If google, in theory, determined that users almost immediately hit the back button when they hit a 404 page, it would make sense to have the page load & render as fast as possible. Right now, the page size is ~1.3k for me.

I also have a feeling that the 404 may be implemented in a different part of the stack than search. In a very unscientific test of data, the response times for 404 pages according to firebug are at least 10% faster than the normal response on the same subdomain. Also, of all the subdomains I tested, only m.google.com, picasaweb.google.com, and docs.google.com had different 404 pages.

Like I said, this is all blind guessing, but my feeling is that Google has reasons for not better utilizing 404s. Having better 404's isn't rocket science and there are google properties that have them (youtube) so there's certainly reasoning behind it.