I think there's similarity to the dynamic too. Not many instances have the means at this time to produce LLMs from scratch and operate them at scale.