Nscale achieves NVIDIA Exemplar Cloud status on NVIDIA GB300 NVL72

Nscale is a fast-scaling AI cloud provider, building gigawatt-class AI factories, and deploying NVIDIA GPUs at incredible scale as an NVIDIA Cloud Partner. Today we are adding another marker to that trajectory: Nscale has achieved NVIDIA Exemplar Cloud status on NVIDIA GB300 NVL72 for large-scale training workloads.

‍

In June this year, we ran the complete set of benchmarking recipes end-to-end on the rack-scale NVIDIA GB300 NVL72 system, powered by Blackwell Ultra GPUs, with every workload running cleanly and reproducibly against NVIDIA’s reference architecture. Every run was carried out at 512 GPUs per workload, equating to seven racks of the GB300 NVL72 connected with NVIDIA Quantum-X800 InfiniBand networking, matching NVIDIA’s recommended scale for this validation. It is a verifiable confirmation that when you train on Nscale, the infrastructure performs the way NVIDIA’s own engineers expect it to.

What NVIDIA Exemplar Cloud status means 

The NVIDIA Exemplar Clouds validation initiative drives NVIDIA Cloud Partners toward the best performance and total cost of ownership relative to the NVIDIA reference architecture, through operational best practices for real-world workload performance. Rather than synthetic peak-performance figures, it measures whether a platform actually sustains the performance NVIDIA expects across a fixed catalogue of modern training workloads.

‍

The validation shows that every part of the system, from power and compute to communications, is working together at the level NVIDIA expects. For an NVIDIA Cloud Partner, achieving Exemplar Cloud is recognition from NVIDIA that a platform meets NVIDIA’s standards for AI workload performance, security and resiliency. For the teams who build on it, it provides a consistent standard of performance across clouds, and the confidence in total cost of ownership needed to plan and execute large AI projects efficiently.

‍

‍

Proving performance on NVIDIA Blackwell Ultra

The benchmarking suite is deliberately broad. On GB300 NVL72, Nscale ran 14 workload configurations built from today’s leading open models, including GPT-OSS, Kimi K2, Qwen3, Nemotron-3 Super, Nemotron-H, Llama 3.1, and DeepSeek-V3, spanning both Mixture-of-Experts and dense architectures. The benchmarks exercised four numerical formats: BF16, FP8 (current-scaling), FP8 (microscaling/MX), and NVFP4, the new 4-bit floating point format that represents the higher throughput for training on the Blackwell architecture. 

‍

Every configuration completed its full run, and the expected performance progression held on real models: the lower-precision formats delivered the throughput gains they promise, all the way down to NVFP4. Just as importantly, we repeated the entire suite across multiple independent 512-GPU allocations and received consistent results, proof of reliable and repeatable performance. Our training logs were sent to NVIDIA and validated. 

Fleet-level results, not a one-off 

Performance and architecture vary widely from one cloud to the next. Differences in firmware, networking, topology, and system configuration can materially affect workload performance, making it difficult to compare AI platforms consistently. For developers and AI leaders, that variability can translate into unpredictable performance and costs, delaying timelines, straining budgets, and slowing internal AI innovation. Standing up each new GPU generation in production brings additional challenges. 

‍

These results were produced on a production GB300 NVL72 cluster at our Verne campus in Keflavík, Iceland, powered entirely by renewable geothermal energy on a 64-rack GB300 NVL72 deployment. While NVIDIA only requires three sets of runs of their test suite(choosing the fastest one for Exemplar Cloud validation), we ran the test suite eight times, across all 64 racks in the data center. All 112 runs passed NVIDIA evaluation standards. This was not a bespoke benchmarking rig, but the same full-stack platform used by our customers around the world.

‍

That matters because NVIDIA Exemplar Cloud reflects the real challenge AI teams face: not simply accessing the latest GPUs, but operating them as a coherent cluster, where placement, topology, collective transport, fleet health, performance, resiliency, and security all work together. Those details determine whether a training run sees the performance the architecture is capable of delivering. 

‍

For the Mixture-of-Experts workloads in the NVIDIA Performance Benchmarking recipes, Nscale’s engineers ensured expert-parallel groups stayed inside the correct NVLink domains, selected the right NCCL and InfiniBand communication path, and repeated the full suite across independent GPU allocations to prove the result was fleet-level, not one-off. 

‍

Bringing a brand-new generation of hardware up to NVIDIA’s standard of performance, and maintaining it reproducibly across the full fleet, is exactly the capability that Nscale has built. This is the value Nscale provides: beyond simple access to Blackwell Ultra capacity, customers get the advantage of a full-stack platform that can bring new NVIDIA architectures up, validate them, and keep them performing predictably at scale. 

‍

‍

Get started with Nscale

For customers, building with Nscale means access to a full-stack AI platform: AI factory capacity, high-density GB300 NVL72 infrastructure, high-bandwidth networking, topology-aware cluster operations, workload validation, and engineering support from bring-up through production training and inference. Nscale gives AI-native companies and enterprises a predictable path from GPU capacity to workload outcomes.

‍

Ready to build on validated Blackwell Ultra infrastructure? Speak to Nscale about dedicated capacity, workload validation, and full-stack deployment support for frontier training, post-training, and inference.

‍