What makes you think the API costs are "the true price" and anywhere near their inference costs? Also what percentage of users do you think actually maxes out their subscription?

I suspect they're subsidizing a lot less than people think.