8-GPU NVIDIA B300 Blackwell Ultra AI Server (HGX B300)

The NVIDIA HGX B300 platform: 8 Blackwell Ultra GPUs and 2.3 TB of HBM3e memory for AI reasoning and the largest models.

Key features

• 8 NVIDIA B300 Blackwell Ultra GPUs with 288 GB of HBM3e each, for a total of about 2.3 TB of GPU memory
• 5th-generation NVLink and NVSwitch: 1.8 TB/s per GPU
• NVIDIA ConnectX-8 networking up to 800 Gb/s per GPU
• Up to 72 petaFLOPS for training (FP8) and 144 petaFLOPS for inference (FP4)
• 2 Intel Xeon 6 processors depending on the manufacturer, up to 4 TB of DDR5
• Rack form factor, power draw of around 14 kW: data-center-grade power and cooling

Use cases

Very large-scale LLM training and inference, reasoning AI, generative and agentic AI, AI factory.

We also offer this server as a branded model: the ASUS XA NB3I-E12 HGX B300 server.

How to order

The server is built by an NVIDIA partner manufacturer chosen based on availability and lead time. The final configuration is validated with you in the quote.

On quotation: describe your project using the form at the bottom of the page (target models and use cases, number of servers, available power and cooling, desired lead time). We get back to you with a priced proposal and the detailed configuration; the deposit is requested by bank transfer once the quote is approved.

Manufacturer's warranty and support from our team in France.