Spam in OP_RETURNs can *decrease* the load on nodes, as the OP_RETURN data can be discarded ('pruned') by the nodes.

The best block, from a storage point of view, is filled with just one transaction that has a huge OP_RETURN taking up the entire megabyte.

(Even better is if the transaction takes many UTXOs as input, and has just a few outputs. But that's a slightly different issue)