Do we have any reason to believe spanning trees will scale to more than a few thousand users? I don't think they even claim that.
In any case, DHTs have a much easier task if they're storing one address for each node in the network. On IPFS it would be an address for each file (or worse if they're chunked). It cannot work.
I wish they put more effort in explaining how their routing stuff works, since that is really the only thing that matters. The last time I checked they had this thing called "bitswap" that in practice meant that every node was at all times sending a bazillion messages to every other node announcing what files they had, meanwhile I couldn't send a file between two computers in the same LAN even after I manually told one to connect to the other with an IP address and node pubkey. It was ridiculous.