> they really, really want to own the entire value chain rather than being relegated to becoming just another "model provider"

This is really the salient point for everything. The models are expensive to train but ultimately worthless if paying customers aren't captive and can switch at will. The issue it that a lot of the recent gains are in the prefill inference, and in the model's RAG, which aren't truly a most (except maybe for Google, if their RAG include Google scholar). That's where the bubble will pop.