Some people will care a lot about keeping their data on-prem, but many others probably won't, and the former can then resell their spare capacity to the latter.

Decentralized storage is much harder, since there reliability matters a lot more as it's inherently stateful. You have to assume data loss, so you have to replicate everything; with inference, you only have to spend extra resources at failover time. Also storage can't be time-shared in the same way as compute; if it's full, it's full even when not actively accessed.

I mean I guess there's nothing left which can settle this disagreement except to see how it turns out. I'll just restate my opinion that, from a user experience and product perspective, decentralized model hosting will look like using the big labs, except worse in every way. I don't expect it to please either the people who want local control or the people who are happy using the products from the big labs now.