Scaling parts of the application independently from each other. If traffic spikes in one part of the application (e.x.: your image categorization web app was posted in a cat forum) more nodes can be added to that one resource to handle the load. Another reason could be worker processes that ran intermittently at midnight every day (e.x.: to email users top pics they missed that day).

The difference is in the network boundaries. I write most of my code as microservices, but it would be incredibly easy to wrap it all up into a single binary.

It's not uncommon for go code to have one binary for separate microservice functions (see OKLog). This way you can change between a microservice, or a monolith (ish) with a CLI flag.