So swapping http://js.arcgis.com/3.6/init.js for http://js.arcgis.com/3.6compact/init.js in a test application increases the amount of http requests by 75... from 132 to 207....yet the total amount of data transferred is 1.5Mb in both cases (the increased number of requests makes up for the smaller size of the build itself).
I believe this speaks to the broader issue that if the CDN is the promoted way to refer to the API, than there should actually be sensible build layers provided that use idiomatic Dojo build practices to reduce HTTP requests.
AMD is a dependency management system which lends itself to good building, but if you don't build AMD slows you down.
I think the JS API team may need to look harder at the issue of build layers and what is best provided by a CDN (ie 4-6 larger files vs. 150 small ones). Also, building small AMD files into layers can reduce them 5-10% just due to inlining, reducing total bytes sent. This would likely be a win for server load also, as there would be less concurrent requests to negotiate.
Alternatively, providing the uncompressed source or even putting the API on github would provide immense benefit to both developers and the API itself. I have fixed several bugs in my own uncompressed source (when it was accidentally released a view versions ago), and I'd love for an easy way to incorporate these back into the community. Doesn't get much easier than a pull request.
Anyhow cheers, and looking forward to more action in this area.