For this particular use case scenario, it would be better to skip the EBS disks and use ephemeral disks with instances that are spawned purely for the build and test, check their results back into the build system, and self-destruct. You could even request spot instances since the workload isn't particularly time dependent.
You're right, if Amazon goes down, you're down without much recourse. But if you've designed your system to use instances that are launched on demand, you just launch them in a different availability zone and/or region. The odds that *every* Amazon datacenter goes down at the same time are extremely low.