The author failed to mention which regions these tests were run. GPU availability can vary depending on the regions that were tested for both Cloud providers.

The author linked to the code at the end of the post.

The regions used are "us-east-1" for AWS [1] and "us-central1-b" for GCP [2].

1: https://github.com/piercefreeman/cloud-gpu-reliability/blob/...

2: https://github.com/piercefreeman/cloud-gpu-reliability/blob/...

This is a big missed point.

At work we run some (non-GPU) instances in every AWS region, and there's pretty big variability over time and region for on-demand launch time. I'd expect it might be even higher for GPU instances. I suspect that a more rigorous investigation might find there isn't quite as big a difference overall as this article suggests.

Just remember this is for GPU instances. Other vm families are pretty fast to launch.

FWIW in our use case of non-GPU instances they launched way faster and more consistently on GCP than AWS. So I guess it is complicated and may depend on exactly what instance you are launching.