Basically, it's really really expensive (from a compute standpoint, and more compute = money and energy) to train a decent LLM just using books and conversations etc. Anthropic, OpenAI, etc spend a truly insane amount of money doing this is in data centers.
Running an LLM like that is also extremely expensive from a compute standpoint, and right now the true frontier models can basically only be run in datacenters.
However, once you have a really good LLM, you can use it to train another LLM. Because it's basically already refined all its original training data (and is capable of further refining it'd output on the fly), training a new LLM from it is vastly easier and cheaper.
Additionally, you can use a much smaller and more efficient model, so it can run on less powerful hardware.
So all local LLMs that are any good that exist right now, were trained using the outputs of existing super expensive frontier models. They could not exist without the frontier models to train them. This is also how OpenAI and Anthropic create their cheaper models. Mythos/Fable is anthropics current frontier model and they distill it into (use it to train) their cheaper models like Sonnet and Opus.