Three Business Lines, One GPU Estate They Can Govern

A global bank had the GPUs. What it lacked was a way to decide who got them, prove who used them, or tell finance what an inference actually cost. Fraud scoring, the relationship-manager copilot, and risk model training all ran on the same private estate — with long training runs crowding out latency-bound fraud scoring, allocation by manual ticket, and one infrastructure bill nobody could attribute. UPC VEKTOR put a control plane over the whole estate: SLA-aware placement protecting fraud inference from training contention, per-business-line tenancy and entitlements, and chargeback per model that turned an unallocated central cost into budgets each line of business owns.