Run your own AI. The easy way.
Pick a model. gpuyard finds a GPU that fits it, sets the machine up, deploys the model and gives you an endpoint that speaks the OpenAI API. It adds machines when traffic grows, and when traffic stops it gives them back, so an idle model costs you nothing.
No Kubernetes, no drivers, no autoscaler to tune.
- Any open model
- Pick it from Hugging Face.
- Your own endpoint
- Works with the OpenAI SDK you already use.
- Scales by itself
- More GPUs when busy, none when quiet.
Your part is two lines long
You
- Pick a modelAny open model on Hugging Face, searched from the Deploy screen.
- Call the endpointThe OpenAI chat API, streaming included. Your SDK already speaks it.
gpuyard
- Reads the model card and works out how much GPU memory the weights need
- Finds the best machine that holds them, across every market and lent server
- Rents it, installs the NVIDIA driver and Docker, and starts the engine
- Routes your requests to it and adds machines when the ones running fill up
- Replaces a machine that goes away, on any cloud or vendor
- Shuts the machine down when requests stop
gpuyard picks the best GPU for your model. What runs where:
| Model | Needs | Runs on |
|---|---|---|
| qwen3.8-27b | 67 GB | 1× A100 80 GB or 1× RTX PRO 6000 96 GB |
| minimax-m2.5 | 277 GB | 4× H100 80 GB |
| deepseek-v4-flash-vision-exp | 369 GB | 4× H200 141 GB |
Scales up when you’re busy, and down to zero when you’re not
When a GPU gets busy, gpuyard starts another one and splits the requests between them. When requests stop, it waits as long as you chose, from 5 minutes to 4 hours, then shuts the GPU down. If a machine goes away, gpuyard finds a replacement on any cloud or vendor.
Have idle GPUs? Earn from them
Turn free data-centre capacity into spot GPUs for AI. One command sets a server up, and one call takes it back.
Or write to hello@gpuyard.cc.
See every GPU and what it costs, live
One screen shows your GPUs, your models, what you spend per hour, and a chat to test any model.
| GPU | Where | $/hr | Status | Model |
|---|---|---|---|---|
| A100 · 80 GB | verda · fi-hel | 0.79 | Ready | qwen3.8-27b |
| RTX PRO 6000 · 96 GB | vast.ai · us-tx | 0.92 | Ready | qwen3.8-27b |
| 4× H100 · 80 GB | lent server · de-fra | 7.40 | Starting | minimax-m2.5 |
| A100 · 80 GB | vast.ai · pl-waw | 0.74 | Shutting down | — |
Active requests by model
Chat
YouWrite a SQL query that finds duplicate emails.
qwen3.8-27b · verda · fi-helSELECT email, COUNT(*) FROM users GROUP BY email HAVING COUNT(*) > 1;