Actually I've got a customer using Ethereum first and then Algorand to timestamp events but it was only a little more than calling a HTTP API for us.

Sure, you will still have backend processes but the exposure to "the app" is greatly reduced.

IMO there’s room for small benevolent communities where people choose to share in the cost of infrastructure (e.g. hosting seedboxes in torrenting communities where you are expected to keep your up/down ratio even) but expecting users to run full nodes is a tall order, computation and bandwidth-wise you’re offloading costs to your users. I always thought something like wikipedia could work like this since there’s thousands of committed super-users that could afford to run a background process, and indeed there’s IPFS mirrors of it, but IIRC the edits still happen on jimmy wales’ computer.

As for "who's going to run the nodes", if cryptoeconomics or users don't take care of it, I assume businesses will. The nice thing is that regardless, the payment for this infrastructure is also facilitated through cryptocurrency transactions on L2 networks. For someone who doesn't want or need to take ownership of that, it can just be proxied through to the client and not need to be the concern of the developer of publisher of the app.

Of course businesses can still choose to pony up for it proactively of they have the finances to support it, potentially running it themselves or through a third party. You can always fall back to "the old ways" for certain things when it makes sense.

It's orhogonal to consensus so works equally well on PoW/PoS.

There are a couple of competing and complimentary solutions for L2 on Ethereum. I'm quite sure it will converge at some point.