We did observe that the availability for spot on GCP is much better than on-demand ones, especially for high-end GPUs, e.g., A100, A100-80GB. For on-demand GPUs, it seems the availability is quite volatile, and it seems GCP has the best availability among the big clouds, while some other clouds, such as Oracle or IBM cloud can also have some GPUs available. We are also adding support for coreweave, fluidstack and runpod, so more GPUs to come.