A health system with a research arm had four AI programmes ready — imaging, drug discovery, clinical documentation, and RAG over the medical record — all stopped at the same gate: PHI couldn’t leave the boundary, imaging data was too heavy to move, discovery jobs made costs impossible to forecast, and point-of-care latency had to hold inside the encounter. Rather than negotiate exceptions to move patient data outward, the system moved the GPUs inward. Year one on UPC AI Factory: 42% lower TCO, 40% faster model training on single-tenant RDMA-native fabric, sub-100ms p95 inference latency for documentation and retrieval, 100% of PHI and model embeddings inside the tenant boundary, and a fixed cost per million tokens by programme that made grant planning realistic for the first time.
