Instancing is really a whole subject on it's own… it can be useful for rendering copies of huge geometry efficiently. The benefits won't be particularly noticeable on simple geometry.
http://www.a3d.co.uk/images/delayed.htm [a3d.co.uk] is a cool read
When instancing a cube, you might not see a huge improvement when instancing over simply rendering it but the improvement will become more obvious as the complexity of the geometry increases to say that of a character or a vehicle. With Delayed Loads, the IFD (which is part of the rendering process) generation will be fast, the IFD will be smaller, and so network bandwidth is less.
Improvements happen because Mantra:
a) only needs to load the model from disk/IFD once
b) can ray-trace it without using additional memory per instance since the geometry may be shared between instances
c) is able to, in the case of micropolygon renders, share the base (prediced) geometry
d) not need to load the geometry if it is occluded or offscreen