How to group inferred instances when computing sustained use discounts

In GCE documentation it mentions “When computing sustained use discounts for predefined machine types, Compute Engine gives you the maximum available discount using inferred instances. An inferred instance combines multiple, non-overlapping instances in the same zone into a single instance for billing.”

But I’d like to know the algorithm to combine those instances, for example, I have multiple projects linking to 1 billing account, does it group the inferred instance by project, or by billing account, or even by admin account?