This post explains why we still run AI workloads close to the database in some production setups. The focus is on data gravity, latency, compliance, and operational cost, not on rejecting the cloud outright. Cloud works well for many teams, especially early on. This piece documents cases where on-prem or hybrid setups stayed cheaper, more straightforward, or more predictable once vector search, embeddings, and retrieval moved into the data plane. Feedback and counterexamples welcome.