It' not unlike Pascal's wager.
( P * G) - C = E
Let P be the probability that your graduate student will produce some Great New Thing.
Let G be the perceived value of the Great New Thing.
Let C be the cost of fellowship and support over the course of the degree.
Let E be the expected outcome for the betters.
---
So Unis are sitting here saying, "man I think Bowties is gonna do something game changing - I'm 33% sure of it" and someone else says "I dunno I think catwoman is going to produce some really great things, maybe not game changing but really great - I'm 90% sure of it". So they all debate the numbers and agree. Then they punch in the equation and the highest E gets an acceptance.
Oh then it gets more confusing. Because they have to weigh how competitive you are in other programs and they have to adjust the cost and the probablity that you'll accept with a given fellowship.
I imagine they have a big super computer that does it all.