Pricing
Pay only for the GPU time you use.
Top up your wallet, run jobs from your notebook or the dashboard, and see the maximum cost before anything starts.
GPU prices
| GPU | Memory | From (per GPU-hour) | Good for |
|---|---|---|---|
| V100 | 32 GB | $0.10 | Learning, small CNNs, fp16 training (no bf16) |
| L4 | 24 GB | $0.44 | QLoRA fine-tunes up to ~7B, bf16 training |
| RTX 4090 | 24 GB | $0.44 | Fast 7B QLoRA, vision models, quick experiments |
| RTX 5090 | 32 GB | $0.63 | Up to ~13B QLoRA, faster training runs |
| A100 | 80 GB | $1.37 | 13B–30B fine-tunes, full fine-tunes, multi-GPU pre-training |
V100
from $0.10 / GPU-hour
- Memory:
- 32 GB
- Good for:
- Learning, small CNNs, fp16 training (no bf16)
L4
from $0.44 / GPU-hour
- Memory:
- 24 GB
- Good for:
- QLoRA fine-tunes up to ~7B, bf16 training
RTX 4090
from $0.44 / GPU-hour
- Memory:
- 24 GB
- Good for:
- Fast 7B QLoRA, vision models, quick experiments
RTX 5090
from $0.63 / GPU-hour
- Memory:
- 32 GB
- Good for:
- Up to ~13B QLoRA, faster training runs
A100
from $1.37 / GPU-hour
- Memory:
- 80 GB
- Good for:
- 13B–30B fine-tunes, full fine-tunes, multi-GPU pre-training
Lowest current price, in USD. Exact rates for each GPU type are shown in your dashboard, and your price is locked when you launch. Updated 21 November 2026.
How billing works
Wallet top-ups.
Add credit to your account; jobs draw from your balance. Minimum top-up: $5.00.
Cost before you commit.
Every job shows its maximum cost, based on GPUs × time limit, before it starts. You're billed for actual time used.
Failed on our side? Not charged.
If a job fails because of a DataSpires issue, you aren't billed. Jobs that fail because of your code or data are billed for the GPU time used.
Deployed endpoints.
[TODO: how deployed models are billed, e.g. per hour while running or per request]
Storage and uploads.
[TODO: whether storage and data uploads are charged]
Example
Fine-tuning Qwen 2.5 0.5B in low (QLoRA) mode on 1 × L4 with a 1-hour time limit costs at most $0.87.
Based on the highest current L4 price.
What fits on a single GPU
| Model size | Mode | Fits on 1 GPU? |
|---|---|---|
| 0.5B–3B | low / medium | Yes |
| 3B–7B | low (QLoRA 4-bit) | Yes |
| 7B | medium | Yes |
| 13B | low | Tight |
| 30B+ | low | Use multiple GPUs |
Not sure? Call client.get_model_requirements(model, mode) before you submit.
Coming soon
Team and cohort pools
One shared GPU budget for a lab, course or accelerator, with a fair share for every member.
Frequently asked questions
Yes. The smallest top-up is $5.00.
10 hours per job. Long runs can checkpoint and resume from where they stopped.
Today, on vetted GPU providers in India and the US, chosen to keep costs low. We'll add African capacity as it becomes cost-competitive.
Ready to run your first job?
pip install dataspires