To be fair: You'd need to make at least several billions of revenue by selling pencil sharpeners while mostly spending the $100b on R&D (not the production of individual sharpeners)!

The market for AI is very clearly there; people are willing to pay more than enough so that providing LLM infrastructure (inference) is profitable.

Just the SOTA model R&D is pretty unsustainably expensive right now, but this only means that if the status quo "bursts" (because people realize that pushing the frontier is not paying off), model improvement will (probably) be slower afterward.

Note: this could also happen because expensive SOTA progress fails to keep ahead of "garage-trained" models, in such a case SOTA-AI lab investments collapsing would not really slow AI progress anyway.