Surprisingly not so much - about 9 gigabytes of the allocated 32. at the same time LW consumes 50 gigabytes on this scene.
But the buffer loading takes about 6-8 minutes for Octane. And if you use LW native renderer, the loading is done in less than a minute.
BTW, the scene is - https://www.nonecg.com/3D-products/tokyo-shibuya/ - very good test for hardware.
The scene for 3DMax's looks like this. But for the LW (I was working for the NONECG), I've optimized it and converted most of the trees into instances.