I'm not an OpenGL expert, but my understanding is that the fixed function pipeline (FFP) is like a set of big, generic shaders and state that everybody had to go through to actually do the work of displaying 3D graphics. You could write shaders a fraction of the size that does just the work you need without having to pay the performance price for features you don't use. FFP-using code looks nice in tutorials, but performs terribly outside of demos, never mind the complexity tax it imposed for implementations as a 3D graphics layer for people who don't understand how 3D graphics works.

Also the old OpenGL API had immediate mode functions which encouraged people to trickle in interleaved data and operations; the exact opposite of what 3D APIs need to run fast.