OpenAI published a follow-up on scaling Habitat’s Azure Cosmos DB storage layer. A shared container was vulnerable to hot keys and noisy neighbors, while dedicated containers pushed capacity-planning burdens onto individual product teams. Hot-key throttling, request priorities and load-based partition splitting enabled a return to shared tenancy by default. After large-scale failover problems, OpenAI and Azure also shared production-scale test environments, metrics and direct response channels. The case illustrates how managed services can require joint engineering and operations at extreme scale.