Cold Start refers to the latency period from when an instance is spun up, the container is loaded, and the model weights are loaded, until it is able to respond to inference requests.