Just about every service I've ever monitored has had some fairly clear inflection points where higher utilization starts to affect performance. It could be things like more time spent on garbage collection, or just unlucky collisions with async work that make things take longer.