Using 1.39 of the plugin, the behaviour I observed was that when the instance cap was set to 2 and 2 slaves were already provisioned (e.g. in the stopped EC2 state rather than  terminated), when a slave is required for a pending queued build, the plugin used to simply start one of the 2 stopped instances as required.

Now, on 1.42, the plugin appears to leave the previous 2 EC2 instances in the stopped state on AWS and provision a brand new slave, leading to it exceeding the instance cap of 2 and there now being 3 EC2 instances (albeit 2 stopped and only 1 running).

For now, my workaround was to go back to 1.39.