The API inference cost to customers is not the actual cost of providing inference, and the cost of providing API inference need not be the cost of providing subscriber inference.
This is correct, they are subsidized but it's the training cost that costs the most with a majority of people hitting cache for most queries for inference.