The main problem is that once training is complete, you only need inference which needs far less compute.