DATACENTER
NVIDIA GB200 NVL72 Rack
A single liquid-cooled rack wiring 72 Blackwell GPUs into one machine with nearly 14 TB of pooled memory. This is what frontier AI labs buy dozens of at a time — one rack alone can serve enormous traffic for the largest open models.
Specs, in plain English
- GPU memory 192 GB per GPU × 72 GPUs = 13824 GB total GPU memory is fast on-card RAM that has to hold a model's parameters plus working data while it runs. If a model is bigger than the GPU memory available, it simply will not load.
- Number of GPUs 72 across 1 node How many GPUs work together as one system. More GPUs pool their memory into one bigger pool, which is how you fit models too large for a single card — but they need fast links between them to act as one machine.
- CPU 36x NVIDIA Grace CPUs (integrated on-board) The CPU prepares and feeds data to the GPUs and runs the software around the model. For running big AI models it rarely limits which model you can run — the GPU memory does.
- System RAM 17280 GB System RAM (separate from GPU memory) holds the operating system, the data pipeline, and anything staged before it reaches the GPU. More headroom here means fewer slowdowns under load.
- Power draw 120000 W Power draw is how much electricity the machine pulls when running flat out. It determines your electric bill, and how much cooling and electrical wiring the room needs.
- Price $3,000,000 The purchase price of the hardware itself — it does not include electricity, cooling, networking, or setup labor.
What this means in everyday terms
This machine draws 120000 watts — about like 100 homes running continuously. Left on 24/7 it uses about 2880 kWh a day, which is roughly 32.0 electric-car batteries worth of energy every day.
Homes = watts ÷ 1,200. Daily energy (kWh) = watts × 24 ÷ 1,000. EV batteries/day = daily kWh ÷ 90.