It can be done on a dedicated PC, ideally one that supports h/w virtualization (VT-x and VT-d). People who want to purchase a new PC would need known-good configs. The suggested devices are relatively cheap (no Windows tax, the Dell one officially supports drivers for RedHat Linux) and can be used as PCs. With virtualization support, one could run a local Windows desktop, analytics in a Linux VM, and NAS storage in a separate Linux/FreeBSD VM - all on the same computer. For defensive security, you want to separate a read-only content store from potentially-vulnerable programs which parse & analyze data in the content store.
As someone said in another comment, the challenge with these solutions is making them usable to a mass audience who won't know or care about Linux. Android & OS X both created user experiences that hid the underlying Unix OS. If someone can only afford a single computer, then virtualization allows that device to play the role of both "server" and "desktop".
With local s/w RAID, the system can be designed so that backup consists of (1) shutting down the computer, (2) removing one hard drive, replacing it with another, (3) take the (encrypted) hard drive to offsite location, e.g. trusted friend/family.
> The only thing left is FLOSS version control for sound and video editing
Coud you expand more on this use case? Do you mean archiving binary blobs and storing their metadata in git? git-annex does this and camlistore can be adapted for this purpose. Is this only for local production workflow, or is there a need for remote collaboration on pre-final artifacts?
> an effective "publish to BitTorrent" feature
Could you spec out what's needed? Are there trackers which specialize in public domain content? Strong DRM is coming to browsers and new hardware. While it will create unexpected user experiences and change the web as we know it today, it could increase demand for public domain content. We are going to need distribution channels for public domain content which treat copyrighted content like viruses, i.e. reject before they can enter the channel.
The same fingerprinting techniques used for copyrighted content can only be used to improve discovery of public domain content. This means central directories (which can be cached locally) of metadata and hash/fingerprints for public domain content (text, audio, video, raw data).
Another feature: support for software agents that take action based on computed result of (remote event + private data). A FLOSS version of IFTTT, Yahoo Pipes, e.g. https://github.com/cantino/huginn . A remote event could be a private signal from a mobile device app. This would increase flexibility since the user has a larger private dataset about themself than any web/cloud service. The user can configure FLOSS algorithms, instead of relying on remote black boxes without appeal or transparent governance.