vRANs are emerging as a key component of wireless cellular networks, making it imperative to optimize their architecture. This paper addresses the problem of vRAN design in its most general form, i.e., minimizing operational expenditures required to serve the expected traffic. Our framework optimizes the number and operation of Cloud Units (CUs), the function split, and the association and routing for each Distributed Unit (DU). We couple a linearization technique with a cutting-planes method to solve the problem (exactly) fast. Our data-driven evaluation reveals that the benefits of departing from single-CU deployments can be as high as 30% for the networks studied, although such gains saturate quickly as we add more CUs.