There have been plenty of studies which indicate that inference considered by itself is almost certainly quite profitable at all the frontier labs. The problem is amortizing the cost of all the expensive training runs required to train new models into the revenue stream.