A related beef though, is that with any reasonable size dataset ontology-based inference is computationally very difficult, you have to cut all sorts of corners and know all sorts of tricks to actually infer across your data. In other words- if semantic data are going to truly become ubiquitous we need to infer in real time across them. Inference takes everything in your dataset into account, so adding a single axiom means if you want to be complete, you have to compute all over again -> slow.