On 26 October 2026 Anthropic filed its registration statement with the US Securities and Exchange Commission. We have already written about what its risk factors say about insurance. This note is about a different part of the same document: what it tells a buyer about the companies an AI supplier depends on, and the companies that depend on it.
Most supplier assessments stop at the first tier. They ask where your data is processed and who the subprocessors are. The filing shows why that is not enough for a model supplier. Behind the model sits a small number of compute providers, bound to it by commitments measured in tens of billions of dollars each, who are also its investors, its sales channels and its competitors.
Three providers, three roles each
The filing says Anthropic relies "on a limited number of CSPs", naming Amazon, Google and Microsoft, and on suppliers such as Broadcom, for compute. It sets out the size of the cloud commitments, each made "regardless of actual usage":
- Microsoft: at least $44.9 billion between November 2026 and May 2033.
- Google: approximately $114.5 billion between April 2026 and July 2033, with a further $125 billion to $150 billion estimated for additional TPU compute from 2028 to 2031.
- Amazon: at least $110.0 billion between May 2026 and April 2036.
Those three companies are not only suppliers. The filing records that they have made "significant equity investments" in Anthropic, that each has AI models and applications that compete with its own, and that they are also how many customers buy. Sales through cloud marketplaces made up approximately 48% of revenue in the first half of 2026, up from 11% in 2023.
On moving away from them, the filing is direct. Migrating workloads to alternative providers "would be costly, time-consuming, and operationally complex, and may not be feasible within timeframes that would avoid material disruption" to the company's model roadmap or its ability to serve customers.
Your model supplier carries the same concentration risk you are trying to manage, one tier further down. Its own filing says it could not move quickly either.
What the customer figures say
The filing also discloses the other side. No single customer accounted for more than 10% of revenue in the nine months to 30 September 2026, and the top five accounted for approximately 25%. It adds that many top customers are not on long-term contracts, and that it has agreements with certain large customers "which provide flexibility around financial commitments, termination rights, indemnification, and security obligations".
Read that as a buyer. Those terms are negotiable, and the largest customers have negotiated them. If a business process depends on a model, the indemnity, the exit terms and the security obligations in your contract deserve the same scrutiny those customers gave theirs.
The model you assessed will not be the one you run
One line deserves more attention than it will get. Model versions released in the six months to 30 September 2026 "accounted for over 80% of our revenue" in the three months to that date. The filing also says the company has changed, and may change, "usage limits, rate limits, or model and product capabilities", and that customers who have integrated older models "may resist migration".
For a buyer that means the supplier assessment is not a one-off. The model you evaluated in spring is unlikely to be the one carrying most of your traffic by autumn. If your records name the vendor but not the model version, the evaluation they rest on has expired and nothing says so.
Read the headline loss correctly
The figure most coverage led with is a net loss of $80.7 billion for the first half of 2026. Of that, $72.4 billion is a change in the fair value of convertible notes and similar instruments, which the filing attributes primarily to the rise in the company's own valuation. Revenue for the half was $16.3 billion, against $1.2 billion a year earlier, and the second quarter produced operating income of $974 million.
The headline figure says less about the supplier's operations than it appears to. For a buyer the more useful question is what happens to you if one of the dependencies above changes: a compute dispute, a marketplace relationship ending, a model withdrawn or its limits changed.
What to do this quarter
Map the second tier. For each model supplier, record which clouds provide its compute and which channel you buy through. If you buy through a cloud marketplace, you depend on both companies. The filing says its contract with you generally survives a change to a platform relationship, but that customers may be unable to move to another access point on acceptable terms or timelines.
Record the model version, not just the vendor. Every evaluation, approval and risk assessment should name the version it was run against, so that a change of model triggers a review rather than slipping past one.
Test the exit. Know which workloads could run on a second model and what it would cost to move them. The supplier has told its investors that moving is hard for it. It will not be easier for you unless you plan it.
Read the contract against the filing. The largest customers have flexibility on commitments, termination, indemnity and security. Ask which of those terms are open to you. Our AI supplier questionnaire covers the first tier; add the questions above to it.
We are not neutral
ORCAHQ sells governance, and we route work across several model suppliers, Anthropic among them. Read our emphasis with that in mind. The figures above are the supplier's own, filed under securities law, and anyone can check them against the source.