Implicit in investments in research data infrastructure is the assumption that data are valuable entities worth preserving, stewarding, sharing, and reusing. This value proposition also implies that research data are useful to others and that others will reuse those data. However, neither outcome is assured. Data practices are local, varying from field to field, individual to individual, and country to country. As the number and variety of research partners expands, so do the difficulties of sharing, reusing, and sustaining access to data. Efforts to develop global research infrastructures are hindered by communities’ lack of agreement on data management practices –or on what constitutes ‘research data.’ This talk argues for a broader focus on knowledge infrastructures, which are robust networks of people, artifacts, and institutions for producing, exchanging, and sustaining knowledge. Technical aspects of infrastructure, from persistent identifiers to compute capacity and storage, are easier to address than are social aspects, such as data stewardship, trust, governance, economics, infrastructure, standards, and science policy. Infrastructures can connect communities when they support local practices, and disconnect communities when they create incompatible silos. Examples are drawn from several decades of empirical research with research communities in environmental sciences, sensor networks, astronomy, biomedicine, social sciences, and digital humanities.