Skip to content
Runs on your own provider accounts

The unified interface for GPU processing

Set up your API keys once. We find an available machine — bare metal or a served model — start your workload on it, and replace it if it fails.

Market pricesUpdated daily
Selected GPU

RTX 5090

32 GB
Listed from
$0.250/hr
Up to
$0.990/hr
Up to 31B models3 clouds listing it4.0× spread
Vast.aiRunPodOne API
34 clouds tracked46 GPU types$0.090/hr lowest listed rate list prices, refreshed daily
What Gridline handles

You ask for a GPU. We handle the rest.

Set up your API keys once, then draw on them from anywhere.

  1. 1

    Add your keys once

    Paste the API keys for the clouds you already use. They are stored once, and every workload after that draws on them — no setup per project.

  2. 2

    Ask for compute

    Bare metal or a served model, whichever the job needs. We pick the machine with enough memory and start it on your own account.

  3. 3

    Replace failures

    If the host drops, we replay the prefix on another machine and finish the stream. No gap, no duplicate tokens.

Host recovery

A failed GPU should not fail your request.

This stream starts on one machine. When that host drops, another machine finishes the response.

POST /v1/chat/completionsstream: true
RTX 3090 · 7f3a1c · $0.109/hrstreaming
A generation that began on one GPU and was completed by another after the first failed, with no duplicated or missing tokens.
Bare metal and inference

Two shapes of compute, one place to ask.

Take a bare-metal box with root and SSH when you want to train, fine-tune, or run your own stack. Ask for a model instead and we serve it behind an OpenAI-compatible endpoint. Tell us the workload and we pick which one fits, and which GPU to put it on.

ModelCurrent route
Gemma 4 E4B8,192 context
RTX 3090106 t/s · 44 streams
$0.090/hrlisted from
Gemma 4 12B8,192 context
RTX 309067 t/s · 15 streams
$0.090/hrlisted from
Gemma 4 31B8,192 context
RTX 509052 t/s · 8 streams
$0.250/hrlisted from
Memory checkedHost checkedFallback ready
Get started

Create an account and send a request.

Sign up, add your provider keys once, and ask for compute — a bare-metal box or a served model — without wiring up a new account each time.