You can make site servers constrained servers, in which case they only download from unconstrained site servers. An unconstrained site server can download from the NS. A constrained site server must download from other site servers.
Let's say you have 14 sites in North America, one is HQ with 40k nodes, and the other sites are across WAN links and have several hundred to a thousand nodes each. You also have 5 sites in South America, one which is 20k nodes and the others, located across a South America-specific WAN, are 2k nodes each.
You would locate an NS in North America, with perhaps five unconstrained site servers, and place an unconstrained site server at each of the other 13 sites. (They are unconstrained because they download directly from the NS. You do not obtain a bandwidth benefit to the 13 sites by asking them to download from other site servers, because this still consumes WAN bandwidth anyway.)
Meanwhile, because there is no NS in South America, you have placed one unconstrained site server at the 20k node site. You've also placed two constrained site servers there, and one constrained site server at each of the remaining 4 sites. (They are constrained because you want them to download from the nearest, fastest unconstrained site server, which will definitely be the one located in South America. You save on WAN bandwidth because, rather than download all packages 7 times across the WAN link between North America and South America, you download it once for the unconstrained server. Of the remaining 6 site servers, the two at your large site use physical LAN links and the four at your branch offices use locally-available, less-strained South America WAN bandwidth.)
I hope I've represented the concept properly. I do not manage any environments where a constrained site server is required.