[hashmp]

There is an argument that block propagation and orphan rates acts as natural barrier to the blocksize.

With a very large blocksize limit, miners who produce blocks which are very big, run the risk of a smaller block from a competing miner winning the race, as it is faster to validate and propagate over the network.

This limiting factor should mean the blocksize converges to a size that the network can comfortably handle.

[dgacmu]

This isn't fundamental. The large-block miner can pre-stage copies of the block (before the nonce has been found) at a few replicas. If they successfully find a nonce, they blast the nonce to all of their replicas, which immediately disseminate the complete block.