← All hardware
Blackwell Ultra rack-scale system tuned for the age of AI reasoning.
Pros
- Highest per-GPU memory (288 GB) for long-context reasoning
- ~1.5x GB200 NVL72 AI performance
- Same proven NVLink/CUDA rack-scale platform
Cons
- 132-140 kW per rack exceeds most facilities' capacity
- Multi-million-dollar cost (est. ~$3.7M-$4M+ per rack) and constrained supply
- Mandatory liquid cooling and 800 Gb/s networking
✓ Where it shines / best for
- Frontier reasoning-model inference and serving
- Hyperscale training of the largest LLMs
- AI cloud operators building next-gen capacity
✕ Not the best fit for
- On-premise small clusters without liquid cooling
- Edge or on-device use
- Cost-sensitive workloads
Features
- ✓ AI inference
- ✓ HBM3E
- ✓ NVLink
- ✓ Rack Scale
- ✓ FP4
- ✓ LLM Training
- ✓ Grace CPU
- ✓ Liquid Cooled
- ✓ AI Reasoning
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| Full NVL72 rack | $3,000,000-$4,000,000 | one-time | Estimated; same Blackwell Ultra rack-scale platform as the GB300 NVL72 listing |
| Cloud rental (per GPU) | ~$12.00-$30.00 | per hour | Early neocloud on-demand rate; reserved commitments significantly cheaper |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| power | 132-140 kW per rack, liquid-cooled |
| nvlink | 5.0, 1.8 TB/s per GPU, 130 TB/s aggregate |
| fp4_per_gpu | 15 PFLOPS |
| architecture | Blackwell Ultra (B300) + Grace, rack-scale |
| gpus_per_rack | 72 B300 + 36 Grace CPUs |
| memory_per_gpu | 288 GB HBM3e, 8 TB/s |
| rack_performance | ~1.1 EXAFLOPS FP4 |
Sponsored
A full review is being generated for this product and will appear here shortly.