So, in its current form, lemmy sends federation packets in serial form. It can send them to multiple instances in parallel, but the feed between any two given instances is serial.

And serial means that the second packet doesn’t get sent until the first packet has been processed. Add in geographic latency, which is relevant at multiple steps of resolving any given AP packet, which adds to the per packet processing time, and now, lemmy.world is producing packets faster than it’s possible for a geographically remote instance to process them, no matter what hardware they’re running on.

The problem would be resolved with parallel sending, but that’s not currently a thing that lemmy allows for, and apparently, is not trivial to implement either.