VHZN,
I think the problem with building the mesh in Arbitrary mode is that for any resolution you select (high, medium, or low), this step requires the same amount of RAM (as you said, up to 45 Gb for 500 cameras). This is because Photoscan builds the mesh based on the density of the point cloud, then it decimates the mesh to the desired resolution setting you choose. Reducing the resolution of the mesh after you've built the point cloud doesn't use any less RAM.
I believe you have at least three choices: 1) split the model into chunks; 2) reduce the bounding box; or 3) reduce the density of the point cloud before building the mesh.
Since you've already aligned the photos, I'd suggest trying option 2) by copying the point cloud 2x or more, and reducing the bounding box in each copy, so you have two or more slightly overlapping bounding boxes in separate copies of the point cloud. Then process the mesh for each of the bounding boxes separately. This is akin to processing the model in separate chunks but wouldn't require you to realign the photos for each chunk. A possible variant of this method would be to select or turn off photos in different regions of the model in each copy of the point cloud to reduce the amount of data being processed into the mesh, but I haven't tried this. Either way, you'd still need to process the mesh for each region from separate copies of the point cloud, and then merge them using Merge Chunks and selecting "Merge models."
Option 1, splitting the model into chunks, is less desirable because you'll need to realign the photos for each chunk, then align the chunks, and build the mesh for each chunk. This often results in slight discontinuities between the chunks that are difficult to align perfectly.
Option 3, reducing the density of the point cloud, would result in a model with lower resolution, but it might allow you to process it in a single chunk, avoiding the pitfalls of option 1. Often, you don't need the highest density of the point cloud to get a model with acceptable resolution. For this option, try re-aligning the photos at medium or high, instead of the highest accuracy. Each lower setting in the alignment step will decrease the resolution of the model by a factor of 4.
Maybe others have better suggestions. Let us know what method you choose and what works the best for you.