Question 1

How much does an H100 SXM cost per hour to rent?

Accepted Answer

Runcrate H100 SXM rental is $1.35-1.65/hr on-demand (avg $1.50/hr). That's significantly below AWS ($4.10), GCP ($3.67), Azure ($3.96), and Oracle ($2.40). Reserved pricing drops the rate another 30%.

Question 2

What's the difference between H100 SXM and H100 PCIe?

Accepted Answer

H100 SXM has 67% more memory bandwidth (3350 vs 2000 GB/s), full NVLink support for 8-GPU clusters, and 18% higher FP32 throughput. PCIe is cheaper and works in standard servers but caps at 4-GPU nodes without NVLink. SXM is the right choice for any multi-GPU workload.

Question 3

Can I rent H100 SXM on demand without a contract?

Accepted Answer

Yes — Runcrate H100 SXM is fully on-demand with per-second billing across 8 regions. No contracts, no minimums, no commitments.

Question 4

H100 vs A100 — when should I pick H100?

Accepted Answer

H100 has ~3x the FP16 throughput of A100 (120 vs 78 TFLOPS) and supports FP8. For modern transformer training, fine-tuning, or any inference workload, H100 is dramatically more cost-efficient. A100 only makes sense for legacy code that can't use FP8 or bf16.

Question 5

How fast can I deploy an H100 SXM instance?

Accepted Answer

Most users deploy in under 60 seconds. Pick a region, choose a base image (vLLM, Comfy, Jupyter, custom), attach a volume, and SSH in. Per-second billing means there's zero risk to spinning up briefly.

NVIDIA H100 SXM.

What you'll actually use this for.

Llama 70B Fine-tune

Mistral 7B Inference

Stable Diffusion XL

Multi-GPU Training

Full specifications.

Other Hopper GPUs.

Frequently asked.

Rent the NVIDIA H100 SXM on Runcrate.