The answer depends on several factors:
1. If the same protocol (primers, read length etc.) was used.
2. If you are using 97% OTUs or denoised sequences (ZOTUs) to make the OTU table.
3. If you have enough RAM to analyze all reads together.
If the same protocol was used, I recommend pooling all samples from all runs to re-generate the OTUs or ZOTUs.
If different protocols were used, then it is still recommended to pool all samples if it is possible to do global trimming to the same position in all of the reads. If this is not possible, e.g. if you are performing a meta-analysis of studies which sequenced non-overlapping regions, then closed-reference clustering
is the only option.