A model with 100k vertices MAY have, not will have, 300k after loading. For obvious reasons, you cannot share two vertices that have different values with the same index, and a cube simply has no shared vertices. Your average model with smooth surfaces is going to have plenty of shared vertices. The problem is sharp corners; at those points you have discontinuities in some attribute, typically the normal, and those are the cases where you cannot have shared vertices anymore. A cube consists entirely of sharp corners.

A model with 100k vertices MAY have, not will have, 300k after loading. For obvious reasons, you cannot share two different vertices with the same index, and a cube simply has no shared vertices. Your average model with smooth surfaces is going to have plenty of shared vertices.