Yeah, that just depends on what you're trying to achieve. Depending on what kind of AI workload you have, you can scale it across 4 GPUs. Or it'll become super slow if it needs to transfer a lot of data between these GPUs. And depending on what kinds of maths is involved, a Pascal generation GPU might be perfectly fine, or it'll lack support for some of the operations involved. So yes, of course you can build that rig. Whether it's going to be useful in your scenario is a different question. But I'd argue, if you need 96GB of VRAM for more than just the sake of it, you should be able to tell...