I've tested it out in the modeller pane and checked all the settings and it seems that, yes, there is functionally a limit to the (graphical) extent of your model. I've run through the source code (insofar as I understand it) and I can't obviously see where this is set, but I suspect if you want to change it you'll have to modify the source code and customise the modeller, which would be a lot of effort.
Since you can zoom out to a wider canvas than you can navigate while zoomed in, I would suggest that a more accessible solution is to keep the bottom half of the canvas as a "construction zone" where you build your model, then zoom out, select the relevant parts, and drag them to somewhere that they're more structurally intuitive. AFAIK this should still let you export visual representations (PDF, JPG) of the model without loss of quality, but I haven't tested this.
An alternative and slightly more technical possible solution is to build your model in stages and save each stage as an algorithm. So if, for example, it breaks into three logical sections, you can construct those sections individually, Export > Export as Python Script, then go to Processing > Processing Toolbox > Scripts > Create New Script, and load and save your previously exported script. Once there, it will be accessible under the Algorithms tab in the Graphical Modeller under the (new, if it's your first script) Scripts tab. This way each part of your model will only take up one "box" of space - BUT to edit each of those parts you will need to edit the original model and re-import the script.
Neither of these is a perfect solution, but hopefully they give you a workaround :)