Beijing probes DeepSeek and Moonshot over data routed to Claude
China's internet regulator has opened an investigation into two of the country's biggest AI labs — over the same allegations a US competitor levelled at them two weeks ago.
The Cyberspace Administration of China is investigating DeepSeek and Moonshot AI over alleged data leaks to Anthropic through its Claude models, and has questioned staff at both companies, The Information reported on Tuesday. The probe follows Anthropic's September 10 threat-intelligence report, which alleged that Moonshot silently forwarded customer requests to Claude and presented the answers as Kimi's own, that DeepSeek ran a parallel relay, and that some exchanges carried sensitive information from individuals, multinational companies and state-affiliated actors. Neither lab responded to the accusations at the time. Neither has commented on the investigation, and its scope has not been made public.
Why it matters: this is the first domestic consequence of the distillation fight, and it runs in the direction the accusations did not. Until now the pressure came from Washington and from the lab that was allegedly distilled from — we covered that report in Anthropic says Alibaba, Moonshot and DeepSeek trained on Claude. A Chinese regulator asking whether Chinese user data transited US servers is a different kind of exposure: it treats the alleged relay as a data-protection failure rather than an intellectual-property dispute, which is the framing Beijing's own law has always used. The timing is not gentle either. DeepSeek has been invited to brief the UN Security Council on AI risks this week alongside OpenAI's Sam Altman, and Moonshot is inside the due-diligence window of a confidential Hong Kong listing at a reported $50 billion valuation.
Our read: the uncomfortable part for both labs is that neither answer helps. Confirming the relay proves the data-protection failure; denying it contradicts traffic Anthropic says it logged.
The cheapest open-weight models now finish a task at roughly one fifth the cost of comparable closed models, and on sparse mixture-of-experts workloads the older A100 serves output more cheaply than the H100, according to a new paper from GPU market-data firm Ornn Data. The study standardizes eleven open-weight and eight closed models by intelligence score on the Artificial Analysis index, then prices a completed task rather than a token. Self-hosting on rented hardware lands at $0.12 to $0.35 per million output tokens at full utilization, and the hardware ranking reverses on gpt-oss-120b — a sparse model with 5.1 billion active parameters — where the paper computes the A100 at $0.29 per million output tokens in its base case against the H100's $0.64.
The rental market is pricing the same conclusion. Ornn's term marks show the five-year A100 price retaining 80.2% of its one-month price, against 43.7% to 59.8% across the Hopper family and 53.8% for Blackwell, while A100 occupancy ran from 74% in March to 90% in September. That is the signal lenders read off the B200's resale premium — Nvidia's B200 now resells for 58% above list — except arrived at from the demand side instead.
The caveats are load-bearing. Open weights price cheaper at several sampled score thresholds, not every one, and the closed frontier still sits above the open sample at the top of the range. The A100-versus-H100 sparse result stitches together two different third-party serving setups, the dense comparison still favors newer silicon, and Ornn sells GPU rentals and licenses this data commercially — a conflict the paper discloses itself.
What to watch: whether the investigation stays an inquiry or becomes a finding, and whether other Chinese labs start routing around US APIs to avoid the same exposure.
When a regulator on your own side opens the file first — is that a problem for the labs, or a rescue? Tell us in the comments.
