The beauty is that IA could offer their own distribution of IPFS that uses their own DHT for example, and they could allow only public read access to it. This would solve the slow part of finding a file, for IA specifically. Then the actual transfers tend to be pretty quick with IPFS.

What's the point of using IPFS then? Others can still spread the file elsewhere and verify it's the correct one, by using the exact same ID of the file, although on two different networks. The beauty of content-addressing I guess.

That isn’t solving the problem, it’s just giving them more of it to work on. IA has enough material that I’d be surprised if they didn’t hit IPFS’s design limits on their own, and they’d likely need to change the design in ways which would be hard to get upstream.

Several minutes sounds more than fine for this purpose ?

Especially if it's about having an Internet Archive backup.

I think the point is that it's already slow at the current amount of data, let alone when you stuff dozens more PB into it