Yah, given that “training models” doesn’t stop when the model is finished and released. Like, a released model needs to be continuously tweaked to keep it up to date or to deal with problems that have occurred. Even if that’s not literally tokens used by customers, it is compute being used to provide service to customers.

And that’s just assuming that they’re not just hiding some compute costs used to service customer demand inside the R&D budget. “Oh, you see, this pool of customers are being served with an experimental version, so any compute here is actually R&D, any API fees or subscription payments made by them of course get counted towards normal revenue.”