Ask HN: Does per-second billing make sense for GPU Instance?

Assume that you can provision a GPU instance in EC2 in 10 seconds, and they charge you by seconds, not hours.

Does this make sense to your AI/ML project? If the job takes hours or days to finish?