If the number of polygons were not an issue, then "welding" vertices would have little consequence, as polygons are constructed of vertices. For aircraft models, most vertices are welded anyway, otherwise you'd likely get discontinuities in the shading of the model. It's not about welding-- it's about how much is in the scene as a whole, and what is being done to render all elements of the scene.
A contemporary graphic card is heavily optimized to render discrete, single-material objects fast, even when the object's data size is large. But if you throw enough objects at it, things will slow down, even if you're running the latest hot gaming rig. And with all those objects thrown at it, it pays to reduce the vertex/polygon count of any given model to what is necessary to visually represent the model, because all those calculations for each vertex consume resources that could be spent doing something else. This is especially true when an object's material uses transparency, as Thorsten describes.
A study of professional models done for professional sims and games reveals that the model's 3D mesh is often surprisingly simple. They use many tricks to reduce level of detail based on viewing distance. For the model, the detail effort and the majority of the art goes into the materials applied to the models. Good texture work dominates the model visually and obscures much of the model's actual simplicity while ensuring that the model has a chance of playing nicely with other objects in the scene on a typical user's rig.
Often modelers consider only their model under the optimized conditions of their particular platform rather than the scene as a whole: your aircraft, the aircraft of other MP pilots, AI aircraft, ground scenery models, cloud models, terrain, etc., etc. Designing with the scene in mind is fundamental to game design, and applies to simulators just as well.
-Buck, who has ranted on this more than once.