Methods, systems, and computer-storage media for intelligently creating and using routing connections between on-premises servers within a client and nodes of a data center are provided. When routing connections being requested by the client are not presently established, a caching service within the data center is programmed to direct a client's cache request over available routing connections, thereby scaling out storage capacity across the data center. When routing connections being requested by the client are established, a load balancer within the data center is programmed to direct the cache request over a requested routing connection, thereby bypassing intervening hops within the data center, such as gateway nodes that host instances of the caching service. To bypass intervening hops, the client network is provided with access to a client-side routing table for identifying an appropriate network key and a connection table for identifying a routing connection mapped to the network key.