An OpenAI-compatible API for open models, paid by the second

RentGPUPower's Compute API speaks OpenAI's language. Point the OpenAI SDK — or any tool built on it — at our address with your account key, name an open model, and you get chat completions, streaming included. Behind it, a GPU serves the model for you, and you pay that GPU by the second instead of paying per token.

Two settings

The API runs on prepaid credits: top up your balance, and calls draw from it.

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://rentgpupower.com/api/v1/compute",
    api_key="gpuapi_...",
    max_retries=20,  # keeps trying while the model starts
)

stream = client.chat.completions.create(
    model="qwen3:8b",
    messages=[{"role": "user", "content": "Write a haiku about GPUs."}],
    stream=True,
)
for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="", flush=True)

curl

curl https://rentgpupower.com/api/v1/compute/chat/completions \
  -H "Authorization: Bearer $GPURENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3.2:1b", "messages": [{"role": "user", "content": "Hello!"}]}'

Which models

• Any Ollama tag, such as llama3.2:1b, llama3.1:8b, qwen3:8b or gpt-oss:20b.
• A vetted vLLM model from Hugging Face: Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen2.5-1.5B-Instruct, Qwen/Qwen2.5-Coder-7B-Instruct, Qwen/Qwen2.5-14B-Instruct or mistralai/Mistral-7B-Instruct-v0.3.

client.models.list() (GET /models) shows what is running for you and what can be started. Not sure how big a model your work needs? How much GPU memory do I need? has the table.

The first call: a GPU starts for you

If none of your instances serves the model you name, one is started on your credits. The call answers 503 with the error type model_loading, an estimate in seconds (eta_s) and a Retry-After header — and the same call is served as soon as the model is ready. The OpenAI SDKs retry a 503 on their own and wait as long as Retry-After says, which is what max_retries is for. A warm machine, with the model already on its disk, answers in seconds; otherwise the model is downloaded first, which takes a few minutes.

Later calls go to the same instance while it runs, so only the first one waits. Meanwhile, GET /instances/ID/ready tells you whether its model server answers yet.

When a call does not go through

Each key can have its own spend limit, and a read-only key can look but never spend — handy for dashboards.

What it costs

There is no per-token charge. You pay the instance's GPU by the second, at the machine's hourly price, for as long as it runs — whether or not you are calling it. An instance started this way is reserved for half an hour: extend it, or stop it early with DELETE /instances/ID (gpurent stop ID from the command line), and the unused time returns to your credits. To choose the length yourself, start one with POST /instances and a duration_s, then call it the same way. The docs list every endpoint.

Embeddings too

POST /embeddings is also available, in OpenAI's format. It is computed by a pool of phones and browsers that earn on /earn, with a small sentence-embedding model, all-MiniLM-L6-v2 (384 dimensions).

FAQ

Do I need to rent a GPU first?

No. Ask for a model and an instance starts for you. You can also start one yourself with POST /instances and call it the same way.

Is it really billed per second?

Yes: GPU time by the second, with no per-token fee. An instance you leave running costs the same whether you call it or not, so stop it when you are done.

Does streaming work?

Yes. Pass stream=True (or "stream": true), exactly as with OpenAI.

Will my tool work with it?

Most tools that let you set an OpenAI base URL and key work unchanged — the OpenAI SDKs for Python and Node, and the libraries built on them. To run a model on a machine you rent yourself instead, see How to run Llama, Qwen, DeepSeek or gpt-oss on a rented GPU.