Search 20,000+ GPUs, launch containers on bare metal, and drive the API gateway — one command surface, scriptable end to end.
$ openlink search offers 'gpu_name=H200 num_gpus>=4 reliability>0.99' -o dph --limit 4
ID GPU NUM VRAM $/HR PERF REGION
847291 H200 4 564GB $3.53 99.1 US-West
851003 B200 4 768GB $4.95 99.4 US-East
849177 RTX PRO 6000 WS 8 768GB $1.07 97.8 EU-West
843320 RTX 5090 8 256GB $0.41 96.2 US-Central
$ openlink create instance 847291 --image nvcr.io/nvidia/pytorch:24.01-py3 --disk 128 --ssh
{"success": true, "instance_id": 9841205}
Query the marketplace by GPU model, VRAM, price, reliability, region, interconnect and 20+ other fields. Sort by performance-per-dollar.
Run any Docker image on bare-metal GPUs. Pass environment variables, expose ports, mount volumes. SSH and Jupyter included.
Instances, volumes, templates, API keys, billing and gateway usage — all from one command surface, all scriptable.
$ pip install openlink
$ openlink set api-key YOUR_API_KEY
✓ Authenticated as acme-labsRequires Python 3.9+. Also available via pipx for an isolated install, and as a static binary for CI runners.
Prices move with supply and demand. Filter on the fields that matter to your job, then sort by dollars-per-hour or by performance-per-dollar.
$ openlink search offers 'gpu_name=RTX_5090 num_gpus>=8' -o dph
ID GPU NUM VRAM $/HR REGION
843320 RTX 5090 8 256GB $0.41 US-Central
843471 RTX 5090 8 256GB $0.44 EU-West
845902 RTX 4090 8 192GB $0.35 US-EastAny public or private registry image runs on bare metal. Mount volumes, expose ports, inject secrets, and get a direct SSH endpoint back.
$ openlink create instance 843320 \
--image nvcr.io/nvidia/pytorch:24.01-py3 \
--disk 128 --ssh --direct \
--env HF_TOKEN=$HF_TOKEN
{"success": true, "instance_id": 9841205}
ssh root@ssh4.openlink.ai -p 41022Check status, stream logs, exec into a running container, and destroy the instance the second the job finishes. Nothing bills while nothing runs.
$ openlink show instances
ID GPU STATUS IMAGE $/HR
9841205 8x RTX 5090 running pytorch:24.01-py3 $0.41
$ openlink logs 9841205 --tail 3
epoch 12/40 loss 1.284 lr 3.0e-4 12.4k tok/s
$ openlink destroy instance 9841205
✓ Destroyed. Billed 4h 12m — $1.72GPU model, VRAM, price, reliability, region, bandwidth, compute capability, interconnect and more.
Public or private registries. PyTorch, TensorFlow, vLLM, custom images — whatever your job needs.
Bid on interruptible capacity, or reserve on-demand. RTX 4090 from $0.35/hr.
Direct SSH into the container, or launch a notebook with --jupyter and get a signed URL.
openlink copy moves datasets and checkpoints between instances, or to and from local.
Every command takes --raw for clean JSON. Pipe to jq, wire into CI, automate the whole loop.
Transparent, programmatically queryable, no contracts and no minimums.
pip install openlink on Python 3.9 or newer, then authenticate with openlink set api-key YOUR_API_KEY. For an isolated install use pipx install openlink. A static binary is available for CI runners that have no Python.--raw and returns JSON on stdout with a non-zero exit code on failure, so it composes with jq and standard shell tooling. Authenticate in CI by setting OPENLINK_API_KEY as an environment variable instead of writing a config file.openlink --help for the full command tree, or openlink <command> --help for flags on any subcommand. The complete reference lives in the docs.One OpenAI-compatible endpoint for 300+ models, with routing and failover.
API Gateway →Search GPUs, deploy containers and script the whole workflow from your shell.
You are hereA typed client for compute and models. Build agents that scale their own GPUs.
Python SDK →No contracts, no minimums. Get an API key and start building.