If you have any kind of workload with IO patterns that are not embarassingly parallel you want your files to be in the page caches, i.e. RAM, to minimize the access latencies and thus the walltime of your workload.

For that to work applications shouldn't be gobbling up all the ram, instead they should leave it to the OS to cache data.

Sure, NVMe drives help, but loading things from RAM is snappier still.

Also, those huge memory footprints themselves can also cause issues. E.g. they can increase GC pauses in GCed languages or decrease cache hit rates when allocations are scattered all over the place.