In Phase 1 we could reliably support a few dozen users on a single reclaimed server, but user process CPU and memory contention made the need to shift users to Cori more pressing. We set aside a single 32-core, 500 GB login node for "shared" Jupyter notebooks but have expanded to four nodes and are running out of hardware options on Cori. In 2020, we instituted resource limits via Linux control groups (cgroups) ensuring overall quality of service for shared-node use, and batch-based "exclusive node" notebook sessions for users with needs exceeding those limits. The lesson here is that HPC centers interested in data and AI user support must make proactive plans centered on Jupyter to avoid playing catch-up.