The speed issue is not as bad as it sounds. CGI uses several tricky methods to defer compilation until absolutely necessary. In addition, if you use the object-oriented interface, you can save the time required to import several dozen symbols.

If you use only what you need and code reasonably good code, you can probably handle a hundred thousand hits per day. Depending on what you do, CGI::Lite may be a better approach.

The bottleneck may instead be the cost of spawning new processes, in which case a properly set up mod_perl will indeed give bigger benefits. It's still a little experimental on NT, though.