EKS nodes full but CPU and memory still idle

Your platform team runs a fleet of EKS clusters hosting roughly 200 lightweight microservices (each requesting 100m CPU and 256 MiB RAM). The nodes consistently show CPU utilization around 18% and memory around 22%, yet the pod scheduler reports all nodes at capacity and refuses to place new workloads. Your autoscaler responds by provisioning additional nodes, which also fill up at the same low utilization ceiling. The EC2 bill is climbing even though the underlying compute is largely idle. Engineers have confirmed this is not a resource-request misconfiguration — the pods genuinely fit in available CPU and memory. Leadership wants a clear recommendation before next month's infrastructure review. What architectural change would you bring to that meeting?