This data is massively outdated... it's before signature caching and ultra-prune, each were easily an order of magnitude (or two) improvements in the transaction dependent parts of propagation delay. It's also prior to block relay network, not to mention the further optimizations proposed but not written yet.

I don't actually think hosts are faster, actually I'd take a bet that they were slower on average, since performance improvements have made it possible to run nodes on smaller hosts than were viable before (e.g. crazy people with Bitcoind on rpi). But we've had software improvements which massively eclipsed anything you would have gotten from hardware improvements. Repeating that level of software improvement is likely impossible, though there is still some room to improve.

There are risks around massively increasing orphan rates in the short term with larger blocks (though far far lower than what those numbers suggest), indeed... thats one of the unaddressed things in current larger block advocacy, though block relay network (and the possibility of efficient set reconciliation) more or less shows that the issues there are not very fundamental though maybe practically important.