On-demand data centers host multiple applications on server farms by dynamically provisioning resources in response to workload variations. The efficiency of such dynamic provisioning on the required server farm capacity is dependent on several factors --- the granularity and frequency of reallocation, the number of applications being hosted, the amount of resource overprovisioning and the accuracy of workload prediction. In this paper, we quantify the effect of these factors on the multiplexing benefits achievable in an on-demand data center. Using traces of real e-commerce workloads, we demonstrate that the ability to allocate fractional server resources at fine time-scales of tens of seconds to a few minutes can increase the multiplexing benefits by 162-188% over coarse-grained reallocation. Our results also show that these benefits increase in the presence of large number of hosted applications as a result of high level of multiplexing. In addition, we demonstrate that such fine-grained multiplexing is achievable even in the presence of real-world (inaccurate) workload predictors and allows overprovisioning slack of nearly 35-70% over coarse-grained multiplexing.