The articles states that it takes approximately 1000 requests to optimize / warm up. I suspect that they had concurrency set at the default 999, so the first 999 requests would spin up new instances.

Does that mean their 15,000 requests were actually 15 requests spread over 1000 instances?