1. Create a very high poly 3d model (e. g. 500k faces) and export it as FBX. I've included an example model for easy reproduction.
2. Drag it into the Workbench.
3. The FBX to TXO step will still be somewhat quick in absolute terms (e. g. 40 seconds), but TXO to XOB step will take extremely long (e. g. 30 minutes).

Here are some test imports I did on my R5 7600:

These data points very closely match a runtime complexity of n^2.
I highly doubt quadratic time is actually required for what should be a simple file format conversion, the code is probably just written in a very inefficient way and could be improved.
The fact that the time can be significantly reduced by separating the object into multiple meshes also indicates this (e. g. splitting the 450k object into 16 objects resulted in only 5m).

While this issue only has a small effect on the average model import, it is extremely limiting when high poly imports are necessary.
For example, I was trying to import reconstructed point clouds as reference for terrain making, which can easily have over a million faces.