China is investigating DeepSeek and Moonshot using Anthropic's own count of their Claude traffic

Disclosure: this newsroom's assistant is built by Anthropic, whose allegations this investigation draws on.

China's Cyberspace Administration is investigating DeepSeek and Moonshot AI over user data that may have reached Anthropic's servers without customers knowing, The Information reported on 22 September, followed by Bloomberg. The probe rests on Anthropic's allegation that some requests to the two companies' services were forwarded to Claude for processing, without users being told that an American system was involved. Anthropic has said the two used Claude more than 35 million times in total: more than 23 million interactions attributed to Moonshot and more than 12 million to DeepSeek. Executives and employees at both companies have been called in for questioning. Regulators are examining what kinds of data were transferred, how they were routed, and whether either company broke Chinese law on moving information abroad. The probe comes days before Xi Jinping is due in Washington, where AI is expected to be discussed.

Why it matters

Anthropic made these numbers public as evidence of misuse — Chinese labs quietly leaning on a US model, the same family of complaint as its allegation against Alibaba. Beijing is now using the same numbers as evidence of something else: Chinese users' data leaving the country. One disclosure, two governments' enforcement agendas, and the Chinese companies caught between them on the same facts.

The structural point applies to any frontier lab. A provider's abuse detection works by logging who called the model, how often and with what. That record is by construction a record of foreign users' data sitting on the provider's servers. Publishing the totals makes a security case at home and hands a data-sovereignty case to the regulator abroad. Neither use is illegitimate, but the second was not the purpose of the first, and there is no version of detecting routed traffic that does not also produce it.

What this does not establish matters as much. An investigation is not a finding; the CAC has not said what it has concluded, and the specific content said to have been transferred is reported so far only thinly — we are not repeating those details until they are confirmed. Anthropic's figures are its own attribution of traffic it observed, and neither company has, as far as we can find, publicly accepted them. The probe tells you Beijing takes the allegation seriously enough to question executives. It does not tell you the count is right.

What to watch

What the CAC concludes and whether it penalises either company, whether DeepSeek or Moonshot contest Anthropic's attribution in public, and whether AI model routing comes up in the Trump–Xi talks.