Write or upload your training code in the browser, pick a GPU and start. GridPup runs it on a cloud GPU, saves your results and shuts the machine down. Train on an RTX 4090 from $0.340 an hour or an H100 from $1.73 an hour, paid per minute.
Training GPUs and their price now
Cheapest on-demand price per GPU-hour across clouds right now, and what a 10-hour run on one GPU costs at that price.
| GPU | Memory | From /GPU/h |
|---|---|---|
| NVIDIA B200 | 180 GB | $6.10 |
| NVIDIA H200 | 141 GB | $3.59 |
| NVIDIA H100 | 94 GB | $1.73 |
| NVIDIA A100 | 80 GB | $0.413 |
| NVIDIA RTX PRO 6000 | 96 GB | $1.07 |
| NVIDIA L40S | 48 GB | $0.380 |
| NVIDIA RTX 5090 | 32 GB | $0.453 |
| NVIDIA RTX 4090 | 48 GB | $0.340 |
| NVIDIA L4 | 24 GB | $0.434 |
Prices from GridPup, collected from GPU clouds every few minutes. Storage is billed separately on most clouds.
How a training run works
- Add your code. Write it in the browser editor or upload a project. Pick an environment: PyTorch, TensorFlow, CUDA or plain Ubuntu, and add packages from your requirements file.
- Pick a GPU and a limit. You see the price per hour and the most the run can cost before you start.
- GridPup runs it. It finds the cheapest verified machine with your GPU, starts your code and streams the log. A machine that fails to start is replaced at no charge.
- Get your results. Everything your code writes to
/outputs(models, checkpoints, metrics) is saved when the run ends, and the machine is shut down.
After training
Benchmark it on real GPUs
Check your model works, gives the right answers and how fast it is on each GPU before you deploy.
Deploy it as an API
Serve your model behind an https API with keys, on the cheapest GPU that fits, billed by the hour.
Start from an open model
See how much GPU memory Llama, Qwen, Gemma or Mistral need before you fine-tune them.
Frequently asked questions
How much does it cost to train a model on a cloud GPU?
You pay for the GPU while your code runs, per minute. Training GPUs range from well under a dollar an hour (RTX 4090) to a few dollars an hour (H100, B200); the live table on this page shows the price now. A 10-hour fine-tuning run on one RTX 4090 usually costs a few dollars.
Which GPU should I use to fine-tune an LLM?
For LoRA on 7B-8B models an RTX 4090 or RTX 5090 is enough and cheapest. For 13B-34B models or full fine-tuning of small models use an A100 or L40S. For 70B models and full fine-tuning use H100, H200 or B200, often several at once.
What code can I run?
Any training code that runs on Linux with an NVIDIA GPU: PyTorch, TensorFlow, JAX, Hugging Face Transformers, Lightning or scikit-learn. You choose a ready environment and add your own packages, for example with a requirements.txt file.
Can a training run cost more than I expect?
No. GridPup shows the price per hour and the most the run can cost before it starts, and the run stops by itself at the cost or time limit you set. You can also stop it at any time; what your code saved to /outputs is kept.
Do I need a cloud account or my own GPU?
No. GridPup finds a verified machine with the GPU you pick, sets it up, runs your code and shuts it down. You can write code without an account; you sign in and add credit when you want to run it.