GitHub is the most visible, but the entire industry is trying to figure out how to charge closer to what the product costs them to provide.

Unfortunately for them, the space is being rapidly commoditized. Training a model is really just about throwing a bunch of money around. The training data is available to anyone. The methods are all published or easily intuited from published work. The hardware is in short supply, but if you're a big company that wants to get into the space, that's solvable.

Then you've got smaller companies using the larger models to distill their own cheap models, and the big guys are having a lot of trouble playing this cat and mouse game. The resulting cheap models are quickly gaining ground with clever context engineering tricks, especially for tasks like writing code where RLVR can be used to fine tune the model weights. You still need a lot of VRAM for inference, but as that gap closes, the price ceiling for coding products descends further. We're not too far from developers being able to generate most code locally on somewhat affordable hardware.

The companies building these data centers really need another killer use case like coding where most people don't care that it was produced by a machine and users can trust or easily test the output. Elon decided to just lease out one of his data centers and wait for that idea to come along, which is actually kind of smart. Of course, he seems to think that use case is a Robot Army.