The chats alone are backbreakingly costly relative to the market mix of ChatGPT.

Rest of the market be damned -- combined with the poor customer mix (low to middle income countries) this explains why there has been such a push by the big labs to attempt to quantize models and save costs. You effectively have highly paid engineers/scientists running computationally expensive models on some of the most expensive hardware on the market to serve instructions on how to do things to people in low income countries.

This doesn't sound good, even for ad-supported business models.