How much does it cost to train an LLM?
Training cost is the number of GPU hours you need multiplied by the hourly rate per GPU. Fine-tuning a small model can be a handful of GPU hours, while training a frontier model from scratch runs into millions of GPU hours across thousands of cards. The cheapest verified H100 in our index is $1.99/hr, which is the number you would plug in for a Hopper-class training run.
The cheapest on-demand H100 offers, per GPU-hour. Spot, community and marketplace-host offers are listed on the H100 page.
| GPU | VRAM | $/GPU-hr | Where | |
|---|---|---|---|---|
| H100 | 80 GB | $1.99 | Voltage Park on-demand | Rent → |
| H100 SXM | 80 GB | $2.25 | TensorDock on-demand | Rent → |
| H100 | 80 GB | $2.43 | Prime Intellect on-demand | Rent → |
| H100 | 80 GB | $2.50 | Hyperstack on-demand | Rent → |
| H100 | 80 GB | $2.60 | GMI Cloud on-demand | Rent → |
| H100 | 80 GB | $2.64 | Spheron on-demand | Rent → |
| H100 | 80 GB | $2.73 | Massed Compute on-demand | Rent → |
| H100 PCIe | 80 GB | $2.89 | RunPod secure cloud | Rent → |
For a concrete estimate, start with the compute the run needs and the throughput of your GPU, then divide to get GPU hours. Multiply by the hourly rate, and multiply again by the number of GPUs if you train in parallel. A short fine-tune of a 7B model might take tens of GPU hours; a full pretraining run of a mid-sized model takes tens of thousands.
The hourly rate is where a price index pays off, because the same H100 hour can differ several times over between a hyperscaler and a community tier. Spot instances cut the rate further if your training checkpoints often enough to survive interruptions.
Do not forget the costs around the GPUs: storage for datasets and checkpoints, egress if you move data out, and the wall-clock overhead of failed runs and restarts. Those are easy to underestimate and can rival the raw compute bill on a long project.
Related GPUs
H100 rental prices · H100 vs A100 · H100 vs H200
Related questions
- How much does an H100 cost per hour?
- How does multi-GPU node pricing work?
- Is renting a GPU cheaper than buying one?
Numbers on this page come from today's verified snapshot. Full table on the homepage; method in the methodology.