Furthermore, engines come at a huge efficiency cost. Instead of knowing your own products, you've got to master someone elses.
Seriously that is the same stupid argument industry newbies and old fogies make for not using the C++ STL, the boost libraries, etc... and is the primary driver of NIH syndrome. Can you actually provide some statistics or real examples of cases where it is more cost-efficient to write your own comparable engine rather than licensing an existing one? I'm sure there are some for niche cases that mainstream engines don't serve (say you want to do primarily sparse voxel octree rendering) but for the most part you can extend an existing engine if there are bits and pieces you need that it doesn't provide.