Optimal architectures can be situation dependent. I would say you question relies on a few things:
1. Are you heavily taxing any of the machines with either AF or Analytics? If one of these is true, I would say putting it on the other machine is the best approach.
2. Are you looking to put it onto one of the existing machines, or are you open to spinning up 1 or 2 more machines for dedicated PI Web API usage?
1. Are you looking to have high availability of the PI Web API? It's very common for the PI Web API itself to be behind an NLB, so it would make sense to piggy back on the AF implementation and the Web API to those machines.
2. Are you going to heavily use the PI Web API, or is it mainly for the PI Vision search pane (and the indexed search crawler)? For people that are primarily using the PI Web API within PI Vision, then installing the PI Web API directly on the PI Vision server(s) is typically the best approach. If, however, you're going to be using it heavily for a custom application/website, then I would consider dedicated machines for the Web API.
Thanks for the info, Kelsey. I mostly wanted to make sure that it would work OK to have the PI Web API on the NLB AF Server machines, but hadn't been able to find anything that said that specifically.