← All hardware
Blackwell Ultra single-GPU module for AI reasoning at scale
Pros
- Major inference uplift over B200
- Huge memory for large context windows
- Broad cloud availability in 2026
- Drop-in for existing Blackwell infrastructure
Cons
- ~$40K-$50K per GPU
- 1000W+ power, liquid cooling recommended
- Allocation still tight at hyperscalers
- Overkill for anything but frontier-scale workloads
✓ Where it shines / best for
- Frontier LLM training and large-scale reasoning-model inference
- Enterprises and clouds building rack-scale GB300 NVL72 clusters
- Memory-bound large-context model serving (288 GB HBM3e)
✕ Not the best fit for
- Budget-constrained or small-scale deployments
- Edge / on-device use (data-center class power and cooling)
- Buyers wanting a fixed public list price per unit
Features
- ✓ LLM
- ✓ API access
- ✓ HBM3E
- ✓ Inference
- ✓ Training
- ✓ NVLink
- ✓ Rack Scale
- ✓ FP4
- ✓ High Bandwidth
- ✓ CUDA
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| GPU / system purchase | Custom quote (~$30,000–$45,000+/GPU est.) | one-time | Sold through OEMs and system integrators in DGX/HGX/GB300 systems; not list-priced individually. Premium over B200. |
| DGX B300 system | Custom quote | one-time | 8-GPU DGX B300 platform sold as a complete system via NVIDIA partners. |
| Cloud rental | Usage-based | per GPU-hour | Available via major cloud providers (CoreWeave, Azure, etc.) at per-hour rates by contract. |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| use | Data center training and reasoning inference |
| power | ~1,100-1,400W |
| memory | ~288GB HBM3E |
| performance | ~15 PFLOPS FP4 (dense), higher sparse |
| architecture | Blackwell Ultra (TSMC 4NP) |
Sponsored
A full review is being generated for this product and will appear here shortly.