Hardware catalogue

Five products, in order from smallest to largest. Every number below is paired with what it means and when it matters.

Before you shop: what these numbers mean

GPU memory (VRAM / HBM): This is the memory that lives directly on the GPU chip. An AI model has to fit entirely inside GPU memory to run at full speed. This is the number that decides whether a model runs at all. Too little GPU memory and the model simply won't load, no matter how good the rest of the machine is.
System RAM: The general-purpose memory a computer uses for everything else it's doing — the operating system, other programs, loading files. It is usually much less important than GPU memory for whether a large AI model runs, because the model itself lives on the GPU, not in system RAM.
CPU: The general-purpose processor that manages the machine and feeds work to the GPU. For running a large AI model, the CPU mostly doesn't decide whether it works — the GPU does the heavy lifting. A weak CPU can slow things down at the edges, but it won't stop a model from fitting or not fitting.
Watts (power draw): How much electricity the hardware pulls while running. This matters for your electric bill, your cooling needs, and whether your building's power circuits can even support the machine. It has nothing to do with whether a model fits — that's GPU memory's job — but it has everything to do with what the hardware costs you to run.
NVLink: A fast, direct cable-and-chip connection between GPUs, built specifically so multiple GPUs can share memory and work together quickly. Datacenter hardware (HGX B200, GB300 NVL72) has it. Desktop and workstation cards (RTX 5090, RTX PRO 6000) do not — if you put several of those in one box, they can only talk to each other over the much slower PCIe slot.

GeForce RTX 5090

Desktop card

32 GB (GDDR7) This is dedicated GPU memory (GDDR7). This is the number that decides whether an AI model fits and runs.

575 W Continuous power draw. Doesn't affect whether a model fits — affects your power bill and cooling. See the everyday comparison below.

575 W continuous draw.

Homes: 575 W ÷ 1,200 W per home ≈ 0.5 homes worth of power, all the time.
EV batteries per day: (575 W × 24 hours ÷ 1,000) ÷ 90 kWh per battery ≈ 0.2 EV batteries every single day.

$2,400 Price for one unit.

No NVLink on this card — it cannot be wired to another GPU with a fast direct link.

DGX Spark

Desktop AI computer

128 GB (unified LPDDR5X) This is unified memory, shared between CPU and GPU in one pool — not dedicated GPU-only VRAM like the other products here.

240 W Continuous power draw. Doesn't affect whether a model fits — affects your power bill and cooling. See the everyday comparison below.

240 W continuous draw.

Homes: 240 W ÷ 1,200 W per home ≈ 0.2 homes worth of power, all the time.
EV batteries per day: (240 W × 24 hours ÷ 1,000) ÷ 90 kWh per battery ≈ 0.1 EV batteries every single day.

$3,999 Price for one unit.

Memory here is unified — shared between the CPU and GPU in one pool, unlike the dedicated GPU memory (VRAM) on the other products below.

RTX PRO 6000 Blackwell Workstation Edition

Workstation card

96 GB (GDDR7) This is dedicated GPU memory (GDDR7). This is the number that decides whether an AI model fits and runs.

600 W Continuous power draw. Doesn't affect whether a model fits — affects your power bill and cooling. See the everyday comparison below.

600 W continuous draw.

Homes: 600 W ÷ 1,200 W per home ≈ 0.5 homes worth of power, all the time.
EV batteries per day: (600 W × 24 hours ÷ 1,000) ÷ 90 kWh per battery ≈ 0.2 EV batteries every single day.

$13,250 Price for one unit.

No NVLink on this card either — like the RTX 5090, multiple units only talk to each other over the much slower PCIe slot.

HGX B200 server

Datacenter server — 8 GPUs

1,440 GB total (8 × 180 GB per GPU) (HBM3e) This is dedicated GPU memory (HBM3e). This is the number that decides whether an AI model fits and runs.

14,300 W Continuous power draw. Doesn't affect whether a model fits — affects your power bill and cooling. See the everyday comparison below.

14,300 W continuous draw.

Homes: 14,300 W ÷ 1,200 W per home ≈ 11.9 homes worth of power, all the time.
EV batteries per day: (14,300 W × 24 hours ÷ 1,000) ÷ 90 kWh per battery ≈ 3.8 EV batteries every single day.

$450,000 Price for one unit.

The 8 GPUs inside are linked with NVLink, a fast direct GPU-to-GPU connection built for exactly this job.

GB300 NVL72 rack

Datacenter rack — 72 GPUs

20,736 GB total (72 × 288 GB per GPU) (HBM3e) This is dedicated GPU memory (HBM3e). This is the number that decides whether an AI model fits and runs.

135,000 W Continuous power draw. Doesn't affect whether a model fits — affects your power bill and cooling. See the everyday comparison below.

135,000 W continuous draw.

Homes: 135,000 W ÷ 1,200 W per home ≈ 112.5 homes worth of power, all the time.
EV batteries per day: (135,000 W × 24 hours ÷ 1,000) ÷ 90 kWh per battery ≈ 36 EV batteries every single day.

$3,700,000 Estimated street price for the full rack.

This price is an analyst estimate, not an official NVIDIA list price.

All 72 GPUs are linked with NVLink across the whole rack. The price shown is an analyst estimate, not an official NVIDIA list price.