My assumption would be that it is a complexity thing. As a consumer of the service having a rate limit that is easy to understand and write retry logic for is a big plus. If the criteria is "x requests per 5 minute window" and I start getting rate limit errors it's very clear what back off behaviour I need to implement. If the criteria is CPU usage of my requests, as a consumer it's hard for me to reason about how much CPU a given request is going to take so my retry logic is going to be fairly dumb.