Crusoe Cloud
by Crusoe Energy Systems
Vertically integrated AI cloud — own the power, own the datacentre, rent the GPUs by the minute
Crusoe Cloud is a GPU cloud for training, fine-tuning and serving AI models, run on datacentres Crusoe builds and powers itself. It offers NVIDIA H100, H200, GB200 and B200 plus AMD MI300X and MI355X with managed Kubernetes, managed Slurm, serverless inference and fine-tuning, published per-minute on-demand pricing and no egress fees — aimed at AI teams that want hyperscaler-grade capacity without hyperscaler contracts.
Crusoe Cloud is the public cloud arm of Crusoe Energy Systems, a Denver company founded in 2018 to burn flared natural gas for computation rather than waste it, and now one of the largest independent AI infrastructure builders in the United States. The distinguishing fact is vertical integration: Crusoe develops and powers the datacentres its cloud runs in, including a flagship 1.2 GW Abilene, Texas campus built for Oracle whose first two buildings are live with six more under construction, a separate 900 MW Abilene campus for Microsoft, a 1.8 GW Wyoming campus with Tallgrass, and a 12 MW Norway facility with Polar that will host Crusoe Cloud. Contracted AI infrastructure capacity approaches five gigawatts. The cloud itself offers NVIDIA GB200 NVL72, B200, H200, H100, A100 and L40S alongside AMD MI300X and MI355X, provisioned in minutes through a console, CLI, REST API, Terraform provider, SDKs and an MCP server. Around that sit Crusoe Managed Kubernetes and Managed Slurm for distributed training, Crusoe Command Center for real-time cluster monitoring and alerting, S3-compatible object storage co-located with compute, InfiniBand networking, and Crusoe Intelligence Foundry for serverless LoRA fine-tuning and managed inference without running clusters. Notably, Crusoe charges nothing for network ingress or egress. Customers include Meta, Microsoft, OpenAI, Oracle and Jane Street, the last on a reported five-year, $13 billion GPU contract. In September 2026 Crusoe raised over $3 billion in a Series F co-led by Atreides Management and Valor Equity Partners with Mubadala Capital participating, at roughly a $30 billion post-money valuation — triple its October 2025 Series E mark ten months earlier — and has met bankers about a potential near-term IPO. SemiAnalysis rates it Gold tier in its ClusterMAX GPU cloud rankings.
The head of AI infrastructure at a model-training or AI-product company that needs multi-node GPU clusters at committed scale and is comparing neoclouds — CoreWeave, Lambda, Nebius — rather than negotiating with a hyperscaler.
Hyperscaler-class GPU capacity at published per-minute rates with zero egress fees, from a provider that owns its own power and datacentres and therefore controls its own supply during a capacity crunch.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, Heads of AI Infrastructure, ML Platform Engineers, AI Research Labs, Enterprise Developers
- Deployment
- Cloud-only, API-based
- Founded
- 2018
- Headquarters
- Denver, United States
- Team Size
- 500+
Key Features
- ✓Published per-minute GPU pricing
H200 at $4.29, H100 at $3.90, MI300X at $3.45, A100 at $2.30 and L40S at $1.50 per GPU-hour, billed by the minute.
- ✓Zero ingress and egress fees
Crusoe charges nothing for network data transfer in or out, which removes the cost that makes moving training data between clouds prohibitive.
- ✓Crusoe Managed Kubernetes and Managed Slurm
Provision fault-tolerant multi-node GPU clusters with either orchestrator; managed Kubernetes bills at ten cents per cluster-hour on top of compute.
- ✓Crusoe Intelligence Foundry
Serverless LoRA-based supervised fine-tuning and managed model serving, so teams can customise open models without provisioning or babysitting GPU clusters.
- ✓Co-located S3-compatible storage
Object storage at six cents per GiB-month sits next to the compute, with persistent and shared block disks for training checkpoints and datasets.
- ✓Full programmatic surface
Console, CLI, versioned REST API, Terraform provider, SDKs and an MCP server, so clusters can be provisioned from infrastructure-as-code or an agent.
- ✓Vertically integrated power and datacentres
Crusoe builds and powers its own campuses — approaching five gigawatts contracted — which is what underwrites capacity commitments during GPU shortages.
Capabilities
Use Cases
- •Frontier and mid-scale model pretraining
An AI lab reserves multi-node H200 or GB200 clusters with InfiniBand interconnect and Managed Slurm to run distributed training jobs for months at committed pricing.
- •Escaping egress costs on a multi-cloud pipeline
A team that repeatedly ships terabytes of training data and checkpoints between environments cuts a recurring line item to zero, since Crusoe bills no transfer.
- •Serverless fine-tuning of open models
A product team runs LoRA supervised fine-tuning through Intelligence Foundry and serves the result, never provisioning or operating a GPU cluster directly.
- •Enterprise inference capacity at committed scale
A company with steady production inference reserves capacity for six months to three years at a discount rather than paying hyperscaler on-demand rates.
- •Sustainability-constrained AI workloads
An organisation with reporting obligations runs training in Iceland or Norway on hydroelectric and geothermal power rather than a carbon-intensive grid region.
Ideal For
Best For
- ✓Multi-node distributed training on H100, H200 or Blackwell-class clusters with InfiniBand networking
- ✓Teams that move large datasets or model artifacts in and out frequently and are being punished by hyperscaler egress fees
- ✓Reserved multi-month or multi-year GPU capacity where supply certainty matters more than self-serve flexibility
- ✓Serverless LoRA fine-tuning and managed inference for teams that do not want to operate GPU clusters at all
- ✓Organisations with sustainability mandates, since Crusoe's sites run on hydroelectric, geothermal, flare-mitigation and renewable-credit power
Not Ideal For
- ✗Small teams wanting cheap interruptible capacity — Crusoe publishes no spot tier, so interruption-tolerant workloads are materially cheaper elsewhere
- ✗Anyone needing Blackwell-generation hardware without a procurement cycle: GB200, B200 and MI355X are all contact-sales with no published hourly rate or self-serve path
- ✗Workloads requiring broad geographic coverage or specific compliance jurisdictions — Crusoe Cloud has five regions (Iceland, Norway, Houston, Virginia, Sparks NV) against a hyperscaler's dozens
- ✗Research labs wanting a polished managed-Slurm experience today: SemiAnalysis found the Slurm-on-Kubernetes offering unusable out of the box, with login pods missing vim, nano, git, python and sudo, and no partition support, RBAC or SSO integration
- ✗Teams that need reservation flexibility, since unused reserved capacity does not roll over and overages bill immediately at on-demand rates rather than averaging monthly
Integrations
Deployment
Market Analysis
Pros
- ✓Owns its power and datacentres end to end, which converts capacity from a procurement gamble into something Crusoe controls — the flagship 1.2 GW Abilene campus for Oracle and a second 900 MW campus for Microsoft are already building out
- ✓Transparent per-minute on-demand pricing on H100, H200, A100, L40S and MI300X, plus genuinely free ingress and egress, which is a large real-world saving on data-heavy training workloads
- ✓Independently rated Gold tier by SemiAnalysis ClusterMAX, with reported cluster uptime of 99.98% on one customer's H100 fleet and proactive fault detection and migration guidance
- ✓Blue-chip demand signal: Meta, Microsoft, OpenAI, Oracle and a reported five-year, $13 billion Jane Street contract, with a $3B Series F at a $30B valuation tripling its ten-month-old mark
- ✓Broad managed surface for a neocloud — Kubernetes, Slurm, Command Center monitoring, serverless fine-tuning and inference, Terraform and an MCP server
Cons
- ✗SemiAnalysis found the Slurm-on-Kubernetes offering not usable out of the box: login pods ship without vim, nano, git, python or sudo, there is no partition support, RBAC or SSO integration, and it is limited to roughly ten researchers per lab
- ✗The same independent review flags reliability defects — repeated NVML driver mismatch errors in containers, shared filesystems randomly unmounting, and widespread link flaps and filesystem unmounts at the Iceland facility before March 2025 — and warns of a downgrade to Silver tier
- ✗Organisational risk called out publicly: top engineers departing the cloud division, middle-management bloat and slow release cycles, with the AutoClusters feature described as having missed its window
- ✗Only five public cloud regions (Iceland, Norway, Houston, Virginia, Sparks NV) against a hyperscaler's dozens, so compliance-jurisdiction and latency requirements are frequently unmet
- ✗No spot tier, no published Blackwell pricing and no self-serve reservations — and reserved capacity neither rolls over nor averages overages monthly, so the commercial terms favour Crusoe on any usage miss
- ✗Kubernetes clusters ship without a default ReadWriteMany StorageClass and require manual OS drive, RAID and InfiniBand PKey configuration, which is real setup labour
Pricing
On-demand GPU
From $1.50/GPU-hr
- ✓L40S $1.50, A100 PCIe $2.00, A100 SXM $2.30, MI300X $3.45, H100 $3.90, H200 $4.29 per GPU-hour
- ✓Billed by the minute
- ✓No ingress or egress charges
Blackwell and MI355X
Contact for pricing
- ✓NVIDIA GB200 NVL72, NVIDIA B200 HGX, AMD MI355X
- ✓No published hourly rate
- ✓Sales-negotiated only
Reserved capacity
Contact for pricing
- ✓Six-month to three-year commitments at a discount
- ✓Direct sales negotiation, no self-serve
- ✓Unused capacity does not roll over
Managed Kubernetes
From $0.10/cluster-hour
- ✓Crusoe Managed Kubernetes control plane
- ✓Charged on top of GPU and CPU compute
Crusoe Intelligence Foundry
$5 free credits
- ✓Serverless LoRA fine-tuning
- ✓Managed model serving
- ✓Complimentary starting credits
Crusoe publishes real per-GPU-hour on-demand rates for its older SKUs and bills them by the minute, but every Blackwell-generation part — GB200 NVL72, B200 HGX and AMD MI355X — is contact-sales with no listed rate, so the hardware most buyers actually want cannot be priced without a call. Reserved contracts run six months to three years at undisclosed discounts, are negotiated directly with sales, and carry two terms worth reading closely: unused reserved capacity does not roll over, and overages bill immediately at on-demand rates rather than averaging across the month. There is no spot tier. CPU runs $0.04 per vCPU-hour general-purpose or $0.09 storage-optimised, object storage $0.06 per GiB-month, persistent disk $0.08 and shared disk $0.07, managed Kubernetes $0.10 per cluster-hour, and network ingress and egress are free. Intelligence Foundry gives $5 of complimentary credits; there is no general free tier.
Security & Compliance
Sources
This page was written from 7 sources, 5 on domains other than crusoe.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Nexthop AI
Ethernet switches and network software co-designed for AI clusters, built for hyperscalers and NeoClouds
InsightFinder
Predictive reliability for AI agents and IT estates — unsupervised anomaly detection that flags incidents before they land
Tensormesh
KV-cache-native LLM inference that bills cached input tokens at $0
OCI Enterprise AI
OpenAI-compatible agents, tools and memory on Oracle's own cloud, with the data staying put