Typical answer for the freemium customers. Not only did the OP say that he read that guide already in detail, but you ignored the fact that he said each were running on their own server.
The OP is 100% legit in asking this question, as there is no recommendation for the Graylog Server itself, nor is their any distinction between the ES_HEAPSIZE and JVM_HEAPSIZE in the wiki. While it’s a starting point, the lack of detail is likely intentional.
The Graylog instance, by itself, doesn’t need as much heapsize as the ES instance. Giving each 6GB/12GB is a fine configuration, although you may consider doubling the RAM on the ES server to 12GB/24GB, if you plan on scaling this out to even a medium-sized company. You may find that GL works just fine with 4GB RAM dedicated to it, and that anything above this has very little impact on performance. At a certain point, the ring size, input/output buffers and underlying hardware, including but not limited to the number of processors, play a lot larger role in the overall performance and stability of your system.