[joshvm]

But you would run expensive machine learning jobs on them, and these (GPU instances) tend to offer more significant absolute savings when they're pre-emptible.