NVIDIA · Platform · Rack system
NVIDIA GB300 NVL72
Exact GB300 NVL72 reference specifications and AI deployment guidance, separated from FusionPointAI commercial inventory.
Current
Reference only
Verified technical specifications
| Manufacturer | NVIDIA |
|---|---|
| Exact SKU | GB300 NVL72 |
| Product family | NVIDIA GB300 |
| Architecture | Blackwell Ultra |
| Generation | Blackwell Ultra |
| Entity type | Rack system |
| Lifecycle | Current |
| Launch date | The manufacturer has not published this information. |
| VRAM | 20000 GB |
| Memory technology | HBM3e |
| Memory bandwidth | The manufacturer has not published this information. |
| Power | The manufacturer has not published this information. |
| Interface | MGX NVL72 |
| Form factor | Rack scale · 72 GPU · Platform |
| Interconnect | NVLink 5 |
| MIG / partitioning | Not supported |
| Supported numeric formats | FP64, FP32, TF32, BF16, FP16, FP8, FP6, FP4, INT8 |
| Software stack | cuda, vllm, sglang, tensorrt-llm |
AI workload suitability
LLM inference
High memory capacity supports large checkpoints and multi-GPU inference when runtime and topology are certified.
Fine-tuning and training
Suitable for limited fine-tuning and development when software support, cooling and memory capacity are verified.
Image, video and embeddings
FP16 acceleration supports common generative and vector workloads within the listed VRAM envelope.
Power and cooling
The manufacturer does not publish a directly comparable board-power value for this entity.
Multi-GPU use
NVLink 5 is the recorded interconnect path; topology must still be validated at node level.
Software ecosystem
cuda, vllm, sglang, tensorrt-llm
Product timeline
Media gallery
Official manufacturer source
https://www.nvidia.com/en-us/data-center/products/
Last verified: 2026-08-24