I'm not awake/invested enough to make up numbers and do the math here, but shouldn't the cost for provisioning workers and value provided at various latencies matter? Like, each worker you provision costs a certain amount and shifts the latency distribution towards zero by a certain amount, and in principle there should be some function that converts each distribution to a dollar value.

Like, suppose you had a specific service level agreement that requires X% of requests to be served within Y seconds over some time interval, and breaking the SLA costs Z dollars. Each provisioning level would generate some distribution of latencies, which can get turned into likelihood of meeting the SLA, and from there you can put a dollar value on it. And crucially, this allows the "ideal" amount of provisioning to vary based off the relative cost of over and under-provisioning; if workers are cheap and breaking the SLA is costly, you would have more workers than if the SLA is relatively unimportant and workers are expensive.