How do we know inference is heavily subsidized? Are all these third party Chinese model providers subsidizing the true cost?