I view the favorable performance characteristics of IPFS at scale (over the current centralized client-server architecture of the present web) more as a general symptom of the pathology resulting from the incomplete, ill-planned architecture of the web in general.

The greatest advantage of an architecture like IPFS instead lies in its friendliness to more a democratic, semantic web, in which users and programs may make use of URI's at a fine-grained and peer-to-peer level. If we can decouple resources from a central server, and then build programs around these static resources, the web will not only become more permanent, but also more robust against walled-gardens robbing us of programmatic control of those resources.

To paraphrase the Plan 9 authors[1], it might make sense to say that a p2p, content addressable web should try to build a web out of a lot of little systems, not a system out of a lot of little webs. (Well, that doesn't completely make sense, and I'm starting to mix my metaphors, but my point is that what we have now is significantly hindered by lack of fine-grained semantics allowing interop. Hypermedia has the potential to be so much more than simply manipulating the DOM with some javascript!)

[1] http://www.plan9.bell-labs.com/sys/doc/9.html