Crusoe Spark AI factories achieve NVIDIA Exemplar level performance

Crusoe has achieved NVIDIA Exemplar Cloud performance on NVIDIA HGX B300 systems for training workloads deployed in Crusoe Spark™ modular AI factories and measured per-GPU performance on par with traditional data centers.

The result confirms something we have believed since we launched Crusoe Spark: that a prefabricated modular AI factory, delivered in months and powered by whatever energy source is available on site, can run the most demanding frontier training workloads at the same level as a traditional data center. NVIDIA's benchmarks now say the same thing.

What is the NVIDIA Exemplar Cloud?

The NVIDIA Exemplar Cloud initiative sets transparent and standardized benchmarking for cloud providers. Rather than testing isolated components, it evaluates real AI training workloads end-to-end using NVIDIA's Performance Benchmarking Suite, measuring throughput across large-scale LLM training scenarios with production-grade software stacks.

Passing this evaluation means Crusoe Cloud delivers workload-level performance aligned with NVIDIA's reference architecture, whether the GPUs sit in a gigawatt-scale campus or a modular AI factory. For AI teams evaluating infrastructure, it is an objective, NVIDIA-validated signal that training jobs are designed to run efficiently and predictably on Crusoe.

Crusoe's Spark validation ran on NVIDIA Blackwell Ultra GPUs across a broad set of frontier models: DeepSeek-V3 671B, Kimi K2 1T, Llama 3.1 405B, Qwen3 235B, GPT OSS 120B, NVIDIA Nemotron 3 Super 120B, NVIDIA Nemotron-H 56B, and Llama 3.1 70B, all trained with the Megatron-Bridge framework across BF16, FP8, and NVFP4 precisions. All model and precision configurations met NVIDIA's Exemplar recipe guidelines.

Crusoe Spark matches traditional data center cluster performance

The Crusoe Spark AI factories that were validated with Exemplar Cloud recipes house NVIDIA HGX B300, delivering 1.5x the dense NVFP4 compute, 2x the attention-layer performance, and 2.1 TB of total HBM3e memory per system, when compared to an equivalent cluster of NVIDIA HGX B200. Networking bandwidth doubles to 1.6 TB/s per system feeding the scale-out fabric.

Those gains come with real physical demands: higher rack power density, more heat to reject, and twice the network bandwidth to provision and keep clean at scale. Every Crusoe Spark unit integrates power distribution, cooling, remote monitoring, fire suppression, and high-density GPU racks into a prefabricated enclosure, and the cluster spans multiple units connected by a single high-performance fabric.

We deliberately did not build a special configuration for Crusoe Spark. We ran the same software stack and procedure we used for our HGX B300 validation in our traditional data center:

• Core Foundation: The cluster foundation remains the same, with the same storage, high-speed, low-latency networking (NVIDIA Quantum-X800 InfiniBand switches and adapters), and core compute architecture (bare metal and virtualization). VM image carries the same Ubuntu 24.04 OS, with matching NVIDIA CUDA and NVIDIA DOCA-OFED drivers.
• Orchestration: The Slurm cluster used to validate Spark is identical, in configurations, to the Iceland cluster, from Slurm to container runtime (Enroot/Pyxis) versions.
• Benchmark execution: The latest version of the NVIDIA Exemplar Cloud suite with Megatron-Bridge. Recipes were run exactly as published, and we tuned only the variables specific to Crusoe Cloud: NCCL settings, network topology, host configurations, and those related to orchestration.
• Evaluation: The same method as referenced in the official documentations, the mean per-GPU model throughput over the steady-state iterations of each run.

Every recipe met NVIDIA Exemplar performance target at the Spark AI factory, when compared to the equivalent GPU-scale cluster Per-GPU throughput measured in traditional data centers. This is what you'd expect from the same GPUs, fabric, and software, regardless of the building they sit in.

Achieving NVIDIA Exemplar-level performance for training across every model and precision in the suite means the fabric between Crusoe Spark AI factories, the cooling inside them, and the orchestration layer on top all held up under sustained frontier-scale load. There is no modular discount on performance. Customers get the full capability of NVIDIA Blackwell Ultra GPUs.

Powered by solar and second-life batteries in Nevada

The HGX B300 systems used for this validation run in Nevada are primarily powered by large-scale solar and second-life EV batteries. This is Crusoe's energy-first approach to AI infrastructure in practice. We are vertically integrated: we find innovative energy sources, build and manage AI factories at every scale, and offer a scalable AI cloud platform on top of them. Crusoe Spark exists because the fastest path to new AI capacity is often not a new hyperscale campus, but where power and ancillary infrastructure already exists.

Crusoe’s three Exemplar Cloud validation achievements on three different energy sources - hydro in Norway, geothermal in Iceland, and solar in Nevada, show this in action. Renewable energy and peak AI performance are not a tradeoff. Neither are speed of deployment and peak AI performance. Crusoe provides the foundational AI infrastructure so our customers can build the future faster.

Built in America, validated in America

Every Crusoe Spark modular AI factory in this cluster was designed and built at Crusoe manufacturing factories in the United States. Building in America is a deliberate choice, and we have been investing behind it. Crusoe is vertically integrated across a domestic manufacturing footprint spanning Colorado, Oklahoma, and Louisiana. Earlier this year we expanded it with the Crusoe Spark Factory in Brighton, Colorado, and earlier this month we opened a second Tulsa facility, bringing our US manufacturing capacity to roughly 1 million square feet. Across our sites, more than 1,800 employees and 15,000 skilled workers build the equipment behind America's AI buildout: welders, fabricators, electricians, and engineers. Owning that supply chain is also why we move fast, designing and building in-house what most of the industry has to source and wait on.

The idea behind Spark manufacturing is that a data center should be built like a high-precision machine, not a construction project. Switchgear, cooling, racks, and the enclosure itself are manufactured off-site in a controlled environment, then shipped and commissioned in as little as 90 days, compared with the year or more a traditional build requires. Because we design the module and operate the cloud that runs inside it, we can co-optimize the physical form factor for the GPUs it houses, down to the power and cooling configuration a specific NVIDIA Blackwell Ultra deployment needs.

NVIDIA Exemplar-level performance on a Spark deployment closes the loop on that approach. Modules built in Colorado, deployed on renewable energy in Nevada, running NVIDIA's most advanced platform at a performance level NVIDIA itself has verified. It is a repeatable unit of AI capacity that can be manufactured, shipped, and stood up in 90 days, and it is designed to perform like a hyperscale facility when it arrives.

What this means for AI teams

For customers building on Crusoe Cloud, this validation delivers practical confidence across three dimensions:

Performance that does not depend on the building. NVIDIA Exemplar-level performance across hyperscale and modular deployments, on consecutive NVIDIA Blackwell generations, in three regions, show that Crusoe's platform and operations can deliver the same result regardless of the facility underneath. Crusoe Cloud is designed to deliver consistent performance, so what you validate on Crusoe Cloud today is built to behave the same way wherever we deploy it next.

Validated on the workloads you are actually planning. Trillion-parameter MoE pretraining, 405B dense models in NVFP4, long-context attention on NVIDIA Blackwell Ultra GPUs. These are the shapes of frontier pretraining and RL post-training, and they run at scale on a modular deployment.

Capacity where and when you need it. Crusoe Spark units are manufactured in the Crusoe Spark Factory in Colorado, deploy in as little as 90 days, and can be deployed on-premise, at the edge, in a Crusoe Edge Zone, or wherever power is available. The same validated Crusoe Cloud stack runs on all of them. NVIDIA HGX B300 is available on Crusoe Cloud now, with trial runs before you commit and solutions engineers who specialize in AI infrastructure to help.

Get started now

To get started with NVIDIA HGX B300 on Crusoe Cloud, visit Crusoe Cloud or contact our team. To learn more about Spark and our data center deployments, explore Crusoe AI Data Centers.