They already do that sort of, B2B pre-trained/post-trained models have their own system prompts/setups.

It depends on the actual implementation. There really isn't anything stopping them from including a "pre system prompt" or "root prompt" or whatever they want to call it, before your system prompt, even for API calls. So the "system prompt" becomes "developer prompt" but the model still receives a provider-authored prompt before yours. (and likely trained to take precedence over whatever you add)