A global HTTP(S) external Application Load Balancer fronts a managed instance group (MIG) running a stateless web tier. During flash-sale events, users report elevated latency for several minutes at the start of each spike, even though the MIG eventually scales out. Metrics show that when traffic surges, backend CPU jumps to 95% before new instances become ready. The team wants to reduce the latency during onset without permanently over-provisioning idle capacity. Which change most directly addresses the onset latency while keeping steady-state cost efficient?