The paper considers the use of caching in mobile access networks and seeks to evaluate the optimal memory for band-width tradeoff at base station (BS), packet gateway (PGW) and a possible intermediate mobile cloud node (MCN). Formulas are derived for the hit rate under time varying popularity and for a novel cache insertion policy incorporating a pre-filter. The analytical model is applied first to demonstrate that reactive caching is not efficient for nodes with low demand due to the negative impact of content churn. This means BS or MCN caches must be managed proactively with popular content items pre-fetched under some centralized control. Quantifying the tradeoff at each level leads us to conclude that limited caching at BS and MCN levels brings significant savings while to store the vast majority of the content catalogue at the Internet edge, at the PGW or in some higher level shared facility, is clearly cost effective.