Is this speculation or proved? It's just as easy to surmise that getting the assets closer to the client will overcome any slow down from an extra connection.
It's speculation to make this claim. The burden of proof should be on them. I'd need to see a benchmark showing that making a new HTTP connection (DNS resolution and all) for a single <100KB file is worthwhile rather than sending it over your already open HTTP 'keep alive" connection.
Asynchronous script loaders like LAB.js (http://labjs.com/) also help alleviate the latency cost of having multiple resource requests by loading them asynchronously.
It all depends on the size of the file and how fast the user's DNS server is.