AI clusters are heavily interconnected, the blast radius for single component failure is much larger than running single nodes -- you would fragment it beyond recovery to be able to use it meaningfully.

I can't get in detail about real numbers but it's not doable with current hardware by a large margin.