[alsodumb]

You are definitely over estimating how much time and effort it needs to build large models.

Sam will get billions of dollars if he starts a new company. So there's no issue of money. In terms of data and training models, look at Anthropic - they did train a reasonable model. Heck look at Mistral, a bunch of ex Meta folks and their LLaMA team lead who spinned up a good models in months.

The only bottle neck i could think of would probably be RLHF data - but given enough money, that's not an issue either.