It's a statistical procedure, so it applies to whatever data you give it. Like "Mean" could be across samples or within a sample depending on how you use it. For an MA plot, that's average across samples by log-fold-change across groups, so I'd guess you're doing LOESS across all M&A values, that is, pooling samples.

If you wanted to do it sample-by-sample, you wouldnt have the log-fold-change (vertical axis on MA plot).

For some contexts it might be good to do one or the other, but with regard to an MA plot, it must be across multiple samples. Or more precisely, applies to the cohort as a whole, samples are already aggregated.