What actually influences CU consumption in Dataflows Gen2?

Hi mrbartuss,

Yes, exactly and the saving is actually proportional to your table count.

Without query referencing (no "Reference" step):
Each of your 19 output tables would independently evaluate its own query all the way to the source. That means 19 separate HTTP requests to SharePoint + 19 full Excel file parses through the Mashup engine. Each one is billed at 12 CU/s for up to 10 minutes.

With query referencing (your current setup) :
The SharePoint read happens once (the base "Source" query), and all 19 output tables branch off that single evaluation. Only the filtering/column selection steps run 19 times which are near-instant in-memory operations.

So in your Dimensions dataflow specifically, query referencing could be saving you up to ~18× the cost of a full source re-read. You're already doing the right thing there.

The remaining optimization lever (point 1) would remove the SharePoint dependency entirely once the file is in OneLake as a Delta table, even that single source read becomes a local Delta scan instead of an HTTP call to SharePoint.