Supermicro NVIDIA GB300 NVL72

Example Use Cases:

• Frontier LLM training
  This is the strongest fit because the rack packs 72 NVIDIA B300 GPUs, 36 Grace CPUs, up to 21 TB of HBM3e, and a 1.8 TB/s GPU-to-GPU NVLink fabric—exactly the kind of configuration needed for training very large foundation models efficiently across many accelerators.
• Massive-scale AI inference for enterprise or hosted AI services
  The system is also ideal for serving large models at high throughput, especially when many users or agents must run concurrently, because it combines very large GPU memory capacity with up to 800 Gb/s fabric and rack-scale integration.

• Mixture-of-Experts and multi-model AI platforms
  MoE models and shared AI platforms benefit from large aggregate memory, fast interconnects, and dense GPU pools, all of which this platform emphasizes through 72 GPUs, 21 TB HBM3e, and NVLink-connected compute trays.

• Multimodal model training and fine-tuning
  Workloads combining text, image, video, and speech typically need both high memory bandwidth and large memory footprint, making this system a strong fit given its Blackwell Ultra GPUs with 288 GB HBM3e per GPU and rack-scale liquid-cooled design.

• Large-scale scientific AI and simulation-accelerated research
  The platform is well suited for computational science, digital twins, and physics/biology workloads that use AI plus large datasets, because it provides substantial GPU density, Grace CPUs, high-speed networking, and enterprise deployment infrastructure in a single 48U rack solution.

Download Data Sheet