Relatedly, I think files / encouraging application layer to marshall/unmarshall from flat bytes was always a terrible idea. IPLD encourages not doing that.
The most important thing is betting people to agree on how we content-address data (see [1]). At that point, we can just let the network protocols duke it out, using bridges to ensure any fragmentation doesn't hurt network affects so bad.
Ultimately, shared addressing scheme -> commodification, something this industry desperately needs, and indeed something humanity needs to make better and fairer use of technology. It's a lot like the identity problem with human-oriented distributed systems, just way easier to solve prior to coordinating better since the actual content addressing (as opposed to agreeing on the scheme) is trustless and coordination free.
[1]: https://www.softwareheritage.org/2020/07/09/intrinsic-vs-ext...
So similar to JSONLD (JSON Linked Data) -- if you know what the schema of a blob it can help with pre-processing... this seems solvable as an add-on but being included is a difference.