My idea is that from the moment that network propagation delays between (important) miners matter, they will form a strong backbone network, because they have all mutual incentives to do so.  If you invest in millions of hardware, you can invest in a strong network link to another big mining pool, *and that pool has also an advantage to set up a strong link to yours*.

This is why I see a PoW system as a strong (almost) fully interconnected backbone of big mining pools (I don't know, 10, 20, 5, 50, whatever), and all the other nodes connecting to them : these back bone miner pool data centers are also interested to get directly the user's transactions (and maybe keep the most juicy ones for themselves unless the user also forwards it to others).

In other words, PoW doesn't favor a P2P network at all, but a big backbone of strongly interconnected miner data centers, serving directly the users ; those funny users that really want to, can use other user's nodes as P2P proxy servers instead of connecting directly to one of the miner pool data centers.

"decentralisation" then simply consists in hoping that the bosses of the miner pool data centers don't collude too much.  This is not impossible, if they are 15 or so, in different countries.

The strong backbone network between important pools is unavoidable, because it is mutually advantageous.  It also kills the idea of selfish mining, because selfish mining counts on your selfish block overtaking potentially the public block when you learn about it: on a strong fully connected back bone, that is impossible: by the time YOU learn about the public block, and you want to publish your selfish chain, all others ALSO received the public block (with their direct links) and your selfish block ALWAYS comes late.  So when network delays matter, automatically a strong backbone will emerge.