Prior research and applications demonstrate that serverless infrastructure (i.e., Functions-as-a-Service, FaaS) can support data management workloads by eliminating the need to provision and manage resources and enabling unlimited elasticity and pay-per-use. Machine learning models can be trained on serverless infrastructure or on Infrastructure-as-a-Service platforms (e.g., that often run code in containers or VMs). In this paper, Jiang et al. implement a platform to facilitate the fair comparison of these environments for distributed ML training. They find that FaaS is most effective for models with reduced communication that converge quickly. Although FaaS can be faster than IaaS, it is not typically cheaper.