Serving an image over HTTPS implies initiating the TCP connection (which requires minimum 3 packets) and the TLS connection (which requires many more, lets say 10).
CDN <-> PoP <-> Datacenter communication doesn't require initiating connections. They reuse them because they centralize requests to serve different users (this is even explicitly called out in the article).
Lets say your closest datacenter is 50ms away, and the closest PoP/CDN node is 10ms. Just initiating the image download through HTTPS all of sudden is 500ms vs 100ms.
Sure, PoP/CDN might need to go to the datacenter to fetch the image (and only if the content is not cached there already) but that only happens once before it gets cached, and there's still a lot of ms to use on that to make the tradeoff worth it.
CDN <-> PoP <-> Datacenter communication doesn't require initiating connections. They reuse them because they centralize requests to serve different users (this is even explicitly called out in the article).
Lets say your closest datacenter is 50ms away, and the closest PoP/CDN node is 10ms. Just initiating the image download through HTTPS all of sudden is 500ms vs 100ms.
Sure, PoP/CDN might need to go to the datacenter to fetch the image (and only if the content is not cached there already) but that only happens once before it gets cached, and there's still a lot of ms to use on that to make the tradeoff worth it.