No GPU compute infrastructure existed. - The data warehouse was a traditional SQL Server-based system optimized for reporting, not for ML training workloads. - No experiment tracking, model registry, or model serving infrastructure was in place.