← All hardware
A rack-scale exaflop AI supercomputer that acts as one giant GPU.
Pros
- Unmatched single-domain scale for the largest LLMs
- Mature CUDA software ecosystem with broad framework support
- Treats 72 GPUs as one coherent accelerator, simplifying very large models
Cons
- Roughly $2.8M-$3.4M per rack and severe supply constraints
- Requires liquid cooling and ~120 kW+ power per rack
- Vendor lock-in to NVIDIA networking and software
✓ Where it shines / best for
- Trillion-parameter LLM training at hyperscale
- Real-time inference on massive mixture-of-experts models
- Cloud providers and large AI labs building exascale clusters
✕ Not the best fit for
- Single-server or small-cluster deployments
- Datacenters without liquid-cooling and high-density power
- Cost-sensitive or experimental workloads
Features
- ✓ AI inference
- ✓ Data-center scale
- ✓ HBM3E
- ✓ NVLink
- ✓ Rack Scale
- ✓ FP4
- ✓ LLM Training
- ✓ Grace CPU
- ✓ Liquid Cooled
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| Full NVL72 rack | $2,000,000-$3,000,000 | one-time | Rack-scale system price via OEM/cloud quote; not publicly listed by NVIDIA |
| Cloud rental (per GB200 GPU) | ~$10.50-$27.00 | per hour | On-demand neocloud per-GPU rate; ~$756-$1,944/hr for a full 72-GPU rack; lower with reservations |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| power | ~120 kW+ per rack, liquid-cooled |
| memory | 13.5 TB HBM3e unified (192 GB per GPU) |
| architecture | Blackwell (TSMC 4NP) + Grace Arm CPU, rack-scale |
| interconnect | Quantum-X800 InfiniBand / Spectrum-X Ethernet |
| gpus_per_rack | 72 Blackwell + 36 Grace CPUs |
| fp4_performance | 1,440 PFLOPS (1.44 EXAFLOPS) per rack |
| nvlink_bandwidth | 130 TB/s aggregate (1.8 TB/s per GPU, 5th-gen NVLink) |
Sponsored
A full review is being generated for this product and will appear here shortly.