ID is the card's address: its host machine's id, "#", and the card's ordinal within that machine ("gpu-1#0"). Stable for as long as the machine is, and the only id a single accelerator has — providers do not name cards.
Memory is the card's VRAM as its own tooling reported it ("122880 MiB") — a display string in the reporter's units, not a byte count. BYO cards carry it (nvidia-smi); Visor's machine object states no VRAM, so a rented card leaves it empty and the console renders "—" rather than a fabricated 0.
Model is the accelerator: the model token read out of the size slug for a Visor GPU droplet ("H100", "MI300X"), or the name nvidia-smi reported for a BYO card ("NVIDIA GB10").
Provider distinguishes a BYO accelerator ("byo") from a Visor-provisioned one (the host machine's real provider). It is what tells a card the org owns from a card the org rents.