IPFS has absolutely nothing whatsoever to do with Ethereum, or indeed any blockchain. It is a protocol for storing distributing and addressing data by hashes of the content over a peer to peer network.
There is however an initiative to create a commercial market for “pinning*”, which is blockchain based. It still has nothing to do with Ethereum, and is a distinct project that uses IPFS rather than being part of the protocol, thankfully. It is also not a “proof of work” sort of waste, but built around proving content that was promised to be stored is actually stored.
Pinning in IPFS is effectively “hosting” data permanently. IPFS is inherently peer to peer: content you access gets added to your local cache and gets served to any peer near you asking for it—like BitTorrent—until it that cache is cleared to make space for new content you access. If nobody keeps a copy of some data you want others to access when your machines are offline, IPFS wouldn’t be particularly useful as a CDN. So peers on the network can choose to pin some data, making them exempt from being cleared with cache. It is perfectly possible to offer pinning services that have nothing to do with Filecoin or the blockchain, and those exist already. But the organization developing IPFS wanted an independent blockchain based solution simply because they felt it would scale better and give them a potential way to sustain themselves.
Frankly, it was a bad idea then, as crypto grift was already becoming obvious. And it didn’t really take off. But since Filecoin has always been a completely separate thing to IPFS, it doesn’t affect how IPFS works in any way, which it continues to do so.
There are many aspects of IPFS the actual protocol that could stand to be improved. But in a lot of ways, it does do many of the things a Fediverse “CDN” should. But that’s just the storage layer. Getting even the popular AP servers to agree to implement IPFS is going to be almost as realistic an expectation as getting federated identity working on AP. A personal pessimistic view.
Ironically, because there’s no UDP in browsers, we can’t actually get proper p2p on the web. WebRTC through centralized coordination servers at best. Protocol Labs has all but given up on this use-case in favor of using some bootstrapped selection of remote helper nodes.