rack cluster

NVIDIA GB300 NVL72

NVIDIA's largest rack-scale system, tuned for serving reasoning models at scale

$6,500,000

Industry-reported pricing for the inference-optimized Blackwell Ultra rack; NVIDIA does not publish a list price.

Specs

  • GPUs in this unit72
  • Total GPU memory20,736 GB
  • Memory bandwidth (per GPU)8,000 GB/s
  • Total power draw140,000 W
  • Compute (dense BF16/FP16)270,000 TFLOPS

72x Blackwell Ultra GPU + 36x Grace CPU, NVLink Switch fabric, fully liquid-cooled 48U rack.

What this actually means for you

Our top-of-the-line offering: 72 Blackwell Ultra GPUs with about 20.7 terabytes of combined memory, purpose-built for serving the largest open models to large numbers of simultaneous users at once rather than for training from scratch.

Solo capacity: at the standard FP16/BF16 rule of thumb (2 GB of GPU memory per billion parameters, plus 20% overhead for context), this unit's 20,736 GB can hold an open-source model up to roughly 8,640 billion parameters before you'd need more hardware.

Power & running cost

Continuous draw
140.00 kW
Monthly electricity
$18,950
Everyday comparison
~115.9 homes

Run 24/7, this unit uses 3,360.0 kWh/day — about the same continuous power draw as 115.9x an average U.S. home (~29 kWh/day, EIA). Monthly cost assumes the 2026 U.S. residential average of 18.8¢/kWh.

Request this build

No payment, no account — just tell us your name and we'll follow up with a real quote.