NVIDIA GB300 NVL72 on Nebius AI Cloud
Run frontier AI workloads at rack scale on NVIDIA’s most advanced Blackwell Ultra system, deployed in production on Nebius AI Cloud.
What makes GB300 NVL72 on Nebius different
GPU performance, optimized in-house
Our GPU clusters are optimized across every layer of the stack — from in-house designed servers to a software layer validated against NVIDIA benchmarks — so your AI workloads run at the highest achievable performance.
AI without operational overhead
We take care of the infrastructure, at any scale. Deploy large-scale training clusters or high-throughput inference systems on Slurm or Kubernetes, without managing rack-level hardware, cooling, or network configuration.
Security and reliability by design
Every Nebius cluster comes with auto-healing that detects and recovers from hardware failures with minimum possible interruption. Our platform is also built on industry security and compliance standards, so your data and workloads always stay secure.
What teams run on NVIDIA GB300 NVL72
Frontier model training
Train the largest and most complex AI models, including MoE models at frontier scale, on a platform with 37 TB of fast memory and 1,440 PFLOPS of FP4 compute per rack. GB300 NVL72 provides the memory capacity and interconnect bandwidth for model sizes and training configurations that go beyond the limits of node-level platforms.
AI reasoning and agentic AI
Run AI reasoning and agentic workloads that require test-time scaling, where models evaluate multiple candidate responses before producing an output, demanding up to 100× more compute than traditional inference. GB300 NVL72 is purpose-built for this class of workload, with an attention engine delivering 2.5× faster attention performance compared to NVIDIA Hopper.
Real-time inference at scale
Serve AI models at maximum throughput and efficiency. The GB300 NVL72 rack-scale architecture is optimized for high token-per-second output, with the memory capacity to hold large models and long KV caches entirely in GPU memory, enabling low-latency responses at scale without offloading.
NVIDIA GB300 NVL72 specifications
Specification
GB300 NVL72
Form factor
Fully liquid-cooled, rack-scale system
Blackwell Ultra GPUs | NVIDIA Grace CPUs
72 | 36
CPU cores
2,592 Arm Neoverse V2 cores
Total FP4 Tensor Core (sparse | dense)
1,440 PFLOPS | 1,080 PFLOPS
Total FP8/FP6 Tensor Core (sparse)
720 PFLOPS
Total fast memory
37 TB
Total memory bandwidth
576 TB/s
Total NVLink Switch bandwidth
130 TB/s
Individual Blackwell Ultra GPU Specifications
FP4 Tensor Core per GPU (sparse | dense)
20 PFLOPS | 15 PFLOPS
FP8/FP6 Tensor Core per GPU (sparse)
10 PFLOPS
INT8 Tensor Core per GPU (sparse)
330 TOPS
FP16/BF16 Tensor Core per GPU (sparse)
5 PFLOPS
TF32 Tensor Core per GPU (sparse)
2.5 PFLOPS
FP32 per GPU
80 TFLOPS
FP64/FP64 Tensor Core per GPU
1.3 TFLOPS
GPU Memory per GPU | bandwidth
279 GB HBM3e | 8 TB/s
Interconnect
Fifth-generation NVIDIA NVLink™: 1.8 TB/s
Source: NVIDIA Blackwell Ultra datasheet. Specifications in sparse unless noted. Projected performance subject to change.
Rack-scale architecture with a unified NVLink domain
The NVIDIA GB300 NVL72 integrates 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs into a single, fully liquid-cooled rack, connected by a 72-GPU unified NVIDIA NVLink domain delivering 130 TB/s of low-latency GPU communication. This architecture eliminates the inter-node communication overhead typical of multi-server clusters, enabling the entire rack to operate as a single high-bandwidth compute platform.
For workloads that span beyond a single rack, GB300 NVL72 systems on Nebius connect via NVIDIA Quantum-X800 InfiniBand at 800 Gb/s per GPU, scaling output across multiple racks while maintaining the high-speed fabric that large-scale reasoning and training workloads require.
NVIDIA Exemplar Cloud Achievement
NVIDIA Exemplar Cloud Achievement
Nebius is a Reference Platform NVIDIA Cloud Partner and has achieved NVIDIA Exemplar Cloud across multiple GPU generations, from NVIDIA H200 to NVIDIA GB300 NVL72. For the GB300 NVL72 specifically, Nebius achieved NVIDIA Exemplar Cloud status for training workloads, demonstrating real-world performance against NVIDIA’s benchmarking standards on this platform.
Frequently Asked Questions
The NVIDIA GB300 NVL72 is a fully liquid-cooled, rack-scale system built on NVIDIA Blackwell Ultra architecture. It integrates 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs into a single platform, connected by a 72-GPU unified NVLink domain, purpose-built for AI reasoning, frontier model training, and large-scale inference.
Get started with NVIDIA GB300 NVL72 on Nebius
Launch your first instance in minutes or talk to our team to find the right setup for your workload.