Forty-eight percent. That is the share of Anthropic’s revenue in the first half of 2026 that came through the cloud marketplaces run by Amazon, Google and Microsoft, according to the S-1 registration statement Anthropic filed publicly with the SEC on October 26. In 2023 it was 11%. Anthropic confirmed the filing the same morning: Class A shares on the Nasdaq Global Select Market under the ticker “ANT,” with Goldman Sachs and Morgan Stanley named on the prospectus and the share count and price range “not yet determined.”
The headline numbers match what the leaked draft already showed. Yahoo Finance’s live coverage puts first-half revenue at $16.3 billion, the first-half net loss at $80.7 billion, and second-quarter revenue at $11.6 billion with $974 million of operating income, after a $1.7 billion operating loss in the first quarter. We walked through those draft figures two days ago in our buyer’s read of the AI IPO race. The public document adds results through September 30, and three of its disclosures change what a Claude customer should take from it.
The concentration risk moved from customers to channels
The draft story was customer concentration: reporting on it said two customers produced a fifth or more of revenue. The public filing reframes that. It says no single customer accounted for more than 10% of revenue in the nine months to September 30, and that about 79% of third-quarter revenue came from outside the top five customers, up from about 64% in the first quarter. It also reports more than 11,900 customers paying $100,000 or more on a trailing twelve-month basis, against roughly 950 a year earlier.
The concentration that grew is in distribution. Marketplace sales went from 11% of revenue in 2023 to 32% in 2024, 47% in 2025 and 48% in the first half of this year, and marketplace receivables were 58% of Anthropic’s net receivables at June 30. The filing describes the mechanics plainly: “These platforms bill end customers, collect payments on our behalf, and remit such amounts to us, net of applicable platform and other fees.” It adds, in the financial statement notes, that “Certain of the CSPs also hold the Company’s non-voting redeemable convertible preferred stock and convertible notes.”
Put those together and the three companies that bill roughly half of Claude’s revenue are also shareholders and creditors, suppliers of most of its compute, and sellers of competing models. The forward-looking statements list “potential conflicts of interest with cloud service providers and chip partners that develop competing AI offerings” as a named risk. For a buyer, that changes the meaning of the procurement choice. If you buy Claude through a marketplace to burn down a cloud commit (the same move we covered when OpenAI’s models arrived on Google Cloud), your contract terms, support path and data handling run through a party with its own model to sell. That may still be the right call. It should be a deliberate one.
The compute bill quintupled in six months
At December 31, 2025, Anthropic’s non-cancelable purchase commitments with terms beyond one year totaled $54.6 billion, mostly cloud hosting. By June 30, 2026, expected fixed or minimum guaranteed commitments under those agreements were about $265.6 billion, with terms of one to 20 years, and the filing says agreements without a fixed minimum are “expected to be significant” on top of that. Separately, between May and October Anthropic signed agreements with SpaceXAI (formerly X.AI) to spend up to $120.5 billion on Nvidia-powered compute through 2029, roughly $3.1 billion a month once all capacity is delivered, which it expects in December 2026.
The filing also records what happens when demand outruns supply. In March 2026, it says, Anthropic had a service outage and periods of degraded performance caused by “a significant and rapid increase in usage demand, compute constraints in certain components of our infrastructure, and other factors.” That is the vendor’s own account of the risk behind every rate limit. December is when the SpaceXAI capacity is due; until it lands, keep a second model qualified on another provider for the workloads you cannot pause. The Broadcom financing sits on top of this stack, and the S-1 shows how much more stack there now is.
The pricing risk factor tells you where contracts are heading
The most useful paragraph for a buyer is in the risk factors, under pricing. Anthropic says it “may be required to reduce our prices, offer shorter contract durations or offer alternative pricing models, such as outcome or task-based pricing.” The next sentence explains why: “our agentic offerings are structurally more compute-intensive than our non-agentic offerings.” If it cannot implement pricing that reflects that, or “if competitive dynamics constrain our ability to price these products at levels sufficient to cover their costs, our compute efficiency could deteriorate as agentic usage grows.”
Read that as a signal, not a threat. Token prices have fallen (the filing itself credits Claude Fable 5.1 with cutting the cost of typical workloads by about 25% versus Fable 5), and a lab that says agentic work may not cover its costs at today’s prices is a lab that will look for a different unit to charge on. Outcome or task-based pricing is that unit. It can be good for a buyer with clean task definitions and bad for one who cannot audit what counted as a completed task.
One more figure explains why all of this matters to your budget: the filing reports Net Revenue Retention of about 900% as of September 30. Anthropic defines it as current-period revenue from qualified customers divided by their base-period revenue, excluding Pro and Max subscribers. On average, existing customers are spending around nine times what they spent a year earlier. Some of that is new use cases. Some of it is agent loops nobody budgeted for.
Three things to do with this before your next renewal. Decide whether marketplace billing or a direct contract better serves your data and support terms, rather than defaulting to whichever burns a commit. Ask your account team how a move to task-based pricing would be measured and audited, in writing, now. And put a ceiling clause in any multi-year agreement, which was the advice in this morning’s look at how discount terms decided the Replit account and holds even more after reading the prospectus. The share count and price range are still blank in the filing, so the valuation headlines will come later. The number that affects your bill is the 48.