In order to deliver the high data rates promised for 5G networks, mobile base stations need to be deployed in dense layouts. This results in increased inter-cell interference, which can be mitigated by leveraging centralized architectures in radio access networks. Nonetheless, centralizing all the processing requires prohibitively high link capacities for the fronthaul network connecting centralized and distributed units. In contrast, a static, partially-centralized architecture yields poor performance as it fails to adapt to instantaneous interference situations. In this work, we show that a dynamically centralized architecture enables drastic interference reductions even when using a very limited fronthaul network. We propose multiple algorithms to find the optimal centralization option and evaluate their performance on operator-grade hardware. In addition, owing to the dynamicity of the problem being solved, we provide a framework to decide on the best algorithm based on the trade-off between performance, cost, and adaptation time.