Crusoe earns NVIDIA Exemplar Cloud validation on Blackwell Ultra

Crusoe has achieved NVIDIA Exemplar Cloud validation for training on NVIDIA HGX B300 systems. This follows our NVIDIA Exemplar Cloud validation on HGX B200 earlier this year, making Crusoe an Exemplar Cloud on both Blackwell generations. The result confirms that NVIDIA GPUs on Crusoe Cloud deliver the consistent, end-to-end performance that large-scale AI training demands, this time on NVIDIA's most advanced Blackwell Ultra platform. It reflects work across every layer of the stack: facility design, GPU orchestration, and networking, all engineered so customers extract the full capability of the NVIDIA Accelerated Computing Platform.

What is the NVIDIA Exemplar Cloud?

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

Crusoe's NVIDIA HGX B300 validation ran on 512 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, Nemotron 3 120B, 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 performance criteria.

This breadth matters, showing it is not a single benchmark on a favorable workload. It validates performance across both dense and mixture-of-experts architectures, from 56 billion to one trillion parameters, in the precision formats AI teams use.

Passing this evaluation means Crusoe Cloud delivers workload-level performance aligned with NVIDIA's reference architecture. For AI teams evaluating infrastructure, it is an objective, NVIDIA-validated signal that training jobs will run efficiently and predictably on Crusoe, with no guesswork required.

Unlocking the full potential of NVIDIA Blackwell Ultra

NVIDIA HGX B300 is built on the Blackwell Ultra architecture and extends what Blackwell made possible. Compared with NVIDIA HGX B200, it delivers 1.5x the dense NVFP4 compute, 2x the attention-layer performance, and 288 GB of HBM3e per GPU for 2.1 TB of total memory per system, up from 1.4 TB. Networking bandwidth doubles to 1.6 TB/s per system, with 800 Gb/s per GPU feeding the scale-out fabric.

For training, these gains land exactly where frontier models put pressure: NVFP4 throughput for mixture-of-experts pretraining, attention acceleration for long-context sequences, and memory headroom that lets larger models and larger batches fit on-chip. NVFP4 is also more computationally efficient: 4-bit values with shared scaling drastically reduce memory traffic and compute cost while maintaining accuracy close to higher-precision formats. Our validation held across all three precisions, on every model in the suite.

Hardware alone does not guarantee performance. Extracting the full capability of NVIDIA Blackwell Ultra GPUs requires infrastructure engineered for their power density, direct liquid cooling (DLC) requirements, and doubled network bandwidth. Crusoe's AI-optimized data centers and cloud platform are purpose-built to do exactly that.

Powered by geothermal energy in Iceland

The NVIDIA HGX B300 systems used for this validation ran in Iceland, powered entirely by geothermal energy, making the cluster 100% renewable. The naturally cool climate also reduces the facility-level cooling load needed to support NVIDIA HGX B300's direct liquid cooling infrastructure, improving overall system efficiency and uptime.

This is Crusoe's energy-first approach to AI infrastructure in practice. We are vertically integrated: we find innovative energy sources, build and manage hyperscale AI factories, and offer a scalable AI cloud platform on top of them. Siting NVIDIA Blackwell Ultra systems where abundant geothermal energy is available, and pairing that with the computational efficiency of the NVIDIA Blackwell Ultra architecture across every precision we validated, means more training throughput for every unit of energy we bring online.

Achieving NVIDIA Exemplar Cloud validation on geothermal-powered infrastructure, following our hydro-powered HGX B200 validation in Norway, demonstrates that renewable energy and peak AI performance are not a tradeoff. They go hand in hand. Crusoe provides the foundational AI infrastructure so our customers can build the future faster.

What this means for AI teams

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

Repeatable proven performance. Exemplar Cloud validations on consecutive Blackwell generations across different regions, show that Crusoe's platform and operations hold up as the hardware changes underneath. What you validate on today will behave the same way when you scale.

Validated on the workloads you are actually planning. Trillion-parameter MoE pretraining, 405B dense models in NVFP4, long-context attention on Blackwell Ultra: these are the shapes of frontier pre-training and RL post-training, and they ran at 512-GPU scale on the full NVIDIA Accelerated Computing Platform, including CUDA, NeMo, Megatron-Bridge, and the libraries millions of developers already use.

Built from energy source to GPU, available today. Crusoe engineers the full path from power to compute, so the platform is ready for Blackwell Ultra's power density, direct liquid cooling, and 1.6 TB/s of networking per system. 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.ai/cloud or contact our team.