Performance of cellular networks is severely limited by intense inter-cell interference due to aggressive frequency reuse among cells. How to characterize the dependency of inter-cell interference on positions of BS antennas and users is a key question for the capacity analysis of cellular systems, which unfortunately remains elusive. In this paper, a comparative study on the uplink ergodic sum capacity of cellular systems is presented, where L <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> BS antennas are either co-located at the cell center or uniformly distributed within each cell. With a large number of users, the inter-cell interference density is shown to be inversely proportional to L <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> if the co-located antenna (CA) layout is adopted. With the distributed antenna (DA) layout, it scales in the order of L <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-α/2</sup> , where α is the path-loss factor, and is much lower than that in the CA case when L <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> is large. Substantial gains on the uplink sum capacity are achieved by the DA layout thanks to the reduction of inter-cell interference level. The analysis also reveals that the inter-cell interference density of each BS antenna is sensitive to its position. With the DA layout, BS antennas at cell boundary areas suffer from much higher inter-cell interference than those at the cell center, which may exacerbate the performance disparity of users in cellular systems. To tackle the cell-edge problem, a distributed antenna system (DAS) is further considered, where L BS antennas are distributed over a wide area, and each user chooses V ≪ L surrounding BS antennas as its virtual cell, i.e., its own serving BS antenna set. A uniform inter-cell interference density is shown to be achieved thanks to the adaptive formation of virtual cells. More importantly, by the use of virtual cell, the number of users served by each BS antenna decreases with an increasing L, implying that much of the signal processing and information exchange can be performed in a local and distributed way. The uplink ergodic sum capacity and the ergodic rate with orthogonal access of DASs with V=1 are further derived, and shown to be close to each other even with a large L. It is in sharp contrast to cellular systems where a significant tradeoff between performance and complexity has to be made when the number of BS antennas is large.