I was thinking along the lines that if you are sending bigger packets then any new data has longer to wait for that packet to finish being sent, before another packet can be. Apparently this is
indeed a thing.
Also like I said before, if you lose a packet you have potentially a LOT more data to re-transmit, though obviously low latency protocols would just continue to use smaller packet sizes to avoid this. But if those smaller packets are stuck behind a queue of larger packets, you're so out of luck unless the router prioritises smaller packets, and then you have the issue that you've negated the speed benefit as larger packets have to wait for smaller packets. Then what's to stop me using smaller packets to get a leg-up on the delivery time?
I get this somewhat applies today as smaller packets are often prioritised, but were talking much smaller packets where if you chose to use those deliberately you would be slowing things down anyway, whereas using 1500 to get priority over say 9000, would likely be worth it.
But the big problem why it probably wont ever happen is while it reduces the CPU load on the client and server, those are the cheapest bits of hardware to upgrade. As there will always be legacy hardware on the Internet, routers having to fragment those packets would likely negate or exceed any benefit from those frames being more efficient when not fragmented. As you know, core routers are darned expensive so anything that might increase their CPU load is a bad idea, its computationally harder to fragment a larger packet than it is to just deliver several 1500 packets equalling the same total size.
The general wisdom these days seems to be to not use jumbo frames, even on a LAN, with the possible exception of over dedicated iSCSI links. Its more trouble than its worth. Now try multiplying that trouble by the entire Internet.
If you never ever were going to lose any packets and your interconnects never have contention, sure it would be more efficient, but that's not how the Internet works. We only need to look at ping graphs to see how often small ICMP packets fail to arrive in a timely manner, even on a connection that is idle.