My last information was, that the inference itself is cheap but its subbsidy comes from ignoring the training.

Really? The rumors are that their margins on tokens is quite healthy. That doesn't seem like loss-leading to me.