Anthropic admits up to $2.33 billion of revenue came from customers it later banned

If you build a product on top of someone else's models, there is one line in Anthropic's new prospectus worth reading. In a document filed with the SEC on November 2, 2026, the company acknowledges that a noticeable share of its money came from people it later labeled as violators. This is not a scandal or an accusation of double standards, but rather an honest look at how moderation works in a large API business.

What the prospectus says

According to the updated IPO prospectus, customers whose use of Claude was later found to violate the rules accounted for about 5–7% of revenue over the first nine months of 2026. Total revenue for the period was $33.242 billion. Simple arithmetic gives a range of $1.66 billion to $2.33 billion. Nobody at Anthropic named this figure directly: it was calculated from metrics the company published.

The company explains plainly why this happens. Detecting a violation, conducting an investigation and blocking an account is a process, not a single button press. Throughout that time the customer keeps using the model and paying. The company recognizes revenue for services already rendered, so a subsequent block does not retroactively cancel it.

Imagine a café that issued a bill for lunch and only afterward found out the guest had paid with someone else's card. The food has been eaten, and the guest will not be let in again. The principle is the same here, only the scale is different.

The gray market: facts versus speculation

The post that the news item draws on places another figure right alongside. According to The Bell, the Russian market for reselling access to foreign AI models grew to 100–150 million rubles per month by August 2026. A reader can easily connect the two figures and conclude that billions of dollars were earned from users in Russia, China and Iran.

That is not a valid conclusion. Anthropic does not disclose what share of the disputed revenue comes from the gray market or from individual countries. A violation of the rules can be anything from circumventing regional restrictions to the abuses the company describes in its threat reports. The Russian access-resale market is just one possible source, and next to billions of dollars its turnover looks modest.

What this means for developers and founders

First. If your product depends on a single provider, the account is part of your infrastructure, and it can be shut off, not instantly but after an investigation. At Anthropic this process clearly takes time, and that should be factored into your plans.

Second. Buying access through intermediaries is a business risk, not just a way to save money. If the account you work through is found to be violating the rules, the block can arrive without warning, and nobody will refund you for tokens already consumed. Storing critical data and prompts in such an environment is a bad idea. This is a conclusion drawn from the logic of the prospectus, not a direct quote.

Third. For those looking at the market as a product manager or investor, the figure is useful as a benchmark. For a large AI provider, five to seven percent of revenue in the gray zone is not an exception but the expected price of scale. Moderation always lags behind growth, and the faster revenue grows, the more money passes through that lag.

Limitations

Keep in mind that we see only a calculation built on a rounded percentage share. Exact sums, the structure of the customer base and the criteria for violations are not disclosed in the prospectus. It is not possible to derive from this an assessment of the company's honesty or the volume of "Russian" revenue. But a practical conclusion can be drawn: rely only on legal access, plan for a backup provider, and do not build a business on an account that can be switched off without your involvement.