Every AI workload the health system proposed — clinical summarization, RAG over the record, imaging support — cleared its clinical evaluation and then stalled at information governance for the same reasons. Using external model services would move PHI outside the boundary; prompts and embeddings included. Fine-tuned weights derived from patient records were ungoverned. Retrieval agents had no enforceable identity or execution limits. And the system couldn’t evidence which data trained which model or who approved it. UPC Sovereign’s AI-Optimized edition placed dedicated GPU clusters inside the sovereign perimeter — training and inference where the data already sits, derived artefacts carrying the same residency rule as the records they came from, bounded agents with enforceable execution limits, and data lineage documented for EU AI Act mapping.
