My understanding is that each user runs a separate Colab instance with its own hardware, it's not the kind of thing you could replicate easily with one server and a couple GPUs.