Hardware / GPU servers / HGX B300
Spec sheet · Blackwell Ultra · 2025
HGX B300 SXM: 8-GPU Server System
Blackwell Ultra: eight B300 GPUs with 288 GB of HBM3e each, NVLink 5, and 800G ConnectX-8 networking.
Full specifications
Compute
- GPUs
- 8× NVIDIA B300 SXM
Memory
- GPU memory
- 288 GB HBM3e per GPU · 2.3 TB total
Interconnect
- GPU interconnect
- NVLink 5 · 1.8 TB/s per GPU · NVSwitch
- Network fabric
- 8× ConnectX-8 800G
Power and cooling
- Cooling
- Liquid-cooled and air-cooled variants
- System power draw
- 15 kW max (DGX B300)
Platform
- Architecture
- Blackwell Ultra
- First shipped
- 2025
Power, cooling and rack density
| Cooling | Per node | Per rack | Rack power |
|---|---|---|---|
| Liquid | 15 kW | 6 nodes | 90 kW |
| Air | 15 kW | 2 nodes | 30 kW |
A 32-node cluster (256 GPUs) needs about 0.6 MW of facility power: a colocation cage or suite. Size your own cluster in the site planner →
What one node can serve
Largest open models that fit on eight GPUs at FP8 with an 8K context, with estimated output throughput.
- DeepSeek-V4 Pro5,906 tok/s
- DeepSeek-V4.1 Flash12K tok/s
- GLM-5.312K tok/s
- gpt-oss-120b51K tok/s
Cost per million tokens in the inference calculator →
How it compares
| System | Architecture | Memory / GPU | Bandwidth | FP8 dense | Max power | Cooling |
|---|---|---|---|---|---|---|
| HGX B300 | Blackwell Ultra | 288 GB | 8 TB/s | 4.5 PF | 15 kW | Liquid / Air |
| HGX B200 | Blackwell | 180 GB | 8 TB/s | 4.5 PF | 14.3 kW | Air / Liquid |
| HGX H200 | Hopper | 141 GB | 4.8 TB/s | 2.0 PF | 10.2 kW | Air / Liquid |
| HGX H100 | Hopper | 80 GB | 3.35 TB/s | 2.0 PF | 10.2 kW | Air / Liquid |
| HGX A100 80GB | Ampere | 80 GB | 2 TB/s | — | 6.5 kW | Air |
| RTX Pro 6000 Blackwell | Blackwell | 96 GB | 1.8 TB/s | 1.0 PF | 6.5 kW | Air |
How it connects to the cluster
Each GPU gets its own network card and its own port on the leaf switch, one rail per GPU. These are the parts on that path.
On this path: ConnectX-8 SuperNIC · Quantum-X800 Q3400 · Spectrum-4 SN5600
Alternatives to consider
What to verify before you buy
- Delivery window and cancellation terms in writing
- Cooling: confirm whether the system is liquid- or air-cooled, and that the site supports it
- Power: up to 15 kW per node; plan PDUs and UPS for it
- Network: ConnectX-8 800G needs 800G-capable switches and optics
Export controls
B300 carries the most advanced export controls in the NVIDIA lineup. Consult legal counsel before placing any allocation order with a non-US delivery address.
Frequently asked questions
How much power does an HGX B300 draw?
Up to 15 kW per 8-GPU HGX node at maximum rated power. Plan site power for that peak: training runs close to it. Inference typically averages 55–65% of it.
Is the HGX B300 air-cooled or liquid-cooled?
Both. HGX B300 systems ship direct-to-chip liquid or air-cooled. Direct-to-chip liquid racks typically hold 6 nodes (about 90 kW). Air racks typically hold 2 nodes (about 30 kW).
How much GPU memory does an HGX B300 have?
288 GB per GPU, 2,304 GB across eight, at 8 TB/s per GPU.
What models can one HGX B300 node serve?
At FP8 with an 8K context, the largest open models that fit on eight GPUs include DeepSeek-V4 Pro, DeepSeek-V4.1 Flash, GLM-5.3, gpt-oss-120b. The inference calculator shows throughput and cost for each.
What kind of site does an HGX B300 cluster need?
A 32-node cluster (256 GPUs) needs roughly 0.6 MW of facility power with direct-to-chip liquid cooling: a colocation cage or suite.
What comes after the HGX B300?
The Vera Rubin. The HGX B300 is Blackwell Ultra architecture, first shipped in 2025.
Sources: NVIDIA HGX platform · NVIDIA DGX B300 user guide. Power is the maximum rated system draw; throughput and site figures are planning estimates from the same models as the Rack Rate tools.
Building with HGX B300?
Send us the cluster spec. We source the full bill of materials, and can finance it.