Dumb question: How would performance be if you took a used server with like 80 lanes pcie 5 and stuck NVMe on them for model run?

So for LLMs, VRAM speed is king. But what if you bought a used server which had, for example, 80 lanes of pcie 5 available, and you bifurcated that to hold 40 SSDs @ 2x lanes, with each NVMe doing 15Gbps, that means a mirror of 40 2TB drives could potentially do 600Gbps for a 2TB model. Or if you di