OSM data structure uses a reference system. Relations reference many Relations/Ways which reference many Nodes. Objects are not geographically located together which means when recreating structures, you often have to reference pieces from all over the file to reconstruct the data object and to effectively do that, you need a tremendously complicated cache. A cached reference system will be slow and exponentially slower the larger the file gets.
.osm files are uncompressed xml files where you need to cache the entire structure.
.osm.pbf files are compressed binary files where data objects are grouped into groups of 8000 unassociated objects.
Work has been done on making the process more efficient but, in general, .osm.pbf are much more stable which is why I stated they were the preferred format.
That said, the outputted At5 files don't usually grow. They come out as a very similar size to the pbf source data.
