Imagine a baker who sells half his bread through two supermarkets. Those supermarkets have lent him money. They rent him his oven. And they sell their own bread on the next shelf along.
That is, with only slight simplification, Anthropic's situation. It is described in the document the company prepared for its stock market listing, which the Reuters news agency has seen.
What the filing reveals
In 2025, 47% of Anthropic's sales, or 2.16 billion dollars, went through Amazon's and Google's cloud platforms. They are the ones offering Claude to their own customers, and collecting payments on Anthropic's behalf.
That share keeps growing: 11% in 2023, 32% in 2024, nearly half in 2025.
Yet Amazon and Google are not mere resellers. They have invested tens of billions of dollars in Anthropic. They supply much of the computing power it needs. And they are developing their own AI, in direct competition with Claude.
Anthropic presents this situation as an asset: thanks to these giants, it reaches customers it could never have reached alone, at a speed no company could match on its own.
Two customers, a quarter of sales
The dependence does not stop there. According to the document, two customers, whose names are not disclosed, each accounted for around 12% of 2025 revenue. Between them, nearly a quarter of sales.
This is exactly the kind of information a company never publishes voluntarily, and that a stock market listing forces it to disclose. We wrote about this in September in our guide to reading the figures of AI companies: it is in this kind of document that the inconvenient truths are found.
In a document intended to convince investors to buy its shares, Anthropic writes that advanced AI could present catastrophic or even existential risks, and details the difficulties of assessing the safety of its own models. It is rare for a company to describe its product as potentially dangerous at the very moment it is seeking to sell it on the stock market.
The defeat against the Pentagon
The same document lands in a difficult week. On 25 September, a federal appeal court ruled in the Pentagon's favour, by two votes to one.
The case dates back to the start of the year. The Department of Defense wanted to be able to use Claude for all lawful purposes. Anthropic refused to amend its contract to authorise mass surveillance and autonomous weapons. The Pentagon then classified it as a "supply chain risk". We had reported the start of that trial.
The court's decision allows the Pentagon to exclude Anthropic from its systems, and to bar its contractors from using its products for defence work. Amazon, Google and Microsoft, by contrast, continue to offer Claude to their civilian customers.
Refusing a customer has a cost. That is exactly the question we raised in our reflection on the right to refuse to sell. The answer is starting to be quantified.
What we take away
Anthropic is one of the most prominent companies of the moment. Its filing shows that it is also one of the most dependent: on two giants that are at once its partners and its competitors, and on a handful of very large customers.
This is not unique to Anthropic. It is the paradox of this entire industry, where everyone is at once the customer, the supplier and the rival of everyone else. What changes with the stock market listing is that we finally have the figures to measure it.