It wouldn’t hurt to grab a tcpdump and make sure the timings in the waterfall are real for the individual requests. The timings are captured from within the browser so if the main thread is busy doing something it will make the individual requests look like they took longer than they really did.
It’s actually accurate from the browser’s perspective because that is when the browser serviced the requests but if you’re looking to make a decision about the network performance of a given server it’s worth validating against the actual packet capture (and make sure to run several runs just to be sure you’re not hitting an empty CND node).