I don't see any reason why someone should train a 1T model, they probably don't have a big enough dataset to make it worth it, so with the same compute or less

I don't see any reason why someone should train a 1T model, they probably don't have a big enough dataset to make it worth it, so with the same compute or less you can get a better model by training a smaller one (the chinchilla scale).