The biggest problem is that they provide no specs whatsoever and you have to use their cloud compiler for your models. Even if you stockpile the hardware, if Google gets bored and shuts down the service, you can’t train anymore. Google also gets to see every model you ever deploy to the thing.