Look at it this way: To render a stack of math to get a pixel-color costs you X.
Part of that X can be precomputed and offloaded to the virtual texture, you can assume you save some cost up front from pre-computation.
There are no draw-calls associated with the stuff rendered into the texture if you do not show the mesh, but you still have to render it, if it’s lit, etc, the cost of the materials, etc. All that ought to be the same as a standalone mesh to get it INTO the VT, but there are no draw-calls like a static mesh.
I can’t qualify/quantify the exchange between draw-calls and instruction-count, it’s like trying to convert weight to time; not really straight forward.
SO! What I can say is you save draw-calls (which can generally be more expensive) and lookup overhead, but you still have (generally) the same rendering cost as the things you are capturing to the VT. So, generally, if you can make it work with a VT, it’s going to come out in your favor.
Does that help?