Because it's not losing money on each token? Aside from most global people using American inference providers to run the models, I suspect the cloud inference products of the Chinese labs are profitable, at least on the inference costs (ie: not including model training, salaries, etc).