I am a heavy skeptic of this thing, but I can see a good use case for it: S3 I/O, ephemeral compute (1 instance), versioned blobs. The first two allow you to abstract the data away from the compute (flexibility), and the third lets you recover from mistakes or bugs quicker (or do immutable migrations easier).

I think the devil's in the details though. I expect a high number of unusual bugs due to the novel code, networking, and multiple abstractions. I'd need to trial it for a year before I called it reliable.