Your team runs a GKE Standard cluster hosting a mix of workloads: several CPU-bound microservices, a few memory-heavy batch jobs, and a GPU-based inference service. You currently manage six manually created node pools with fixed machine types, and you notice that many nodes sit at low utilization while some workloads are frequently Pending because no node pool matches their resource requests. You want to reduce this operational overhead and lower cost while ensuring workloads always get appropriately sized nodes. What should you do?