You don’t build an amp model using components or anything like that.
Let’s say you have a physical amp setup in your house, complete with mics and cabinets set the way you want. You then “train” the neural modeler using the training program they supply. I haven’t gone through this process myself, so don’t pretend to know the details, but that’s what they mean by “build a model.”
You can also build such a model using existing models. So let’s say I had a TH-U based rack that I really like the sound of. I could “train” NAM to that rackspace. Then I can share that “model” on the site and anyone else can load that and use it. Sort of like impulse responses in a sense.