Your team of data scientists occasionally needs access to a Google Kubernetes Engine (GKE) cluster under your management. They require GPU resources to execute certain lengthy, non-restartable tasks. To manage this while minimizing costs, what approach should you take?