Given the nature of averages these remaining 5% of requests would have to have insane response times. Let’s say, charitably to your argument, that 95% of requests take 20 ms and that all remaining requests take 1000 ms, that’s still an average of 69 ms.
This is all ignoring that “95th percentile” doesn’t mean all of those requests will be that exact number. Some of them will be lower, making the average even lower.
