NVIDIA GB300 NVL72

Every number, explained

GPU memory

20,000 GB HBM3e class (72 GPUs, pooled over NVLink)

GPU memory is the workbench — the whole model must fit here or it doesn't run at all. Regular RAM cannot substitute: it sits on a link far too slow for the GPU, which would spend nearly all its time waiting instead of computing.

Memory bandwidth

576,000 GB/s TODO: verify

How fast the GPU re-reads the model from its own memory. To generate every single word, the GPU re-reads the ENTIRE model — all of it, every word. So faster memory means faster answers: double the bandwidth and the model talks roughly twice as fast.

Power draw

120,000 W — about 100 homes (a home draws ~1,200 W around the clock); one day (2,880 kWh) drains about 32 EV batteries (an EV battery holds ~90 kWh).

Electricity burned every second the machine runs — this is a real monthly bill, not a spec-sheet footnote. We translate it below into homes and electric-car batteries so you can feel the size of it.

Form factor

Complete datacenter rack — 72 GPUs wired as one giant machine

The CPU is the traffic director — it does almost none of the AI math. Don't overspend here; the GPU does the work.

Price

~$3,000,000

What it's for: A whole rack that behaves like one computer — for serving frontier-scale models to thousands of users at once.