Well, using server side calculation, you can just wrap the kernel density tool in a tool service in order to do heat map.
What the silverlight example does, though, is execute kernel density server side and produce output using a png encoder. The intensity slider corresponds to the kernel bandwidth.
I've always been opposed to the concept behind the heatmap api though, because kernel density requires a lot of consideration of the bandwidth and the kernel shape in order to be used appropriately. A Gaussian kernel is not always the right way to go; a fixed kernel is not always best. The bandwidth requires a lot of examination of the dataset to get correct, and a variable bandwidth may or may not be appropriate.
Pushing all of this down into a single set of variables, or even into user side settings when the user does not have a spatial analysis background, is going to put out bad results more often than useful results.
So, it might be better to use server side python scripts to produce the output you want with appropriate decisions in place on the kernel density output (even if esri only has a Gaussian kernel density tool right now - more versions are coming at 10). Push that output to a map service and server that out instead.