Best way to Load-Transform-Cache for Power BI Report fetching data from AWS S3 Bucket

Option 2 is the right direction for this scenario.

Using Dataflow Gen2 only as an ETL layer and landing the curated fact and dimension tables into a Lakehouse gives you a clean separation between ingestion, transformation, and consumption. It also makes the data transparent and reusable, since you can inspect and query the Lakehouse tables directly in SQL or Spark without having to reverse-engineer logic from a semantic model.

Building a single semantic model on top of the Lakehouse keeps your reporting layer simple and avoids duplicating refresh logic, transformations, and security rules across multiple models. Incremental processing is easier to control upstream, and your semantic model refresh becomes more stable and predictable.

This approach also scales better as data volume grows and aligns well with Fabric’s intended architecture, especially when you have multiple facts, shared dimensions, and role-based security requirements.