However, I have seen plenty of example where that is not true.

Enterprise databases are a great example. While the DBA can often do alot to compensate for dev's bad db code ... sometimes that just is not enough.
A fulltime employee can work on optimizing code for an entire year before it pays to add another node to the cluster. (server hardware, hba's, fiber switch ports and most importantly oracle licenses). Interestingly enough PHB's often times choose expanding the infrastructure eventhough the ROI is not there citing opportunity cost, etc.

Another favorite are processes running on new hardware ... the dev just simply did not consider parallelizing the tasks and opted for a monolithic single process approach.

From my experience most of these issues dont go away by horizontally scaling and eventually they come back to haunt you one way or another.

Surprisingly, the added cost of administrating added infrastructure is never part of the calculation ... the only consideration is the cost of hardware.

all jvms except for the 32 bit win32 version default to -server for this type of hardware.
server class hardware for the jvm is defined as min. 2 cpu's and 2GB memory both of which are present in the test machine.

I saw no mention in the article whether the vista install was 32 or 64 bit.

If it was running with default vm options it could make quite a difference on vista.