If you've ever had to scale capacity on AWS (or any other on-demand cloud service, really) you'd know. On the page they say 300 seconds to bring up a Linux instance, but really that can even take longer from time to time.

When you need to scale in a hurry (or in this case semi-automatically), the ability to launch persistent or temporary VM instances in a much shorter timeframe is critical.

Granted (and as they note here) this is sort of demo-only, and probably not particularly useful out of the box. But as a demonstration of how fast things could work it's perfect. It doesn't hurt that it's functional.

Scalable MapReduce, for example, super scalable web servers tolerant for any possible spikes of load, including abuse-resistant hostings, personal virtual appliances... there are many possible usecases, indeed

The folks at ZeroVM have a list of motivations[1] for this type of execution and hosting model. The only problem is that it's hard to tell how active the project is from their main page alone.

EDIT: Their github commit is fairly active for a solo developer[2].

[1] http://zerovm.org/motivation/

[2] https://github.com/zerovm/zerovm/commits/master