overheads of pipelines

The NVDL discussion has got me thinking about the overheads of pipeline usage.

Let us say I am building an application and I want it to be easily
extensible in the future, so I want to use a pipeline, but currently
the application only has two steps. I'm wondering if a pipeline in
that context wouldn't seem resource intensive for no good reason?
But then, neglecting to build my application as a pipeline, in someway
this smacks of premature optimization, when in reality using the
pipeline when unnecessary would actually be premature extensibility.

So, questions:

1. what are the initial resource usages of a generic pipeline, and how
do pipelines generally scale - is there a theory, algorithm etc. or is
it all implementation specific.

2. should one always build the most extensible application even if the
current needs do not require extensibility - my personal opinion is
yeah

Cheers,
Bryan Rasmussen