How to Read an AI Company’s S-1: The 7 Numbers That Matter

An IPO prospectus is the one document in which an AI company has to tell you the truth on penalty of securities law. It is also long, repetitive and written by lawyers whose job is to make sure nobody can later claim they weren’t warned. Fortune reported that the draft of Anthropic’s prospectus, which leaked to Reuters in late September, spends more than a third of its length on risks.

Nobody reads all of it. You don’t need to. This guide is for investors, founders, employees with equity and anyone covering the wave of AI listings who wants to get from the cover page to the numbers that matter in an evening. We picked seven figures that tell you whether an AI business works, and each comes with a worked example from a real filing or a company’s own disclosure. The examples use figures as reported in September 2026; the method holds for any filing.

A note on the main example: Anthropic’s prospectus is a draft that Reuters reviewed, not a public filing, and its numbers can change before the company lists. We’ve covered it in our analysis of the leaked S-1 and in The Great AI IPO Wave, Explained.

1. Run rate versus GAAP revenue: ask for the window

Revenue under generally accepted accounting principles (GAAP) is what a company actually booked, audited, over a defined period. A run rate is a projection: take a recent slice of revenue and multiply it out to a year. Private AI companies love run rates because they grow faster than anything else on the page.

The method decides the number. AI video startup Higgsfield says plainly that its US$1 billion (CA$1.39 billion) run rate is its last four weeks of revenue multiplied by 13. That’s honest, and it’s also the best four weeks the company has ever had, extended across a year it hasn’t lived. AdExchanger pointed out that Anthropic has used a similar one-month lookback, and that short windows exaggerate seasonal peaks.

The S-1 forces the reconciliation. Anthropic’s audited 2025 revenue was nearly US$4.6 billion, according to Reuters. Its draft shows US$11.5 billion in the second quarter of 2026 alone, up from US$4.73 billion in the first, per Fortune. Both are real. But only the GAAP line tells you what happened, and the gap between them tells you how much of the story rests on the most recent month.

2. Compute commitments: the bill that’s already signed

Look in the notes for “purchase obligations,” “contractual commitments” or lease obligations. For an AI lab this is where the cloud and data-center deals live, and they often dwarf revenue. Reuters reported that Anthropic’s draft lists about US$518 billion (CA$720 billion) in planned spending on cloud, computing and infrastructure in the years ahead.

Set that against nearly US$4.6 billion of 2025 revenue and you understand the bet. The question to ask is how much of the commitment is cancellable and over what period it falls due. Those details usually sit in a table in the notes, not in the summary.

3. Customer concentration: who could sink the quarter

Companies must disclose customers that account for a significant share of revenue, typically 10% or more. Search the document for “concentration” and “significant customers.”

Reuters found that nearly a quarter of Anthropic’s revenue came from two customers. On its own, that’s common for a company selling through cloud partners. Paired with item 7 below, it becomes the most important sentence in the filing.

4. Gross margin after inference: what each answer costs

Software investors are used to gross margins of 70% or more because copying code is free. AI is different. Every query runs on chips somebody pays for. Find “cost of revenue” on the income statement, divide gross profit by revenue, then read the footnote that defines what’s in the cost line. Does it include only inference, the compute that answers customers, or training too? Companies make different choices, and the choice can swing the margin by tens of points.

Reuters reported Anthropic spent US$7.33 billion on compute and infrastructure in 2025, triple the year before, against nearly US$4.6 billion in revenue. That figure blends training and serving, which is exactly why you need the cost-of-revenue line rather than the headline.

While you’re on the income statement, check the net loss for non-cash items. Anthropic’s 2025 net loss was US$42 billion, but Reuters reported that roughly US$34 billion of it was an accounting charge tied to the estimated value of convertible financing. The operating loss, excluding writedowns, was a little over US$8 billion. Still enormous, and a different kind of problem.

5. Contracted backlog: ask how it becomes revenue

Under GAAP, backlog is reported as remaining performance obligations: contracted revenue not yet recognized. Companies also invent their own metrics, such as “total contract value,” which may include options, renewals or contracts that haven’t started. Always find the definition. It will be in the “key metrics” section of the management discussion.

The cleanest worked example of 2026 is Nscale, the London-based AI infrastructure company that filed on Sept. 18 to list on the NYSE as NSCL. It reported “active and contracted total contract value” of US$103.4 billion as of Aug. 31, up from US$38.0 billion at the end of 2025. Its revenue for the first half of 2026 was US$140.6 million, with a net loss of US$1.02 billion.

Our arithmetic: double that half-year revenue and the contracted value is roughly 368 times a year of current sales. That isn’t a red flag by itself, since data-center contracts are long. It does mean almost all of the value depends on facilities that still have to be built, powered and filled with chips, and the filing’s construction timeline matters more than its revenue.

6. Dual-class shares and who’s really in control

Check the “Description of Capital Stock” section for share classes and votes per share. Public buyers often get a class that carries fewer votes, or none that matter, while founders keep control. In AI there’s a twist: mission structures.

Anthropic’s Long-Term Benefit Trust, set up in 2023, holds a special Class T stock that lets five financially disinterested trustees elect a growing share of the board, ultimately a majority. Fortune flagged that the draft keeps leadership in control “to promote public good over market forces,” a CNBC headline it linked. For a buyer, the practical question is simple: if the trust and the shareholders disagree, who wins? The answer is in the charter, and the S-1 must summarize it.

7. Risk factors: the boilerplate and the real ones

Most of the section is defensive and generic. Competition, regulation, key personnel and cyberattacks appear in nearly every tech filing and tell you little. The real risks share two traits: they’re specific to this company, and they connect to a number you’ve already seen.

Anthropic’s draft has a lot of each. Reuters quoted warnings that models have shown unexpected behavior, including “sabotaging code, assisting fraud and manipulating information,” which is unusual language for a prospectus. The line that matters more for the stock is quieter: “many of its largest clients were not locked into long-term contracts and could cut or stop spending.” Put that next to item 3. Nearly a quarter of revenue from two customers, and many of the biggest customers free to leave.

Sources

• Fortune: Anthropic's leaked S-1 prospectus income statement
• Devdiscourse (Reuters): Exclusive: Anthropic's IPO prospectus shows sweeping AI vision, surging costs
• Yahoo Finance: Nscale files for IPO as pipeline passes $100 billion
• Renaissance Capital: Nscale (NSCL) IPO profile
• AdExchanger: Higgsfield's $1 billion run rate comes with an asterisk
• Higgsfield: Higgsfield crosses $1B annualized revenue run rate
• Anthropic: The Long-Term Benefit Trust