Check out this benchmark where they did an analysis: https://github.com/GoogleCloudPlatform/vertex-ai-samples/blo....
At the bottom, it shows the calculations around the 30% cost efficiency of TPU vs GPU.
Our range of 30-70% is based on some numbers we collected from running fine-tuning runs on TPU and comparing them to similar runs on NVIDIA (though not using our code but other OSS libraries).