NVIDIA GB200 NVL72 Rack-Scale Compute Unit

The NVIDIA GB200 NVL72 is a rack-scale AI supercomputing system integrating 36 Grace CPUs and 72 Blackwell GPUs into a single NVLink-connected compute unit. Built on NVIDIA's fifth-generation NVLink fabric, the system delivers 1.4 exaflops of AI compute (FP4) and 720 petaflops of FP8 performance, enabling training and inference workloads at a scale previously achievable only by multi-rack HPC clusters. The 72 Blackwell B200 GPUs within the rack share a unified 8.6 TB aggregate HBM3e memory pool, eliminating traditional PCIe bottlenecks and allowing model parallelism across the entire rack as a single logical GPU domain.

The GB200 NVL72 is engineered for the most demanding large language model (LLM) training and inference workloads, including models exceeding one trillion parameters. Each B200 GPU incorporates the second-generation Transformer Engine with FP4 and FP8 mixed-precision support, dual-chip Blackwell GPU architecture per die, and fifth-generation NVLink providing 1.8 TB/s of bidirectional bandwidth per GPU. The 36 integrated Grace CPUs (Arm Neoverse V2 architecture) are connected to their paired B200 GPUs via NVLink-C2C, delivering 900 GB/s of coherent CPU-GPU interconnect bandwidth — eliminating the PCIe memory copy overhead that constrains conventional accelerated servers.

Designed for liquid-cooled data center deployments, the GB200 NVL72 utilizes a direct liquid cooling (DLC) architecture to manage the thermal demands of rack-scale AI compute. The system is well-suited for deployment in modern AI factories and hyperscale data centers equipped with rear-door or direct liquid cooling infrastructure. NVIDIA's NVLink Switch System within the rack enables all-to-all GPU communication at full bandwidth, making the GB200 NVL72 a foundational building block for sovereign AI infrastructure, cloud service provider GPU clusters, and enterprise-scale AI platforms across the UAE, GCC, EMEA, and APAC regions. Omnixon Global offers worldwide supply of this system to qualified enterprise, government, and research organizations.

Ideal for

• Training and fine-tuning large language models and multimodal foundation models with parameter counts exceeding 100 billion, leveraging the unified 8.6 TB HBM3e memory pool and 1.4 exaflops of FP4 AI compute
• High-throughput LLM inference serving for enterprise AI platforms requiring maximum token-per-second throughput with reduced per-token energy consumption compared to prior-generation GPU systems
• Sovereign AI and national AI infrastructure deployments requiring rack-scale, self-contained AI supercomputing capability with predictable, high-bandwidth all-to-all GPU communication
• Scientific simulation and digital twin workloads in energy, climate modeling, and life sciences that demand extreme floating-point throughput and large unified memory capacity
• Cloud service provider GPU-as-a-Service (GPUaaS) buildouts targeting AI model training and inference services, where rack-level NVLink connectivity maximizes utilization and reduces multi-tenant overhead
• Genomics, drug discovery, and computational biology research requiring simultaneous large-scale data ingestion and deep learning inference pipelines within a single coherent compute domain

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.