Disclosed herein are enhancements for operating a content delivery network to load balance origin requests to origin servers. In one implementation, a method of operating a content delivery network comprising a plurality of cache nodes that cache content between end user devices and origin servers includes, in a first cache node of the plurality of cache nodes, obtaining distribution information indicative of how each cache node in the plurality of cache nodes has distributed content requests to the origin servers. The method further provides, in the first cache node maintaining a load balancing profile for the plurality of origin servers based on the distribution information, and distributing a content request to an origin server in the plurality of origin servers based at least in part on the load balancing profile for the plurality of origin servers.