They are not artificially cheap, they are still cheap even when hosted by independent inference providers. Are all providers subsidizing their open-weight models?

Nobody's making profits right now, not because they're selling tokens for less than their cost but because they're always investing in the next bigger model.