Generally I think the question is whether they actually are selling it at a markup. Inference is easier to reason about. I think the problem is financing training. The issue is the future value of inference seems to be being massively overstated to justify present day expenditures on training (particularly since the value of training today evaporates extremely quickly)