In any stable system, the average number of items in the system (L) equals the average arrival rate (λ) times the average time each item spends in it (W): L = λW. It holds regardless of arrival distribution or service discipline.
Then
To cut latency (W) without adding capacity, cut work-in-progress (L). Cap concurrent requests — a shorter queue is faster than a faster server.
Unless
Backpressure / Admission Control
L = λW binds the three quantities, so you cannot move one without moving another. Arrival rate is set by your users, not by you. That leaves one lever you actually control: cap the work in progress, and latency falls out of the arithmetic — no faster server required.