Research and the Clinic, Sharing One Estate Safely

Imaging, drug discovery, clinical documentation, and record retrieval all competed for the same GPU estate — and clinical workloads were losing. Multi-day discovery training runs were pushing documentation latency past the encounter. Capacity allocation was informal negotiation between teams. Governance was manual evidence-gathering that took weeks. And GPU consumption couldn’t be attributed to a specific research study or grant. Rather than physically split the estate and strand capacity on both sides, UPC VEKTOR introduced a control plane that guarantees clinical inference priority by policy, isolates clinical and research as distinct tenants with separate entitlements, and attributes GPU-hours and token consumption by study and program — turning research computing from an overhead pool into accountable, fundable capacity.