I'm not sure how different the table structures between AX and D365 are, nor what the data volumes are you are looking at.
I would extract the different D365 sites into a reporting database as well as the AX sites, keeping the load on the source database as limited as possible (and possibly using some incremental loading - reload last x days). Using views you can then align these schema and load each into the semantic model. You can choose to union those tables in PowerBi Query editor (or even with a view in the reporting database) and report on them.
In this case I wouldn't opt for 3 semantic models, as you'll only report on 1 of them and add complexity without a specific benefit (at first sight).
Refreshing a powerbi semantic model by ingesting all data from all AX and D365 sites might be a big strain on the operational system, especially because all that data has already been read the day before (with exception of the new/modified data).
The views can also allow you to implement a filter at the cut-over date when migrating from AX to D365, so that accidental late entries are not considered.
