Am I understanding it correctly that the gist of your argument is that OpenAI, Google etc have run out of available data to train LLMs?

What's your estimate for how many tokens that represents?