The reason for FPKM is mostly historical as there are practically only disadvantages in distributing the data this way.

• There are several posts and publications showing that FPKM is inferior to other units.
• FPKM is not directly compatible with most DE packages.
• Providing raw counts would instead allow anyone to compute the transformation they wanted (CPM, TPM, FPKM), while the FPKM transformation is not easily reversible.
• FPKM manifests biases and errors in the gene prediction, especially it is not suitable for draft genomes where the exons are often not well annotated.
• FPKM need to be represented as floating point values, introducing unnecessary rounding errors and maybe data volume, while the counts can be represented by integers.

R(F)PKM/TPM values are used to normalize read counts by library size (total number of reads you have in a given RNAseq experiment) and the length of the feature (gene/transcript). But remember that commonly used software for differential expression analysis (DESEQ2/EdgeR) are using raw counts instead of normalized values (they do their internal normalization steps).

Okay, I just found that link which says that FPKM makes it possible to compare Gene A to Gene B even if they are of different lengths, and to compare Sample 1 and Sample 2 even if they have different library sizes.