Unfortunately, it does depend. It depends at least on the enrichment of the antibody over background and on the binding pattern of the TF across the genome. Antibodies with lower (only a few fold) enrichment or TFs with a promiscuous binding pattern (bind all over the place) will require more reads. However, without any other information, a few million reads (a lane, for example) is a place to start. From there, you can do some virtual experiments to look at the added benefit of more reads by subsampling your data at various levels and observing the increase in numbers of sites picked up with increased coverage. If you see a plateau in the sites versus reads, then you are probably close to the number of reads that you need.