What happened
Anthropic has filed confidentially with the SEC for an initial public offering. On September 29, 2026, Reuters reported on a copy of the draft prospectus it had seen. The document is not public, so what follows is Reuters' reading of it. Anthropic did not comment on the reporting. A few details matter for buyers. Reuters reported that 47% of Anthropic's 2025 revenue was sold through cloud partners Amazon and Google, up from 11% in 2023 and 32% in 2024. Those partners also collect customer bills on Anthropic's behalf. Two unnamed customers each accounted for 12% of revenue, and Anthropic warned that many of its largest customers are not bound by long-term contracts and could reduce or stop spending. The draft also says Anthropic expects usage-based revenue to remain the substantial majority of its business, and it warns that its business could suffer if the compute it gets from third parties is "curtailed, repriced or terminated."
Why it matters for business owners
This is a rare case where a vendor describes, in its own words, what it depends on. Most AI buyers only see the product page and the invoice. A prospectus has to name the risks, and several of them land on the customer side. Three stand out. First, how you buy matters. If a large share of a vendor's sales runs through a cloud marketplace, many customers hold their contract, billing, and support relationship with a cloud provider rather than the model maker. Second, the vendor says its own supply of computing power can be repriced. Prices you pay are downstream of that. Third, the vendor says it needs a continuous cadence of new model releases to stay competitive, which means the model you tested may not be the model you are running next year. None of this is unusual for a company at this stage, and none of it says the service is about to break. It says the ground can move, and a business that has built workflows on one vendor should know which way.
What owners should not misunderstand
Do not read a risk-factors section as a prediction. These sections list everything that could go wrong, because the law requires it. Treating every line as likely is as wrong as ignoring all of it. Do not assume this is only about Anthropic. Any AI vendor that sells through cloud marketplaces, rents computing power from larger companies, and prices by usage has a version of the same dependencies. This filing happens to be public enough to read. Most vendors' equivalents are not. Do not assume that buying through a big cloud provider removes vendor risk. It can simplify procurement and billing. It also adds a second party between you and the model maker, with its own terms, and Reuters notes that those cloud companies are also investors, suppliers, and competitors of the vendor. And do not treat a draft as final. The filing may change before it is published, and the figures are Reuters' account, not something we have read directly.
The operational lesson
A vendor's service has more moving parts than its product page shows. Your price, availability, and model behavior sit on top of a supply chain you do not control. The practical response is not to avoid AI vendors. It is to know what you are relying on and to keep your own options open. The same logic the filing applies to Anthropic's big customers applies to you in reverse. Anthropic says its largest customers are free to leave without long-term contracts. Your business should hold the same freedom: workflows that can move to another model, data you can take with you, and no single prompt, plugin, or integration that only works with one vendor.
What a serious business should do next
Take your two or three most important AI tools and answer five questions in writing. Who is the contract with, the model maker or a reseller? What happens to your price if the vendor's costs change, and what notice do you get? Which model version are you relying on, and how will you hear about changes? What would it take to move the workflow elsewhere? Who on your team owns that answer? Then check your terms for the points that decide the outcome: price-change notice, model-change notice, data export, and termination. If you buy through a cloud marketplace, check which party handles support and billing disputes. Finally, avoid long prepaid commitments to a single AI vendor until you have tested the workflow on real work. Flexibility is cheap now and expensive later.
The Atlacis view
Atlacis helps owners slow down before an AI decision, map the workflow, understand what a tool depends on, and choose based on the business instead of the announcement. A vendor's own risk language is useful input to that work, because it shows where the dependencies are before they turn into surprises. If you are relying on one or two AI vendors and are not sure how exposed your workflows are, that is a good conversation to have before the next renewal.
The short version
- On September 29, 2026, Reuters reported on a confidential draft of Anthropic's IPO prospectus. The filing is not public, and Anthropic did not comment.
- Reuters reported that 47% of Anthropic's 2025 revenue was sold through Amazon and Google cloud marketplaces, and that two unnamed customers each made up 12%.
- Anthropic's draft warns that its compute could be curtailed, repriced, or terminated, and that it needs a continuous cadence of model releases.
- The lesson applies to any AI vendor: know who your contract is with, what your price depends on, and how you would move a workflow if you had to.
- Risk language is a checklist, not a prediction. Use it to ask better questions before a renewal or a large commitment.
Where ATLACIS can help
Sources
- Reuters: Anthropic IPO prospectus lays bare deep dependence on Big Tech partners (Echo Wang, Krystal Hu, September 29, 2026, via MarketScreener)
- Reuters: Anthropic's IPO pitch embraces AI's promise and peril (October 1, 2026, via The Straits Times)
- Fortune: Anthropic IPO S-1 prospectus has leaked (Jim Edwards, September 29, 2026)