We detect communities, or groups of firms having mutually close business relationship, in a huge inter-firm transaction network and study their geo-graphical perspectives. We apply a random-walk based method proposed in network science for the community detection to a Japanese inter-firm network having around 0.7 million firms. We found that the community size follows a power-law distribution, and high-ranked communities in terms of size are nonlocalized nation-wide communities. Although the in-dustrial proximity is generally important for the community formation, whether the geographical proximity matters depends on the type of busi-ness. Localized communities tend to be medium-sized. We also shed light on the hierarchical structure of community. We found that some of nation-wide communities are divided into several localized subcommunities.