Nobody knows how the future will be. There are lots of options. There are self hostable open models today - they are not as good, but some report they are good enough. Right now I don't have the hardware to try them (I could - if I'm willing to accept a small number of tokens per minute - useful means more tokens per second as I can get in an hour). However I expect in a few years I'll be able buy consumer hardware with good enough GPUs (or perhaps an AI chip???). Will anyone other than a few hobbyists buy them? time will tell.
The moat is Being Able To Do Things No One Else Can and as both hardware and software commoditise that moat disappears.
Open LLMs are already only a few months behind the frontier models and as soon as someone works out that accessible and cheap domain-specific custom training has the potential to be a gamechanger the Big Names will die.
Math puzzles only get you so far.
LLMs are in their Big Compute phase, and it's only a matter of time before that ends - because the intelligence of the frontier models isn't increasing fast enough and providing big enough benefits to make the models more valuable than cheaper, smaller competition.