My primary concern with using the top ~1% or so in upper quantile normalization is that it's going to be prone to the same robustness issues that RPKM/FPKMs run in to. That is, if for whatever technical reason you have to have a fair bit of variability in a couple very highly expressed genes (typically rRNAs, but one can imagine other genes) and the set of genes going into the UQ normalization is composed primarily of those then the normalization results will simply reflect irrelevant technical variability.
Of course, this all assumes that it's even appropriate to normalize scRNAseq data with a standard method like this. In some cases it is (e.g., when comparing control vs. treatment with a largely homogeneous cell population), but if you're dealing with very different cell types then I worry that the standard method is going to wash away things like differences in bulk RNA content between cell types. Presuming there are spike-ins, they should have the same normalization applied and then checked to ensure there's nothing greatly amiss. Having said that, if you're going to be doing tSNE and want to look for cell types, then odds are good you'll still be able to see most of them (after all, it's unlikely that RNA quantity is the only difference between them, though not removing that difference would make things easier).
The problem of global shifts in expression
The concept of global shifts in expression is not unique to scRNAseq, in fact the most common example of this is Myc up/down regulation in relation to cancer. To explain the issue, let's suppose that you have the following (very small) matrix of counts:
sample1 sample2
gene1 1 2
gene2 10 20
gene3 15 32
gene4 100 198
As you can see, sample2 has roughly twice as many reads for each gene as sample1, so any normalization is going to cut all of its counts in half (or the reverse for sample1, or perhaps a mixture of the two...depending on the method). That's usually fine, since the assumption with all common normalization methods is that differences like this are simply due to differences in sequencing depth between the samples. But what happens if sample2 actually has twice as much RNA as sample1? Well, then you've just normalized away all of the differences between the samples (I encourage readers to go through the literature on Myc, since that's a very nice example of this, where a mostly-global shift in expression led to apparent down-regulation of genes with constant expression). Such a situation doesn't occur very commonly in bulk RNAseq, but when it does you need to use something like spike-ins (dosed to match cell number or DNA quantity) to maintain the global difference.
If this is only rarely an issue with bulk RNAseq, then why is it more likely to be an issue in scRNAseq? The answer is that bulk RNAseq benefits from the averaging out of many different cell-types, such that ensemble averages/medians don't tend to change much due to treatments...but this isn't going to be the case at the single-cell level, since cells can vary vastly in size (compare a sperm and an oocyte...every RNA should be differentially expressed between them). This difference in base RNA content between different cell types will be washed away by common normalization methods, which can lead to either losing differences or creating differences in genes where there shouldn't be (again, see examples of this in the Myc literature). Whether this really becomes an issue will depend on what your goals are with your data and how much of a global expression difference you actually have and the type of analysis you want to perform.
At the end of the day, just remember that common RNAseq normalization methods assume that (A) there's no global expression change between samples and that (B) if there are very many DE genes that the direction of change is not heavily skewed in one direction. If either of these assumptions are violated then you can run into issues.
