Radio astronomy is undergoing a transition in which efficient extraction of scientific information from ever-larger data volumes, rather than data acquisition, is becoming the limiting factor. The Square Kilometre Array, the next-generation Very Large Array, and the ALMA Wideband Sensitivity Upgrade will produce data at rates that make human inspection, and even human labelling, impractical. In this invited review we argue that self-supervised representation learning provides a general framework for applying artificial intelligence (AI) across the entire radio data lifecycle: from visibilities, through image reconstruction and large-scale survey analysis, to exploration and anomaly detection at the scale of an observatory archive. We illustrate this with existing work on self-supervised learning methods and foundation models, particularly in the radio domain. We then describe how these ideas could be extended to the observatory level through activities at the NSF–Simons AI Institute for Cosmic Origins (CosmicAI). We close with lessons learned, open challenges, and the need for shared radio datasets, benchmarks, and foundation models.