In this paper we present Celleration, a novel gateway-to-mobile Traffic Redundancy Elimination (TRE) system, designed for the new generation of data-intensive cellular networks. Cellular TRE needs to account for the mobile device's limited battery power and the characteristics of the cellular network such as users' mobility, high packet-loss and long round-trip delays. Celleration is based on a novel TRE technique, in which the cellular gateway observes the forwarded chunks to identify the beginning of a previously observed chunk chain, which in turn is used as a reliable predictor to multiple future chunks. These predictions establish an ad-hoc gateway-to-mobile TRE learning mechanism that leverages the gateway's history records and the user mobile device's cached content for an efficient TRE operation for both the backhaul and the wireless last-mile. We present a data analysis of captured cellular traffic from 130 cellular sites and a long-term study of a social network. Finally, we analyze Celleration redundancy elimination and performance under high packet loss.