The Power BI dataset is tabular and, as such, Deneb processes and generates the supplied Power BI data as a tabular dataset in the Vega view. There isn't the ability to type data from the semantic model in a way that allows us to detect if it's JSON (and infer that it's spatial). We also can't have multiple queries per visual due to constraints in Power BI, so the tabular dataset was prioritized for greater flexibility for developers. If you are using the certified visual, you can currently only use the hard-coded approach that @davidebacci has identified.

I have some ideas for "post v2", where we could supply a valid scalar TopoJSON object from a semantic model via a conditional formatting property in the properties pane, which circumvents the query limit but allows us to potentially treat such a property and inject it as a spatial dataset (provided it parses). We'll also need to consider what this means for reuse via templates, as it creates an additional kind of dependency beyond the traditional tabular dataset that we currently assume.

Note that v2 is still under active development. As we're discussing potentially after this timeframe, it is not a valid short-term (or possibly medium-term) solution. However, I wanted to let you know that I'm aware of it as a limitation and am considering how to solve it in a future iteration.