But no Chinese company has copied and distributed an American LLM.

Even if they had wanted that, it would have been impossible, because Anthropic et al. do not let anyone access them directly.

Querying the available API can be used to extract only an extremely small part of the information stored in an LLM.

That part can be used for the post-training of another LLM, to obtain some desirable properties, but the extracted information is far too little to be called "copying" or "distributing". Moreover, after the post-training it does not appear anywhere in the weights verbatim, so its use is at least as transformative as the training of the original American model.

Besides these facts, there is no evidence that the Chinese companies have actually done this, even if it sounds plausible.