We address the problem of computing the sum, with zero error, of messages that are observed at different sources in a directed acyclic graph network. At each time instant the source messages are i.i.d. uniform over a finite field, and every terminal wants to compute the finite field sum of the messages corresponding to each time instant. We use network codes to solve a sum-network problem, and define its associated computation capacity. In the talk, I will describe a procedure to construct sum-networks using combinatorial incidence structures. We will evaluate the computation capacity of these sum-networks and see that it depends on the characteristic of the finite field alphabet. This dependence is strong; there are sum-networks that have a rate-1 solution over one characteristic but capacity close to zero over a different characteristic.