The most revealing figure in Anthropic’s registration statement is not a product line. It is a distribution channel.
Nearly half of the company’s revenue, 48%, arrives through cloud marketplaces operated by Amazon, Google and Microsoft, according to the S-1 made public Oct. 26. In 2023 the same measure was 11%. A customer buying Claude is increasingly billed by a cloud provider rather than by Anthropic, which means the model developer’s growth is partly a function of three other companies’ sales forces.
The revenue itself: $16.29 billion in the first half, against $11.6 billion a year earlier, an increase of roughly thirteenfold. Cost of revenue was $2.38 billion, leaving a gross margin near 85%, a level more typical of licensed software than of a business that rents computing by the hour.
The loss line needs reading in parts. Anthropic reported a net loss of $80.7 billion, of which $72.4 billion came from non-cash fair-value changes on convertible notes and derivative liabilities. Those entries move with the company’s own valuation, so a higher private price produces a larger paper loss, and a public listing will make the swings visible every three months.
Growth is slowing. Sequential revenue growth decelerated from 145% in the second quarter to 45% in the third, according to the filing. Some investors have cut valuation expectations from $2 trillion to about $1.5 trillion, people familiar with the discussions said.
The offering is being built for individuals as well as institutions. JPMorgan is leading the U.S. retail tranche, and the company has discussed allocating 15% to 20% of new shares to individual investors. That is well above the share usually set aside in large technology offerings, and it is a bet that retail demand for AI names will absorb stock that institutions price more carefully.
Then comes the purchase schedule. Anthropic has committed at least $110 billion to Amazon Web Services through April 2036, about $114.5 billion to Google Cloud, and about $44.9 billion to Azure. A separate agreement with SpaceX covers up to $120.5 billion of GPU capacity through 2029, which works out to roughly $3.1 billion a month at the ceiling.
Those dates read as a calendar of obligation rather than a forecast. The Amazon agreement runs a full decade. The Microsoft commitment reaches into 2033. The SpaceX schedule ends in 2029 and is priced monthly, which makes it the most visible of the three and the easiest for an investor to model.
The marketplace channel explains why the customer mix looks the way it does. Selling through a cloud provider lets a model developer reach enterprise buyers without building a global sales organization, and it lets the buyer attach AI spend to a contract already approved by its procurement and security teams. The trade is margin and ownership of the relationship: the cloud provider sets the billing, bundles competing models into the same commitment, and knows who the customer is.
That arrangement has a consequence for pricing. When three companies sit between a model developer and its buyers, discounting decisions get made in someone else’s deal negotiation, and revenue becomes sensitive to the renewal calendar of contracts the developer did not sign.
The commitments turn a variable cost into a fixed one. A company that reserved capacity it cannot fill still pays for it, and obligations of this kind do not shrink if usage comes in below the reservation. That is a reasonable structure when demand is compounding at triple-digit rates. It is a harder structure to carry when quarterly growth falls to 45%.
Compute is not the only fixed claim. Headcount, research budgets and data acquisition all scale with the capability race, and the competitive logic of the sector rewards spending ahead of revenue because a model that arrives late is worth less than one that arrives first.
What the numbers also show is how much of the AI business has become balance-sheet work. The largest line items in a modern model developer’s filing are not salaries or marketing; they are multiyear capacity reservations and the accounting treatment of instruments whose value depends on the private price of the company’s own equity.
For public investors, that creates a valuation problem with no clean precedent. A company can report 85% gross margins, thirteenfold revenue growth and an $80.7 billion loss in the same document, and all three are accurate. The first two describe an operating business. The third describes how the company is financed.
The retail allocation adds a second question. Individual investors buying a listing of this size will be reading a document in which a majority of the loss is an accounting entry and most of the future cost is a contract signed years in advance. Institutions will price the discount rate on those obligations; individuals are more likely to price the growth rate.
Anthropic’s bankers have an argument available. Cloud distribution is how enterprise software has been sold for a decade, and the marketplaces brought the company into accounts it could not have reached alone in the time it took to grow thirteenfold. The counterargument is that a channel which handles half of revenue and is owned by three potential competitors is not a channel; it is a dependency with a renewal date.
Both readings will be tested in the same roadshow. The company has disclosed enough for investors to choose, which is more than most listings at this scale manage.
