Crusoe is developing sustainable AI infrastructure using an energy-first approach, utilizing excess natural gas as well as renewable energy to produce power for high-performance computing and AI applications. The company develops and operates hyperscale AI data centers as well as provides the Crusoe Cloud, a cloud-based AI environment that has been developed to support AI development and deployment at scale. As a vertically-integrated organization, Crusoe also provides customers with the ability to develop AI solutions at scale while ensuring that they are utilizing reliable performance and low-carbon energy.
Crusoe was founded by Chase Lochmiller and Cully Cavness and is based on the concept of implementing an energy-first approach to utilize excess natural gas for high-performance computing.
Funding from Series A investment established the financial base to execute the mission of providing sustainable computing.
$128M in Series B funding was used to expand Digital Flare Mitigation technology throughout the United States.
$350M in Series C funding was utilized to accelerate expansion of Crusoe Cloud operations began mid-2022 with initial A100 deployment in Montana.
Launched the sustainable geothermal and hydro-powered Crusoe Cloud in Iceland and published its first Environmental Social Governance report.
$1B+ in Series E funding; launched Crusoe Managed Inference with MemoryAlloy; acquired Atero; announced plans to build a 1.8 GW data center in Wyoming.
Hyperscale AI data centers with renewable-powered facilities in US (MT, VA, TX, WY) and Iceland; containerized and modular deployments
Purpose-built for AI training/inference with proprietary MemoryAlloy GPU optimization, managed inference services, and AutoClusters for large-scale model deployment up to 20x faster
Based on official website product descriptions and company announcements
Crusoe supports NVIDIA H100, H200, A100 (40GB / 80GB), L40S, A40, GB200, B200 and AMD MI300X GPUs. Customers can select a variety of configurations ranging from one or two GPUs, up to eight cluster nodes with a maximum total of 946K GPUs supported across all of the Crusoe data centers.
On demand pricing begins at $0.90 per GPU-hour for A40, $1.45 for A100 PCIe, and goes up to $3.90 per GPU-hour for H100. Discounted prices are available for spot instances, such as $1.00 for A100 and $0.50 for L40S. Additionally, reserved pricing is available for customers that commit to a specific amount of usage over time.
Crusoe is focused on supporting AI workloads by providing a highly specialized GPU-based architecture with low carbon footprint, low operating expenses, and up to 81 percent cost savings over traditional hyperscalers. In addition to the cost savings, Crusoe also provides 99.98 percent uptime versus traditional general purpose clouds, and provides customers with optimized RDMA networking and AutoClusters to help optimize their workflows.
Yes, Crusoe provides a full suite of enterprise grade infrastructure features, including 99.98 percent uptime, redundant systems, 24/7 support, advanced cooling systems in each of our data center locations, and AI-optimized networking systems. Additionally, Crusoe is compliant with SOC 2 standards, similar to many other cloud platforms that provide a secure environment for customers to deploy their applications.
Yes, we provide managed Kubernetes, Slurm clusters and fault tolerant AutoClusters. We also integrate with Saturn Cloud, which allows customers to utilize JupyterLab and RStudio, and enables elastic scaling of GPU-based workloads.
Enterprise 24/7 support with 100% customer satisfaction is provided through Command Center, which offers a combination of orchestration, observability and collaboration features. For custom service level agreements (SLA), or for having dedicated resources, please contact sales.
No information about the existence of a free public tier was found; new customers are to contact sales to receive a trial license and/or proof-of-concept. A pay-as-you-go model with spot/on-demand pricing will be used.
Some virtual machine (VM) configurations may have geographic restrictions (e.g., some regions); for access to these configurations, please contact sales.
Ephemeral storage is available on GPU instances. The primary function of Crusoe is to offer specialized infrastructure for AI workloads using high-performance GPUs from NVIDIA and AMD, as well as to provide purpose-built data centers that can achieve cost savings of up to 81% compared to hyperscalers. Crusoe also has strong reliability (99.98% uptime), with tools such as Command Center and AutoClusters that allow the company to deliver an enterprise-level platform to support large-scale AI training and inference. In addition, the company's focus on renewable energy and its ability to rapidly expand its capacity allows it to compete effectively in the crowded AI cloud market.
Crusoe is best suited for large scale GPU-based AI factories that prioritize performance, cost savings, and sustainability above the versatility of a general-purpose cloud provider.
Summary: Crusoe Cloud provides NVIDIA H100/A100/H200 and AMD MI300X GPUs with an interesting on-demand pricing structure ranging from $0.90-$3.90/GPU-hr (with spot discounts they don’t disclose that go up to about 67%). Their AI infrastructure is built around purpose-built AI data centers spanning ~9.8M sq ft, delivering 99.98% uptime on automated hardware-level SSD and NIC failover, with renewable power and air-cooling infrastructure. Some of their tooling looks interesting including Command Center, AutoClusters, and managed Kubernetes. They sell mostly to enterprise companies (their public pricing API and developer docs references are scant, and require emailing their sales team).
Good - detailed specs from official docs and pricing pages, third-party benchmarks. Limited API/integration documentation and no public FAQ/customer stories.
Their purpose-built AI cloud is spread out to about 9.8M square feet of space (multiply that by about 4 for 466K GPUs) and deliver 3.4GW of electricity to 946K GPUs. They do advanced features like air-cooling, RDMA networking with racks with 1GPU-6GPU density, and focusing on renewable energy sources.
“We recently launched a new unified operations platform, available to all AI workloads. We focus less on orchestration and more on deep observability, collaborative support experiences, and optimizing GPU uptime.”.
Partners with Saturn Cloud to employ a GPU-powered JupyterLab, RStudio, Dask clusters, etc., along with managed Kubernetes, Slurm, and AutoClusters to automate MLOps.
They power their data centers on renewable energy sources, stating they are “designed to significantly reduce the carbon footprint of AI workloads with advanced energy efficiency features”, and indicate they’re for “high-density GPU deployments with precision power distribution who traditionally bought infrastructure with hyperscalers”.
“Rapidly scaling and deploying infrastructure to support needs for GB200 and B200 next-gen GPUs” and engineer solely for AI.
| Workload Type | Recommended Hardware | Performance Advantage | Configuration Example | Use Case Notes |
|---|---|---|---|---|
| Generative AI Inference (Text) | NVIDIA L40S, A100 | 30% faster inference vs A40/A100 | l40s-48gb.1x to l40s-48gb.10x | LLM inference, chatbots, text summarization |
| Image Generation (SDXL, Stable Diffusion) | NVIDIA L40S | 3.3 images/sec fp16, 2.8 images/sec int8 | l40s-48gb.1x or higher | SDXL inference, native fp8/int8 support |
| Fine-tuning & Training | NVIDIA L40S clusters (10x GPU) | Completed 32k context Mistral 7b finetune in 3 days | l40s-48gb.10x instance | Cost-effective training: ~$1000 for multi-GPU finetune |
| Multi-Model Serverless Deployment | NVIDIA L40S (10x GPU, high memory) | Fast model switching, multiple models in system memory | l40s-48gb.10x with 1470GB total memory | Serverless providers, per-customer fine-tunes |
| Large-Scale Model Training | NVIDIA H100, H200, B200 | High-bandwidth interconnect, extreme scaling | Multi-node H100/H200 clusters | Trillion-parameter models, distributed training |
| Instance Type | vCPU Count | Memory | Network Bandwidth | Typical Use Case |
|---|---|---|---|---|
| c1a.2x | 2 | 8 GB | 1 Gbps | Development, testing |
| c1a.4x | 4 | 16 GB | 2 Gbps | Small workloads |
| c1a.8x | 8 | 32 GB | 5 Gbps | Medium compute |
| c1a.16x | 16 | 64 GB | 10 Gbps | General purpose |
| c1a.32x | 32 | 128 GB | 20 Gbps | High-memory workloads |
| c1a.64x | 64 | 256 GB | 35 Gbps | Large-scale processing |
| c1a.128x | 128 | 512 GB | 70 Gbps | Extreme-scale workloads |
| c1a.176x | 176 | 704 GB | 100 Gbps | Maximum throughput |
| Instance Type | vCPU Count | Memory | NVMe Storage | Network Bandwidth | Use Case |
|---|---|---|---|---|---|
| s1a.20x | 20 | 176 GB | 1x 12.8 TB | 25 Gbps | Data processing, storage-intensive AI |
| s1a.40x | 40 | 352 GB | 2x 12.8 TB | 50 Gbps | Large dataset handling |
| s1a.60x | 60 | 528 GB | 3x 12.8 TB | 75 Gbps | Multi-tier data workloads |
| s1a.80x | 80 | 704 GB | 4x 12.8 TB | 100 Gbps | Enterprise-scale storage |
| s1a.120x | 120 | 1056 GB | 6x 12.8 TB | 150 Gbps | Distributed training data |
| s1a.160x | 160 | 1408 GB | 8x 12.8 TB | 200 Gbps | Maximum storage density |
The proprietary technology provided by Crusoe enables ultra-low latency and scalable throughput in large context AI workloads.
Crusoe’s simple container orchestration, based on automated scaling and fault tolerance, minimizes operational overhead.
Job Scheduling for Distributed Computing.
Automatic cluster management that eliminates operational overhead to provide fault-tolerance.
High Bandwidth and Low Latency GPU-to-GPU Communication.
Integration with High Performance Storage Solutions for AI Workloads.
Crusoe supports FP8, FP16, FP32, BF16 and INT8 Quantization on L40S and later generation GPU Architectures.
Crusoe supports hardware accelerated sparse tensor operations on supported GPU types.
