Why? My wife has a ChatGPT $20/mth account, rarely uses now the novelty has warn off (I asked she said twice this month), the only reason she keeps paying for it is to retain the chat history. That's a tiny very rarely used docker container on a server somewhere. Maybe I'm thinking about this wrong, but I would just run this the same way we ran digitalocean, all the $5/mth accounts with no activity sit on a box configured for that, once it starts to do stuff it moves to a box with less VMs on it, etc, this is how subscription levels in cloud work, sure it cost us $100MM to get there, but now it just prints cash. If OpenAI decided to stop training models and focused on unit economics, I don't see how they couldn't get it to profitability quite quickly, even with all the competition.

would you mind elaborating why you think their assumption doesn't hold? I have a 6 year old, back then 400EUR (consumer price mind you) GPU that could happily serve 10 "normal" users. sure, we have to add other hardware too, like chassis, power supply, CPU, RAM, ssd etc. so lets 4x that. Now we're looking at 1600 investment + power cost. My setup comes down to 0.04 eur per hour of operations, so about ~30 bucks a month (mind you electricity is very likely cheaper wherever else you are). We also need maintenance and operations, eyeballing in large scale at ~2EUR/user/month.

On a 3 year amortisation period we're looking at

investment: 1600

monthly cost to operate: x * 3 + x 2 + 1600/36

monthly revenue: x 20

For x users = 10

monthly cost: 50 + 44.4

monthly revenue: 200

net month result: +105.6

net total result: +3801.6

i know i know thats a napkin calculation but if anything hardware cost will be lower for big providers, which would give them even more profits. just looks like a feasable business model to me. as long as you can sell those subscriptions