Managed Compute

Managed compute rents GPU and CPU machines from providers such as RunPod, Vast, and Lambda, billed at cost to your organization's prepaid balance. Experiments can use it as the openresearch backend, and you can also create standalone instances from the web or the orx CLI.

Sign in

orx login
orx orgs

orx login signs in through your browser, then offers to register ~/.ssh/id_ed25519.pub (creating it if needed) so managed machines accept this computer. orx orgs lists the organizations you can bill. See Install if orx isn't on your PATH yet.

Use it for experiments

In the app, open Compute > OpenResearch to check your login, organization, and SSH key, or make it the default backend. Each run on it rents the machine you ask for when it starts and releases it when it finishes, so you only pay for the run. Pass a GPU id from the catalog (optionally with a count) or a CPU flavor with its vCPUs, or save a default with orx compute default set openresearch --flavor.

orx exp run <expId> --backend openresearch --flavor h100_sxm
orx exp run <expId> --backend openresearch --flavor cpu5c:8

A supervisor on the machine handles cleanup, so closing your laptop doesn't leave it running. Copy anything you need into the run's logs or artifacts: files left only on the machine are lost when it's released.

Browse offers

# GPU offers
orx compute
orx compute catalog --gpu H100_SXM --count 1 --provider runpod

# CPU offers
orx compute catalog --cpu

GPU identifiers are the exact values in the GPU column. The web dashboard's Compute tab shows the same live prices and availability.

Create a standalone instance

Get your organization ID from orx orgs or the /orgs/<orgId> URL. Omit --provider to pick the cheapest matching GPU offer.

# GPU (disk defaults to 100 GB)
orx instance create <orgId> --gpu H100_SXM --count 1 --disk 100

# CPU: flavor cpu5c, cpu5g, or cpu5m with 2, 8, or 32 vCPUs
orx instance create <orgId> --cpu cpu5c --vcpus 8

List and terminate instances

orx instance list <orgId>
orx instance delete <instanceId>

Standalone instances keep billing until you delete them. Deleting destroys the provider machine, so copy anything you need first. The web dashboard's Instances tab shows the same machines with their SSH details.