The multidimensional model needs the relational model to be available in star schemas (that is what you call "denormalized model") for loading the data. And in many cases, there is some processing like combining data from different sources, keeping the data for reporting longer than it is needed in the OLTP world, keeping historical views like old regional or department structures available for analyzing which are not necessary and hence overwritten in the OLTP world. Hence, this intermediate step makes sense in many cases. You might also want to have clear cut of times, i. e. always report data for complete days (or, in some cases, months), and not have some data for the last day available and some not, which makes comparison of numbers for a day easier than comparing e. g. the sales of today containing only the data up to 10 o'clock with the sales of the whole day yesterday.
In some simple cases, the intermediate relational data structure need not be available physically. A few days ago I prepared a prototype cube where the star schema was just a set of views on the source data. In this case, of course, the data was only physically available in the original source form and in the cube. The structure of the source data did not make the views that inefficient, and thus data loading to the cube was fast enough for the prototype.