Two identical cleartext values would likely not encrypt to the same ciphertext value (for example, you could get around that easily on the client end by simply incrementing any duplicate value by 1 before sending and then decrementing it by 1 again on return, assuming that is done undoably given the other operations happening); any comparison operation would also likely be encrypted and thus unknown to the server; so the server couldn't just linearly compare any two encrypted values to make deductions.
Usually the cloud will have access to your model. That poses a problem if your model is highly sensitive. (Imagine the NSA wanting to run a model on North Korean servers. NK would immediately snatch up that model.)
With FHE, you can theoretically avoid that. Someone can upload an encrypted model to the cloud. The cloud can do some computation on it (inference) and deliver an encrypted result. Then you can decrypt the result in the comfort of your own government^Whome.
Obviously this is a bit of a stupid example, but just think of all the scenarios right now where you'd want to offload your computation on someone else, but you don't want to let them see the computation.