On AWS, whenever the set of nodes changes, or when controller-manager is restarted, controller-manager calls DescribeInstances once for every service to get detailed information about the nodes. This, together with other AWS API calls, causes the AWS API request limit to be hit when you have enough services.

The call chain is

nodeSyncLoop --->
updateLoadBalancerHosts --->
(for each service) lockedUpdateLoadBalancerHosts --->
s.balancer.UpdateLoadBalancer --->
s.getInstancesByNodeNames