productivity. 5,300 employees; TTM revenue per employee of $6.6m (from $1.4m a year earlier). The company says Claude wrote over 90% of the code contributed to its own codebase in September 2026.

2. The unit economics: an 85% gross margin, with a caveat

The income statement groups costs by nature (compute, people, sales, other) rather than by function. That makes the gross margin hard to see on the face of the P&L. Note 8 to the financial statements gives the functional view, and it is the most informative table in the filing:

Serving customers (the compute that answers their requests, plus a little staff and overhead) cost $2.4bn against $16.3bn of revenue: an 85% gross margin, which is software-grade. A year earlier it was 12%, and in 2024 it was negative. Most of the compute spend is in R&D ($9.7bn, 60% of revenue), which covers training, research and the company's own internal use of Claude. The economics are therefore: inference is very profitable, and the profits go into the next model. The filing gives two reasons serving got cheaper: “compute multipliers” (algorithmic gains) and the ability to run on four chip families (AWS Trainium, Google TPUs, NVIDIA and AMD GPUs) and route each workload to the cheapest one. The prices customers see fall too: Claude Opus 5.5 cut cost per task by ~40% versus Opus 5, and the filing says this was done while “preserving our margins”.

The caveat is timing. Compute is contracted 12–18 months before it arrives, and it arrives in large blocks. When a block goes live, the expense starts at once, but the revenue it will support builds up over the following months. Margins are therefore at their best just before a large delivery, and Q2 2026 was such a moment. The filing warns that margins “may compress” as new capacity lands and as the company buys more expensive short-term compute to meet demand. The scale of what is about to land is visible in the contracts. Once fully delivered (expected December 2026), the SpaceXAI arrangement alone runs at about $3.1bn a month. That is more than the entire Q2 2026 compute line ($5.6bn a quarter, ≈$1.9bn a month). Expect the next few quarters to look worse than Q2 even if the business is healthy.

Two smaller points. Sales & marketing jumped to 23% of revenue on the functional basis (19% on the nature-of-expense basis), mostly fees paid to AWS, Google Cloud and Azure for reselling Claude through their marketplaces. And revenue is 80% consumption-based (pay per token) and 20% subscriptions. Section 5 explains why that split matters.

3. How a profitable quarter loses $66bn

Q2 2026 had $974m of operating income and a $66.3bn net loss. The full half-year loss of $80.7bn breaks down as follows:

The large line comes from how Amazon and Google funded the company. From 2023 Anthropic issued them convertible notes: loans that convert into shares at a fixed price. $6.2bn of Amazon's notes are still outstanding. They convert at $2.09 and $5.05 per share into ≈1.84bn non-voting shares, and the latest private round priced shares at ≈$58.90. Economically, a fixed-price conversion right is a call option the company wrote. When the share value rises ten- or twenty-fold, the option goes deep into the money, and accounting rules require the company to record the higher value of what it owes as a loss. So the faster Anthropic's valuation rises, the larger its reported loss. Of the $132.9bn accumulated deficit, $113.2bn is this mark.

No cash leaves the company, and the notes are expected to settle in shares at the IPO. But the expense is not imaginary either. It measures how much of the company early strategic backers received for their money, and that value now belongs to them and not to new shareholders. The useful profitability figures are the company's adjusted figures, which strip it out: Adjusted Net Income of +$2.7bn for H1 2026, and operating cash flow of +$3.9bn, the first positive half-year. One adjustment should not be ignored: stock-based compensation was $2.2bn in H1 2026 (13% of revenue). It is a real cost to shareholders, and it is why GAAP operating margin (−4% for the half) and Adjusted operating margin (+10%) differ.

4. The balance sheet and the financing web

At June 30, 2026 the company held $131.7bn in cash and short-term investments. It had raised $141.1bn in total, $110.8bn of it in H1 2026 alone, including the $45.7bn Series H. It also has access to more money, much of it from the same companies that sell it compute:

• $15bn revolving credit facility (Morgan Stanley agent, to 2031), with a covenant to keep at least $5bn of liquidity.
• $77bn of note-purchase capacity: Google $20bn, Amazon $15bn and Broadcom $42bn. Each unlocks when compute milestones are hit, and each converts into non-voting stock near the market price. Google also has an option to buy up to $10bn of stock at the IPO price.
• AMD commits to buy up to $5bn of stock in increments as its GPUs are delivered and deployed.

These suppliers are paid to deliver chips and capacity, and they are also investors who are paid back in Anthropic shares. The Broadcom facility is the most expensive of these arrangements. It exists solely to finance Anthropic's rent on Broadcom equipment, accrues 9% interest paid in kind (13% if revenue conditions are missed), and converts at a 10% (or 15%) discount to the market price. Amazon's and Google's stakes are capped at 12% and 14.99% respectively, unless they waive the caps. This is the vendor-financing pattern described in The Hyperscaler Debt Machine, seen from the borrower's side.

5. The compute book: long-term obligations, short-term revenue

This is the core of the filing. The aggregate commitment figure is blank, but the prospectus names each major arrangement:

The named contracts add up to ≈$637bn. With the estimated TPU and AMD capacity outside the table, the total is roughly $780–810bn, mostly payable over the next decade, with some running to 2045. The more useful number is the rigid core: the Google, Amazon and Microsoft take-or-pay commitments plus the Broadcom finance leases, ≈$431bn payable “regardless of actual usage”. Three ratios put that in context:

• ≈9 years of current revenue. $431bn ÷ $46bn run-rate ≈ 9.3.
• ≈3.3× the cash pile. $431bn ÷ $131.7bn.
• The durations don't match. The obligations run for 5–10 years. The revenue is 80% consumption-based, and the filing says “much of it is not subject to long-term contractual commitments”. Subscriptions are 12 months or less. More customers are signing annual or multi-year minimum-spend contracts, but the committed revenue is small next to the committed cost.

A bank that borrows long-term and lends short-term is exposed to a run. Anthropic's position is a version of the same problem: its costs are locked in for years and its revenue can change within a month. Two disclosures show how quickly revenue moves. Model versions released in the previous six months produced over 80% of Q3 2026 revenue, so revenue depends on staying at the frontier with each release. And customer spend fell ~20% over two weeks of the 2025 holidays when developers stopped committing code. The company does have some flexibility. The SpaceXAI capacity becomes cancelable on 90 days' notice in 2027. Several GPU deals charge only for delivered capacity. And “Anthropic Direct” (leasing chips into data centres run on its behalf) is designed so the company commits early to buildings, which are the cheaper part, and late to accelerators, which are the expensive part. That flexibility applies to the softer half of the book. The $431bn core has none of it.

6. Customer concentration: lower than it was

The five largest customers produced about a quarter of revenue over the first nine months of 2026, and the trend is downward. Revenue from outside the top five rose from ~64% (Q1 2026) to ~79% (Q3 2026), even as each of the top five reached a record. No single customer exceeds 10% of revenue. The filing does not name the big five. Penetration is also low: for its Fortune 500 customers, Anthropic revenue is on average under 0.1% of their total expenses. Even for the ten most-engaged it is under 1%. If enterprise AI spend is going to grow, Anthropic has plenty of room to grow with it.

7. Governance: public buyers get economics, not control

Anthropic is a Delaware public benefit corporation. That means its directors are legally required to weigh shareholders' financial interests against those of everyone materially affected by its conduct, and against a stated purpose: “the responsible development and maintenance of advanced AI for the long-term benefit of humanity”. After the IPO the company will have five share classes:

So a trust picks most of the board, and the founders hold a fixed majority of the remaining vote however many shares are issued. Holders of Class A shares can neither elect a board majority nor force a sale. Anthropic will be a Nasdaq “controlled company”, which exempts it from some independence rules. The founder control does end eventually: the Class F share is redeemed for nothing once fewer than three of the seven co-founders remain for 90 consecutive days. The trust has no such end date.

The filing also commits shares to public-benefit purposes:

• The Anthropic Public Benefit Fund received 846.8m Class N shares, 5% of the fully diluted company. That is ≈$50bn at the Series H price.
• A dilution rule tied to the share price. Each time the share price stays 50% above the last milestone (on 90 of 120 trading days, starting from the IPO price), the board may issue up to another 2% of the company to the Fund. After k milestones the price is up 1.5^k× and an existing holder owns at most 0.98^k of their original stake. If the price triples (≈2.7 milestones, so 2 completed), dilution is ≈4%. If it rises tenfold (5 milestones), dilution is ≈9.6%. This is small next to the gains that trigger it, but it is a deliberate transfer that ordinary companies do not make.
• The co-founders have each pledged 80% of their personal equity to philanthropy. The employee donation-matching programme cost $541m in 2025.

For an investor, the PBC structure is a promise that the company will sometimes choose to earn less. The filing names examples. In April 2026 Anthropic released its most capable model, Mythos, only as a gated preview to selected cybersecurity organisations through Project Glasswing, because of its ability to find software vulnerabilities. Fable 5 was released generally but with safety restrictions. The filing says future measures “may become more frequent”. Most companies treat delaying a product as a risk to avoid; this one tells investors it is part of the plan.

8. The risks particular to this company

The risk-factor section runs to roughly 90 pages. Most of it is standard (competition, talent, outages, regulation). Four items are specific to Anthropic:

• The Department of War dispute. In February 2026, after Anthropic said publicly that its models should not be used for mass domestic surveillance or fully autonomous weapons, the President directed federal agencies to stop using its technology. The Secretary of War moved to designate it a “supply-chain risk to national security” and said no military contractor could do business with it. Anthropic sued. On August 27, 2026, a federal district court largely granted it summary judgment, finding First Amendment retaliation. On September 25, however, a 2-1 panel of the D.C. Circuit upheld a separate supply-chain-risk designation under the procurement statute (FASCSA). Anthropic has asked the full court to rehear the case. The filing also discloses that the Department suggested it could use the Defense Production Act to force removal of the usage limits. The commercial exposure covers more than direct government revenue, because the designation reaches every company that does business with the military.
• Copyright. The Bartz class action over training data was settled in September 2025 for $1.5bn (approved July 2026). The filing warns that courts adopting Bartz's reasoning in other cases could create further liability.
• Suppliers that are also competitors. Amazon, Google and Microsoft supply compute, resell Claude, invest in Anthropic, and sell competing models. The filing describes the risk of conflicts “that affect our access to compute, its cost, [and] the prioritization of our workloads”.
• Model behaviour. Few prospectuses include a risk factor like this one: “Our AI models may behave in unintended, misaligned, or autonomous ways, including by taking harmful actions that no user or third party directed.”

9. Valuation: what a $1.5–2tn price has to assume

At the Series H price, the 16.9bn fully diluted shares implied by the Public Benefit Fund disclosure (846.8m ÷ 5%) are worth ≈$1.0tn, consistent with the ~$965bn post-money widely reported for that round. The press-reported IPO target of $1.5–2tn would be:

• 32–43× the Q2 2026 annualised run-rate ($46bn);
• 76–101× trailing-twelve-month revenue to June ($19.7bn).

As 10×, 20×, 30× revenue argues, a revenue multiple is a DCF run backwards, so the useful question is what it implies. A deliberately simple version with three assumptions: (i) by 2031 Anthropic is a mature company valued like one, at 25× earnings; (ii) it earns a 30% operating margin, which gives a value of 7.5× revenue; (iii) public investors want a 10% annual return. A $2tn price today then requires 2tn × 1.1⁵ ≈ $3.2tn of value in 2031, which means ≈$430bn of revenue, 9.3× the current run-rate, or ~56% compound growth for five years. At $1.5tn the requirement is ≈$320bn and ~47% a year.

Compared with the past two years, those growth rates are modest. Compared with any company of this size in history, they are extraordinary. Nothing has grown 50% a year from a $46bn base for five years. The compute book points to similar figures: committed spending in the late 2020s implies revenue in the low hundreds of billions a year just to keep a reasonable margin. The IPO price and the compute contracts are effectively the same bet.

The verdict

The bull case is in the numbers. The product earns an 85% gross margin, gets cheaper to serve with each generation, and is spreading through enterprises at a pace no software category has recorded. Customer concentration is falling. Operating cash flow has turned positive. With $130bn+ of cash, the company is not short of money in any near-term scenario.

The bear case is also in the numbers, mostly in the commitments. The company has locked in a decade of fixed costs against revenue that is almost entirely variable, and that revenue depends on releasing a frontier model every few months, which it has promised to delay when safety requires. Its main suppliers are also its investors, resellers and competitors. Public shareholders get a fully diluted economic claim and no control. The $80bn loss can mostly be set aside; the $431bn of take-or-pay obligations cannot.

What to watch in the S-1/A and the first two quarterly reports: the price range and the aggregate commitment figure, both blank today; Q3 2026 revenue; gross and operating margin through the December SpaceXAI ramp, which is the first real test of the timing problem; the trend in top-five concentration; the share of revenue on committed contracts; and the en banc ruling in the Department of War case.