A few months ago I did an analysis of the cost various Google Cloud Platform system configurations for deep learning, and found that due to the price savings offered by premptable VMs, it may actually be more cost-effective to use CPUs instead of GPUs. (HN thread: https://news.ycombinator.com/item?id=14727673)
A 36% price cut, however, might be enough to tip the scales back in favor of GPUs. I’ll rerun the experiment with updated VMs soon and see what happens.