The problem is that model doesn't work very well for probably the most common computationally intensive task consumers do: games.
It also doesn't work well with tasks involving non-trivial datasets. If you've got say 20GB of video from a Gopro you want to transcode, shipping it out to a datacenter is going to be really painful.
I'd say it's even more likely that anything computationally expensive won't be done in the same room with you to begin with 10 years from now, but will be done by an anonymous compute node in a datacenter that you're leasing for the 5 seconds you're using it.