Anthropic's S-1 explained: $16.3B in sales, and a bill that can't be cancelled

Anthropic's IPO filing went public on 26 October, and the number everyone quoted was a loss: $80.7 billion in six months. The more revealing number is how fast the company can lose its customers compared with how long it is locked into paying for chips. Read carefully, the S-1 describes a business with astonishing growth and a cost structure that behaves like a bet on usage never dipping.

Disclosure: this desk uses Claude models in its drafting process, and Claude is made by Anthropic. We have tried to weight outside critics at least as heavily as the company's own framing, and we label which numbers come from the filing and which from analysts.

The plain-English version

Imagine you run a restaurant that has become the hottest in town. Every month more people come, and the till is overflowing. To keep up, you sign leases on three more kitchens, each opening in a year or so, and the leases cannot be cancelled. Your diners, though, are not season-ticket holders. They pay per meal. If a rival opens down the road, or they simply eat out less, they can stop coming tomorrow.

That is, in simplified form, what Anthropic's filing describes. The "restaurant" is the company that makes the Claude AI models. The "meals" are tokens, the small chunks of text an AI reads and writes. The "kitchens" are compute: the chips and data centres that run the models.

What the filing says, according to reporting by Fortune:

• Revenue for the first six months of 2026 was about $16.3 billion, compared with about $1.2 billion in the same period a year earlier. About 80% of it is usage-based, mainly from businesses.
• Anthropic says it earns more than $2 of revenue for every $1 it spends on compute, and the second quarter showed an operating profit of about $974 million.
• It has also signed compute contracts worth hundreds of billions of dollars, typically contracted 12 to 18 months ahead of when the capacity arrives. Estimates of the total differ widely (roughly $475 billion to $750 billion) depending on what is counted.
• Customers can often reduce usage or end contracts with little notice. The company warns of "periods of excess or underutilized compute and committed cash outflows that do not produce immediate revenue."

Why the giant loss then? Most of it is bookkeeping. About $72.4 billion is a re-valuation of financial instruments (such as convertible notes held by investors) that rose in value as the company's own worth soared. It reduces reported profit without any cash leaving the building.

So a fair way to explain this to a friend: Anthropic is growing extremely fast and, at the operating level, is close to break-even or better. The risk is not that it is losing money today. It is that its costs are fixed for years while its income can change quickly. Whether that is a good bet depends on whether people keep finding more uses for AI faster than AI gets cheaper.

How it actually works

Three ideas explain most of the filing.

1. Revenue per compute dollar is the unit economics. Training and serving models needs chips. The more revenue each dollar of compute produces, the healthier the business. According to the Filing Insights summary, compute and infrastructure expense was about $8.9B in the first half, or 55% of revenue, versus about $2.3B (201% of revenue) a year earlier. A year ago the company spent two dollars on compute for every dollar it earned; now it earns roughly two for every one it spends.

2. Costs are lumpy and early; revenue is smooth and late. According to Fortune, Anthropic usually contracts compute 12 to 18 months in advance. Capacity arrives in big chunks, and its cost hits the books when it comes online, whether or not demand has caught up. Anthropic says margins "may compress in those periods," per the Fortune takeaways piece. The company also says it has underestimated demand before, which left it short of capacity. So it faces risk on both sides: too little compute and customers go elsewhere; too much and it pays for idle machines.

3. Everything depends on who the counterparties are. Per Fortune's compute analysis, Anthropic buys capacity from Amazon, Google and Microsoft, all of which also compete with it, plus Nvidia GPUs owned by SpaceX (up to about $120.5 billion through 2029) and, per Mostly Metrics, newer providers such as Nscale and Lambda. It is also starting its own "Anthropic Direct" capacity, where partners secure land and power and Anthropic leases the compute. Its software routes work across Google TPUs, AWS Trainium, Nvidia and AMD chips to keep costs down and avoid dependence on one supplier.

A rough walkthrough of the risk:

The mitigation in step 4 is real but still small. Fortune reports remaining performance obligations (contracted revenue not yet recognized) of about $10.8 billion, tiny next to the compute commitments.

The deep dive

What the primary document is, and what we could and couldn't read

The registration statement is on SEC EDGAR (accession 0001628280-26-067168, filed 2026-10-26); the main document is about 14 MB. Our fetch tool could not load the full document, so the figures below are the filing's numbers as reported by Fortune, Filing Insights, MLQ and Mostly Metrics. Where those secondary sources disagree, we say so. Check the filing before quoting any single figure.

The income statement, decomposed

One discrepancy: Fortune's compute-strategy piece cites $5.6B of compute expense and $11.6B revenue for Q2 specifically (a ratio just over 2), while the $8.9B figure above is for the half. Those are consistent: about $3.3B in Q1 and $5.6B in Q2 would sum to roughly $8.9B. We derived that split ourselves; it is not a reported number. (Another Fortune piece describes $5.6B as the six-month figure, so outlets disagree; the $8.9B half-year number comes from Filing Insights.)

Operating leverage, with a toy model

Let revenue be $R = u \cdot p$ (usage times price per unit) and compute spend $C$ be fixed over the commitment window. Define revenue per compute dollar $\rho = R / C$. The share of revenue left after compute is

At $\rho \approx 2$, half of each revenue dollar is left for everything else: personnel, sales and partnerships, stock compensation, legal costs. Now suppose usage is unchanged but price per token falls by a fraction $q$:

At $\rho = 2$, a 50% price cut takes $\rho' = 1$, so compute alone absorbs all revenue. This is an illustration (it ignores that cheaper tokens usually raise usage, that Anthropic can improve efficiency, and that some compute is trained rather than served), but it shows why the filing lists price competition as a risk, and why the author of Mostly Metrics frames the central question as whether use cases grow faster than per-task costs fall. That analysis notes Opus cost per task reportedly fell about 40%, and Anthropic's argument is that cheaper tasks drive more usage.

The same leverage runs in the other direction. If usage grows faster than planned, the incremental revenue carries little incremental cost until the next tranche of capacity arrives. That is the mechanism behind a ratio that Mostly Metrics calculates as a roughly 5x improvement in revenue per compute dollar over five quarters.

The commitment numbers

Published totals vary because the categories differ:

• At least $475B of future cloud purchases; up to $750B with chip-leasing deals (Fortune). Another Fortune piece sums named deals higher still, up to about $808B, depending on what is counted.
• $542-567B per a separate Fortune analysis, which includes a $125.2B five-year lease (Broadcom) that is "non-cancelable except on default" per Filing Insights, which also warns that a default on certain commercial contracts could trigger cross-defaults on credit agreements.
• Named contracts: SpaceX up to about $120.5B through 2029; Nscale up to about $44.6B through 2033; Lambda about $41.3B through 2034; AMD more than $20B from 2027 (Fortune, Mostly Metrics).
• The filing's own aggregate reads "more than ____" in key passages (Mostly Metrics), which is probably why press totals do not agree.

MLQ adds a useful warning: commitment totals are not a current debt balance, since they include contingent items, leases that have not started, and obligations running as far as 2045.

Customer and retention metrics

As of 30 September, per Filing Insights: more than 11,900 customers above $100,000 in trailing-twelve-month revenue, more than 1,900 above $1 million, about 250 above $10 million; no customer above 10% of revenue; top five about 25% of revenue over nine months. Cloud marketplaces accounted for 48% of H1 revenue and 58% of receivables.

Net revenue retention is about 900%, defined (per Mostly Metrics) as current-quarter revenue from customers who spent at least $1 in the same quarter a year earlier, excluding Pro and Max subscribers. Snowflake (158%) and GitLab (152%) are the comparison that analysis uses. It is internally calculated and not independently verified.

Run-rate revenue

The filing defines run-rate as trailing 28-day consumption revenue times 13 plus monthly subscription fees times 12, but reportedly gives no figure. Earlier public figures ($30B in April, $47B in May, $65B in a July Bloomberg report) are not in the filing, and Mostly Metrics notes OpenAI disputed the $30B figure as overstated by about $8B, partly because cloud revenue is booked gross.

Why it matters

For everyday users. The pricing and availability of Claude depends on this arithmetic. A company that must fill committed capacity has strong reasons to cut prices for committed buyers and to push usage-heavy products such as coding agents. A company that is short of capacity rations it. The filing warns of both situations.

For developers and builders. The filing describes a push toward multi-year minimum-spend contracts with discounted token rates. If you build on any frontier API, expect more of those offers. Also note the structural dependency: Anthropic says the hyperscalers that supply its compute treat it increasingly as a competitor, and they gain "visibility into our pricing and commercial terms."

For companies. About 80% of revenue is usage-based, mainly from businesses. Over 50% of the Fortune 500 spend $100,000 or more with Anthropic, but Mostly Metrics calculates that is under 0.1% of their total expenses on average, implying large headroom, and also that spend can be shifted without large switching costs.

For the field. A frontier lab's SEC-filed numbers, rather than leaked run-rates, are now in front of the public. It makes AI's unit economics checkable: revenue per compute dollar, compute as a share of revenue, and the gap between committed spending and contracted revenue. Expect OpenAI and others to be compared against it. It also offers a read on the AI infrastructure chain: a lab whose revenue scales with usage is the demand signal that the cloud providers' data-centre bets depend on.

Second-order effect: Anthropic's governance (a trust that elects a board majority, and a founders' share class holding 50.1% of votes, per Filing Insights) means public shareholders cannot easily force a change of direction if the economics sour. That affects how quickly commitments could be renegotiated. For how the company's public-benefit equity pledge fits in, see our earlier piece on Anthropic's 5% Public Benefit Fund.

What to be skeptical of

Which "revenue". Mostly Metrics and others argue that sales through cloud marketplaces are booked gross, and that net of the cloud providers' cut, H1 revenue would be closer to $13.6B, with implied gross margin in the low 60s rather than the ~85% that cost of revenue ($2.4B on $16.3B) suggests. That is an analyst's estimate, not a company disclosure. Anthropic may well account for it correctly under the rules; the point is that comparisons with companies that report net could mislead.

The 900%. NRR excludes consumer plans, starts from a small base, and in a pay-per-use business partly measures price structure. Nobody outside has cohort-level retention data.

One profitable quarter. Q2's $974M operating profit is the only quarterly operating profit in the six quarters the filing presents (MLQ). Mostly Metrics expects GAAP results to turn negative again as roughly $23.7B of stock compensation is recognized over about 2.1 years after the IPO.

Missing items. Fortune asks why Q3 results are absent, how much the convertible notes mask true expenses, and what the payment schedule for compute is. The filing also says Anthropic stepped up enforcement against policy-violating customers in August and September and expects that to reduce revenue and growth; it estimates 5-7% of this year's revenue came from customers later found violating its policies.

Growth will slow. The company says so itself: "We expect our revenue growth rate to decline over time and to fluctuate from period to period."

Source quality. Several figures differ between outlets (cash balance, commitment totals), and we could not read the full filing ourselves. Treat any single number as provisional until you check EDGAR.

What to watch next

1. The next amended S-1. The registration statement is not yet effective. Look for the blank aggregate compute-commitment figure to be filled in, and for Q3 numbers. Fortune flagged their absence.
2. IPO pricing and timing. Fortune reports shares are expected to begin trading between 9 November and Thanksgiving, under the ticker ANT; the filing leaves size and price undetermined.
3. Revenue per compute dollar next quarter. If it holds above 2 as new capacity comes online, the leverage story is working; if it dips, the "margins may compress" warning has arrived.
4. Price moves. Watch whether Anthropic and rivals cut token prices again. OpenAI just launched a cheaper computer-use model, GPT-6.1 Terra, at $1 and $5 per million input and output tokens (OpenAI-announced pricing, as reported by technology.org, MarkTechPost and others); that is the kind of pressure the filing names.
5. Anthropic Direct. How fast its own leased capacity grows relative to hyperscaler capacity.
6. Safety disclosures. The filing reportedly includes safety-related disclosures (Filing Insights lists a state-sponsored misuse of Claude Code and models behaving differently under evaluation). For related earlier coverage, see our Mythos 5 GitHub-attack explainer.

How this was made: researched and written by an AI model (Claude) from the primary sources listed above, then checked claim-by-claim against those sources in a separate AI fact-check pass. Spotted an error? Email [email protected] and we correct it publicly. Our process.