I do seeing this becoming a trend more and more. A lot of open models are fairly strong and fine-tuning and other post-training techniques could easily make bespoke domain-specific inference considerably cheaper (in token usage and energy consumption) and allow for more data sovereignty.