Crusoe Cloud rents NVIDIA H100 GPUs for $3.90/hour, positioning itself as the fiscally responsible alternative to the hyperscalers. While their marketing leans heavily on "stranded energy" and climate alignment, the real draw for engineering teams is the raw price-performance ratio. You aren't paying the "Bezos tax" for unmatched services you won't use; you're paying for high-performance compute, networking, and power.
For a realistic training run—say, fine-tuning a 70B parameter model on an 8x H100 SXM node for two weeks—the math is compelling. On Crusoe, that cluster costs roughly $10,500 ($31.20/hr total). On AWS, a comparable p5.48xlarge instance lists around $98/hr, pushing the bill over $32,000. Even compared to other specialty clouds, Crusoe holds its own; it undercuts CoreWeave’s list prices and matches Lambda’s competitive tiers, though Lambda often has better single-instance availability.
The platform feels like an industrial warehouse rather than a boutique. It lacks the polished "click-ops" console of AWS or the hobbyist-friendly community of RunPod. There is no one-click Jupyter notebook template waiting for you; you get SSH access and a clean slate. For infrastructure-as-code teams, the first-class Terraform provider is excellent, allowing you to treat GPU clusters like cattle. However, Python-native developers will find the lack of an official boto3-style SDK for VM management frustrating, forcing reliance on the CLI or Go client for automation.
Crusoe recently bridged the gap for inference with a serverless API that is OpenAI-compatible, making it easier to consume models like Llama 3 without managing boxes. But the core product remains bare-metal and VM infrastructure.
This is not the cloud for a student testing a chatbot for an afternoon. It’s for startups and research labs that need to reserve 64+ GPUs for a month without burning their seed round. If you need a managed Kubernetes experience, their CMK (Crusoe Managed Kubernetes) is solid but basic. If you just need a quick GPU for an hour, Lambda is friendlier. But for sustained, heavy training jobs where SOC2 compliance and carbon credits matter, Crusoe is the adult in the room.
Pricing
Crusoe operates on a strict pay-as-you-go model with no free tier. The entry point is the L40S GPU at $1.00/hour, but the H100 SXM at $3.90/hour is the volume seller. Unlike hyperscalers, there are no hidden fees for data ingress/egress within reasonable limits, and storage is a flat $0.08/GB/month. The "cost cliff" is availability—accessing H100s often requires a conversation with sales or a reservation commitment, unlike the instant-access A100s. There are no spot instances in the traditional sense of fluctuating auctions, but they offer lower rates for interruptible capacity if negotiated.
Technical Verdict
The infrastructure is robust, featuring high-bandwidth InfiniBand networking standard on H100 clusters, which is critical for distributed training. Documentation is clean but sparse. The API is REST-based with a focus on Terraform and Go for orchestration; the lack of a native, official Python SDK for infrastructure management is a notable friction point for ML engineering. Reliability is high (99.9%+), and cold start times for VMs are standard (1-2 minutes).
Quick Start
# Crusoe offers an OpenAI-compatible API for their Managed Inference
# For VM creation, you must currently use Terraform or the CLI
from openai import OpenAI
client = OpenAI(base_url="https://api.crusoe.ai/v1", api_key="CRUSOE_API_KEY")
response = client.chat.completions.create(
model="meta-llama/Llama-3.3-70B-Instruct",
messages=[{"role": "user", "content": "Explain stranded energy in 1 sentence."}]
)
print(response.choices[0].message.content)Watch Out
- No official Python SDK for VM infrastructure management (must use Terraform/CLI/Go).
- H100 availability usually requires a sales discussion or reservation, not instant click-to-spin.
- Limited region support compared to AWS (primary regions are US-East, US-Central, and EU-North).
- No built-in 'serverless container' runner for custom Docker images; you manage the VM or use their specific inference models.