In many instances, an image service will suffice well, but for power users, for ad-hoc queries and analysis, the full geometry and attribute data is often needed. And in the case of a distributed enterprise, here is one place where a purely SOA-oriented approach begins to break down.
Things becomes a bit more difficult when it comes to vector geometry, as most GIS clients are still only geared toward consuming and processing vector data in one chunk.
Further, vector geometries can't well be broken into tiles without causing other breakage - polygons and linear features need to retain their topological integrity in order to work.
Yes, one can certainly cache vector feature services, provided the underlying data is relatively static, or apply constraints limiting the amount of data that one can fetch at a time, but is there any possibility, looking down the road, of utilizing more efficient serial or multiple parallel processes to rapidly and efficiently stream large and complex vector datasets?
I think there will need to be, and I don't yet see OGC, ESRI or anyone else looking at this. I'd be interested in hearing other folks' thoughts and experiences on this...