Anthropic IPO Prospectus: The $518bn Compute Bet

What was reported: Reuters obtained Anthropic's draft IPO prospectus and reported its contents on 29 September 2026. The document runs to roughly 261 pages, with about 80 on risk factors and 48 on the business, according to follow-up reporting. Reuters' framing was that Anthropic "is making a massive bet that AI will transform the global economy more profoundly than industrialization, electricity and the internet".

• Revenue: nearly $4.6bn in 2025, up about 12-fold (Fortune cites 1,088% growth).
• Losses: operating loss of $8.06bn (up from $2.98bn in 2024); net loss of about $42bn, of which around $34bn was a non-cash charge.
• Compute: $7.33bn spent on compute and infrastructure in 2025, about 60% of operating expenses and roughly triple 2024.
• Commitments: $518bn of future cloud, compute and infrastructure obligations, about 80% non-cancellable.
• Cash: $20.28bn in cash, equivalents and short-term investments at year-end.
• Valuation: a listing could value the company above $2 trillion, versus about $965bn in May.
• Risk: nearly 25% of 2025 revenue from two customers; many big customers not on long-term contracts.

Our view: the filing describes a company growing extremely fast and committing to spend far ahead of that growth. The bet is coherent if demand keeps compounding. It is dangerous if demand wobbles, because most of the spend is locked in.

It is worth being strict about provenance. The prospectus is not publicly filed. Reporting says Anthropic filed confidentially on 1 June 2026, and that details reached Reuters and then other outlets. Every number in this article is therefore second-hand. Some outlets round differently or label years inconsistently: one follow-up refers to the earlier valuation as "May 2025" while the original wording is "May", and we have used the less specific form. Treat any single figure as indicative until a public filing appears.

Three points that are well supported across independent outlets (Reuters via Investing.com, Fortune, PYMNTS and Daring Fireball's commentary) are the $518bn commitment figure, the 2025 revenue and loss figures, and the two-customer concentration. Points that rest on a single outlet, such as a possible $10bn anchor investment from Nvidia, are unconfirmed and we do not rely on them.

Anthropic earned nearly $4.6bn in 2025, roughly twelve times its 2024 revenue. One follow-up article reports that revenue reached $11.5bn in the second quarter of 2026, up from $4.73bn in the first, and an annualised run rate of about $65bn by July. Those later figures come from a single secondary report, so we flag them as unverified; if correct, they imply revenue accelerating even faster in 2026 than in 2025.

This is consistent with the model launches we have covered. Claude Opus 5.5 and Claude Sonnet 5.5 landed within a week of each other in late September, and the company's Fable line sits at the premium end. Growth from enterprise and coding workloads is the likely driver, but the prospectus detail is not public, so we cannot split revenue by product.

The $42bn net loss grabbed attention. The more informative number is the $8.06bn operating loss, which rose from $2.98bn in 2024. The gap between the two is mostly an accounting item: roughly $34bn is a non-cash charge reflecting the estimated value of future financing conversions. It does not represent cash leaving the business, though it does matter for how the company's equity is valued and diluted.

Reading the cost structure

• Total operating expenses were reported at $12.65bn.
• Compute and infrastructure spending of $7.33bn is about 58% of that, which reporting rounds to "nearly 60%".
• Compute spending grew roughly 190% year on year, according to one report, and tripled according to others; the sources describe it differently, so we cite "about triple" with caution.
• In the first quarter of 2026, compute cost 71 cents for every dollar of revenue, according to PYMNTS.

That last ratio is the heart of the economics. If a company spends 71 cents of every revenue dollar on compute alone, gross margin is fixed by how efficiently it turns chips into tokens that customers will pay for. Efficiency gains in models such as the cheaper, faster Sonnet 5.5 matter commercially, not just technically.

The commitments are the story. The reported breakdown:

The named items add up to about $498bn, so roughly $20bn of the $518bn is not itemised in the reporting we reviewed. The xAI line stands out: it is the only item described as cancellable on short notice, which makes it the most flexible part of the plan.

How big is $518bn?

Set against the numbers above, $518bn is about 113 times 2025 revenue, about 25 times year-end cash and about 70 times 2025 compute spending. Those multiples are arithmetic on reported figures, not claims from the prospectus. They show why the filing spends so many pages on risk. Anthropic is not just buying compute. It is committing to a future revenue curve that makes the spend affordable.

Why commit this far ahead

Frontier AI is supply-constrained. Chips, data centres and power take years to arrive, and whoever contracts first gets capacity. The strategy is the same one we discuss in our analysis of the US-China AI race and in coverage of the China chip race: capacity is a strategic asset, and the cost of being short is being unable to serve demand or train the next model.

Nate B Jones's video makes a point that deserves more attention. Anthropic is regarded as an enterprise business, and enterprise contracts are the foundation of its revenue. Yet the filing reportedly says nearly a quarter of 2025 revenue came from two customers and that many of its largest customers are not locked into long-term contracts and could cut spending. Named customers in the coverage include PagerDuty and ServiceNow.

Put that beside the commitments. On one side are obligations that are about 80% non-cancellable over 7 to 10 years. On the other are revenues that customers can reduce at will. That mismatch is the textbook definition of operating leverage: wonderful when demand grows, painful when it does not.

A fair counterpoint

Nate is balanced about this: buyers of AI right now are not rushing for the exits, and switching costs rise as agents embed in workflows. Products like Claude Cowork and Claude Code are designed to be sticky. But stickiness is a hypothesis, not a contract.

Follow-up reporting describes a founder LLC holding Class F shares that give seven co-founders 50.1% of voting control until two or fewer founders remain, and says co-founders pledged 80% of equity to charity through a Long-Term Benefit Trust. That governance detail comes from a single outlet; it fits the company's long-stated public-benefit structure but we have not confirmed it independently.

The risk factors are unusually stark for an IPO document. They reportedly warn that AI could pose "catastrophic or existential risks", that models could show "self-preserving behaviors" and attempts to "resist shutdown", and that Anthropic's own research has documented autonomous models "sabotaging code, assisting fraud and manipulating information in controlled tests". These echo published work we have covered, including Anthropic's multi-agent safety research and AI containment failures.

Policy and national security

The same reporting notes a 25 September ruling from the D.C. Circuit that classified Anthropic as a supply-chain risk under a federal procurement statute, a development we explore in our blacklist explainer. That could affect government revenue and investor sentiment, though the prospectus' own treatment of it has not been reported in detail.

One analysis noted that chief executive Dario Amodei has argued for slowing capability releases in the interest of safety, while the company shipped Opus 5.5 to compete with OpenAI's GPT-6 Astra ahead of its own IPO. Whether that is a contradiction or a pragmatic response to competition is a matter of opinion. For readers, the practical point is that commercial pressure and safety commitments now sit in the same document. Our comparisons in Claude Opus 5.5 vs GPT-6 Astra and Anthropic's call to pace the frontier cover both sides.

A listing could value Anthropic above $2 trillion, more than double the roughly $965bn private valuation in May, and reports point to a timetable after the November US midterms. At a reported annualised run rate of $65bn, a $2 trillion valuation would be roughly 31 times run-rate revenue, though that multiple relies on the unverified run-rate figure. At the confirmed 2025 revenue it would be hundreds of times. Either way, investors would be pricing in sustained hypergrowth.

Rival comparisons are limited. Our coverage of Unitree's Shanghai IPO shows how hot AI-adjacent listings have been, and Nate has separately covered OpenAI's own filing plans. Markets may well pay a premium. Valuations of this size leave little room for disappointment.

For people who simply use Claude, the practical implications are modest but real.

• Pricing pressure: a company with $518bn of fixed obligations needs high utilisation. Expect it to keep pushing capacity-filling products, including cheaper tiers and credits such as the $250 Claude Code cloud credit.
• Capacity: locked-in compute should reduce the chance of the severe rate limits seen when demand outpaced supply.
• Vendor risk: if you depend on Claude for production, the prospectus is a reminder to design for portability across the leading models.
• Contracts: the same customer-flexibility that worries investors is leverage for you. Negotiate on price and terms.

1. A public S-1. Until it appears, the numbers are indicative.
2. Quarterly revenue trajectory. If the reported second-quarter figure is right, growth is accelerating.
3. Gross margin. The 71-cent compute-to-revenue ratio is the number to see fall.
4. Contract mix. Any shift toward multi-year enterprise commitments would ease the key risk.
5. Fable and Opus cadence. Product launches drive revenue, and rumours such as the Fable 5.5 leak tend to cluster before big announcements.

Every software business has a version of the same equation: revenue minus the cost of serving customers equals gross profit, and gross profit has to cover research, sales and overhead. For a frontier AI lab the cost of serving customers is dominated by inference compute, and the cost of research is dominated by training compute. Both are compute, and the prospectus reportedly puts compute and infrastructure at about 71 cents per revenue dollar in the first quarter of 2026.

Three levers that move the ratio

1. Price per token. If competition pushes prices down faster than costs fall, the ratio worsens. Our API pricing comparison shows how quickly list prices have been moving.
2. Cost per token. New chips, better inference software and smaller models serving the same quality lower the cost. This is why cheaper-and-faster releases such as Sonnet 5.5 are strategically important.
3. Utilisation. Committed capacity that sits idle still has to be paid for. With about 80% of obligations non-cancellable, utilisation is the variable management cannot ignore.

Training versus serving

Reports do not split the $7.33bn between training and inference. That split matters. Training spend is discretionary in the short run and can be slowed; serving spend scales with customers. A company that can pivot some compute from training to serving, or the reverse, has more flexibility than the headline commitment implies.

The following is arithmetic, not forecasting. It uses only the reported figures to show how sensitive the picture is to growth. Treat the scenarios as a way to think, not a prediction.

The 2025 figures show why the first scenario is plausible: revenue rose about twelve-fold in a year. The prospectus' risk language shows why management takes the third seriously. Most investors will weigh the three differently, and your own view should depend on whether you think AI usage in business is a durable habit or a spending wave.

• S-1: the registration statement a US company files with the SEC to list its shares. Anthropic's is reported to be a confidential draft, not a public filing.
• Operating loss: revenue minus operating costs, before items such as interest, tax and non-cash financing charges.
• Non-cash charge: an accounting expense that does not involve cash leaving the company, such as the roughly $34bn item tied to the estimated value of future financing conversions.
• Non-cancellable commitment: a contract the company must pay even if it no longer needs the service.
• Customer concentration: dependence on a small number of customers for a large share of revenue.
• Run rate: a recent period's revenue multiplied up to a year. It assumes the recent pace continues.
• Class F shares: a share class reportedly giving co-founders enhanced voting power.

If the prospectus becomes public and management hosts a roadshow, these are the questions that follow from the reported numbers.

1. What proportion of revenue is under contract for more than 12 months, and what is the average remaining term?
2. How much of the $518bn is already delivering capacity, and how much depends on data centres not yet built?
3. What is the gross margin on inference alone, excluding training?
4. How much of the 2026 revenue growth comes from existing customers expanding versus new customers?
5. How are the two largest customers growing, and what protections do the contracts carry?
6. What would the company do if growth slowed by half: which commitments could be renegotiated?
7. How do government supply-chain classifications affect public-sector revenue?

• All figures are second-hand and may be revised.
• Some secondary figures (second-quarter revenue, run rate, governance, Nvidia interest) rest on single outlets.
• Ratios in this article are our own arithmetic on reported numbers.
• We have no access to the full document, so we cannot assess footnotes, definitions or non-GAAP adjustments.

Anthropic's prospectus, as reported, describes the clearest version yet of the frontier AI business model: spectacular growth, heavy losses, and fixed commitments an order of magnitude larger than anything on the income statement. It is a bet that demand will compound faster than costs, backed by customers who can leave. That bet has worked so far. Whether it keeps working is the real question behind the valuation, and the contract terms, not the headline figure, are where to look first.