I wouldn't even be surprised if they were losing money on paying ChatGPT users on inference compute alone, and that isn't even factoring in the development of new models.
There was an interesting article here (can't find the link unfortunately) that was arguing that model training costs should be accounted for as operating costs, not investments, since last year's model is essentially a total write-off, and to stay competitive, an AI company needs to continue training newer frontier models essentially continuously.
