Yes, when anyone proposes building our tools on top of these services I ask “what will happen to this when they start charging what it really costs to run these models?”
In general I agree with you, but llms are the one exception where it’s not practical and not cost effective to run them locally. If you want to use them, the better option is by far to pay someone for the service.
Then second best option is an inference provider for open weight models, so at least if they raise the price or stop offering it you can get it from someone else or eventually upgrade to self hosting.
Good example of why I don’t rely on technology I don’t control. I want my workflow to be future-proof and have a predictable cost.
Yes, when anyone proposes building our tools on top of these services I ask “what will happen to this when they start charging what it really costs to run these models?”
In general I agree with you, but llms are the one exception where it’s not practical and not cost effective to run them locally. If you want to use them, the better option is by far to pay someone for the service.
That’s because now is the phase where they let you try the good stuff cheaper to hook you up.
That’s true, but there are also massive economies of scale when you can afford the big Nvidia gpus and share them among many people.
Then second best option is an inference provider for open weight models, so at least if they raise the price or stop offering it you can get it from someone else or eventually upgrade to self hosting.
I agree. I use openrouter myself.
https://openrouter.ai/