I ran your snippet multiple times on different Colab instances. Output is consistent with nvidia-smi output. I see full 12GB GPU RAM available.

Thinking that this might be linked to heavy load on Google data centers in your region ( I am in Bangalore, India btw), I used a US google account from a US machine with Google Colab. Still results were exactly same as above!
I am not sure whats going on :) Have you tried using a different Google ID?