A very accessible short article by Greg Wilson in American Scientist provides a nice summary of the results of the survey he and his colleagues conducted recently. He concludes that the survey shows two things are clear:
The first is that if funding agencies, vendors and computer science researchers really want to help working scientists do more science, they should invest more in conventional small-scale computing. Big-budget supercomputing projects and e-science grids are more likely to capture magazine covers, but improvements to mundane desktop applications, and to the ways scientists use them, will have more real impact.
My second conclusion is that we’re not doing nearly enough to teach scientists how to use computers effectively as research tools. (…)
If you care about scientific computing the article is well worth your time.
Catenary 
One thing that I love is the graphic power that even a simple PC can deliver using SaaS to explain research finding and statistical data.
http://www.gapminder.org is a site founded by Sweden Hans Rosling. I use their service to obtain data for several projects including one of mu blog’s post.
I recommend is lecture about “Debunking myths about the’third world’” that you’ll find here:
http://www.gapminder.org/videos/ted-talks/hans-rosling-ted-2006-debunking-myths-about-the-third-world/
You’re right, and just like it’s easy to use PCs to visualize information, it’s easy to use them to run relatively modest scripts and models, so there’s rarely a need for most scientists to go and invest their time learning to deal with grids and parallel programming.