The 5G radio access network (RAN) features a partially centralized architecture, in which a subset of the network functions are deployed in a centralized unit. Centralizing these functions reduces operating costs and enables coordination techniques. However, the more functions are centralized, the more capacity is needed on the fronthaul network connecting centralized and distributed units. In addition, the required fronthaul capacity also depends on the instantaneous user traffic, which varies over time. Therefore, in order to optimize its performance, the 5G RAN should be able to dynamically adapt its centralization level to the user traffic. In this paper, we present the design of an adaptive RAN that can switch between two different centralization options at runtime. We provide design objectives and challenges, as well as measurement results from a working implementation.