H100 on Crusoe: Pricing, Specs & How to Rent

Deploybase · May 18, 2025 · GPU Pricing

Contents

Overview

As of March 2026, Crusoe rents H100s at competitive rates with a focus on sustainability (lower power bills). Pricing not listed publicly-contact sales for quotes. Good for large-scale training/inference if developers need custom arrangements.

H100 Specs

  • 80GB HBM3 memory
  • 3.35 TB/s bandwidth
  • NVLink 4.0
  • FP32: 67 TFLOPS
  • Tensor: 1,457 TFLOPS
  • 700W max power

Handles 70B+ models in distributed training. Standard stuff.

Pricing

Crusoe doesn't list H100 pricing publicly. Contact sales for quotes. Expect:

  • Spot/interruptible: 40-60% savings
  • 6-12 month commitments: Discounts available

Reference: RunPod H100 is $2.69/hr.

How to Rent

  1. Sign up at crusoe.energy
  2. Search for H100 availability
  3. Contact sales if stock is low (common)
  4. Configure CPU/RAM/storage
  5. Deploy and monitor via API

Setup: 1-2 weeks (custom quotes take time).

Performance Benchmarks

H100 performance varies by use case:

LLM Training (7B Model)

  • Throughput: 2,100 tokens/second (single GPU)
  • Training time (Llama 7B): 5-7 days on single H100
  • Scaling efficiency: 92% with 8x H100 cluster

Inference Latency (13B Model)

  • Time to first token: 8-12ms (batch size 1)
  • Throughput: 850 tokens/second (batch size 32)
  • Peak utilization: 95%+ under load

CUDA Operations

  • Compilation time: 30-45 seconds per kernel
  • Memory allocation overhead: <5ms
  • Context switching: minimal impact on performance

Comparison with Other Providers

H100 availability spans multiple cloud providers with varying pricing models:

ProviderH100 Pricing ($/hr)MemoryAvailability
RunPod$2.6980GBHigh
Lambda$3.78 (SXM)80GBMedium
CoreWeave$49.24 (8x bundle)640GBHigh
CrusoeCustom quote80GBDirect contact

Lambda Labs H100 pricing offers competitive rates for shorter workloads. CoreWeave bundles deliver better unit economics for large-scale training. For budget-conscious teams, RunPod remains the most affordable starting point, though Crusoe's sustainability focus appeals to ESG-aligned teams.

FAQ

Can I use H100 on Crusoe for fine-tuning small models? Yes. Single H100 instances are suitable for 7B-13B model fine-tuning, though cost per token may exceed Lambda Labs for shorter 1-2 hour sessions.

Does Crusoe offer spot/preemptible H100 instances? Crusoe's pricing model differs from cloud providers. Availability depends on their energy production and is not traditional spot pricing. Contact sales for interruptible capacity options.

What's the typical lead time to provision H100 on Crusoe? Custom H100 arrangements may take 5-15 business days depending on capacity and commitment terms. Standard on-demand availability is typically same-day for smaller instances.

How does H100 perform versus H200 for LLM inference? H200 adds 141GB HBM3e memory versus H100's 80GB, enabling larger batch sizes and longer context windows. H100 remains cost-effective for models under 70B parameters.

Are there CUDA compatibility issues between Crusoe and other H100 providers? No. NVIDIA's CUDA runtime and libraries maintain binary compatibility across all H100 deployments. Code written for RunPod H100 will run identically on Crusoe.

Sources