30% is a conservative estimate (to be precise, we went with this benchmark: https://github.com/GoogleCloudPlatform/vertex-ai-samples/blo...). However, the actual difference we observe ranges from 30-70%.
Also, calculating GPU costs is getting quite nuanced, with a wide range of prices (https://cloud-gpus.com/) and other variables that makes it harder to do apples-to-apples comparison.