In this paper we develop a new notion called the end-to-end capacity in terms of input and output processes of an end-to-end path. The end-to-end capacity is defined for a path of interest, and we show that the end-to-end path can be represented by a single-node model with the end-to-end capacity in the sense that the single-node model is equivalent to the original path in terms of the queue-length and departure traffic. This allows us to estimate the end-to-end delay distribution based on endpoint measurements. We also investigate the applicability of this approach to admission control. This is the first attempt to estimate the end-to-end delay distribution itself. 1