Who Used the GPU? Building Per Tenant Telemetry, Showback, and Capacity Evidence for AIaaS

This article addresses the challenge of attributing GPU usage to specific tenants in shared AI-as-a-Service platforms. It explains why simple utilization graphs fail for accountability and outlines an evidence chain connecting business ownership to scheduler decisions, workload identity, hardware telemetry, model-server outcomes, and financial records. Key components include NVIDIA DCGM, Kubernetes metadata, Prometheus, Grafana, and a separate financial evidence store for showback/chargeback. The article distinguishes between allocation, utilization, productive work, and financial consumption, providing a framework for defensible GPU showback.