Onboard anywhere · Managed for free · Get paid for idle

Bring your own capacity

WEEKENDFleet capacity you already pay forMonTueWedThuFriSatSun100%0%one training run saturates SaturdayYour teamsIdle. Every hatched second is one we serve inference on and pay you for.69% average utilization.The other 31% of your GPU-hours is what earns.hourly · one week · a fleet nobody would call underused

Onboard your reserved fleet from anywhere. We run it for nothing, and pay you for the hours it sits idle.

Open the one-pager

One page, printable, safe to forward internally.

Included

We operate it. You are not billed for it.

Nobody on your side gets paged because a node wedged at 2am, and nobody has to plan the next driver upgrade.

There is no platform fee, no per-node charge, and no support tier. The inference running on your idle hours pays for the orchestration, and the rest comes back to you.

What your team gets

KubernetesLive
SlurmRolling out
GPU VMsLive
PodsLive
EndpointsLive

Idle capacity

nothing to configure

There is no switch to flip and no policy to write. A chip nobody on your side is using picks up inference, and the moment your team wants it back it is theirs again, near-immediately.

What we operate

Included · $0

  • Security patching, with the kernel held so apt cannot reboot a node under a running job
  • GPU health checks, with unhealthy cards delisted before they can fail a job
  • Node agent updates on a jittered timer, so the fleet never restarts at once
  • VFIO binding and IOMMU config, boot-persistent across reboots
  • Base image and VM template distribution, content-addressed and checksum-verified
  • Scheduling, routing, and metering per user and per project
  • The reconciliation that cleans up after all of it

What you get

One view of the whole fleet

Every node you onboarded, what is running on it, and who it belongs to. The nodes currently serving inference are marked, so the hours you are being paid for are never a mystery at the end of the month.

app.openrelay.inc/fleet

Nodes onboarded

24

3 sites

Your utilization

69%

trailing 7 days

Idle hours served

1,284

this month

NodeGPUsRunningUtilization
gpu-ord-018x H100vision-lab / train-7bKubernetes98%
gpu-ord-028x H100vision-lab / train-7bKubernetes97%
gpu-ord-038x A100nlp / eval-sweepPod62%
gpu-ord-044x L40Splatform / ci-runnersVM31%
gpu-sfo-018x H100OpenRelay inferenceEarning88%
gpu-sfo-028x H100OpenRelay inferenceEarning91%
Illustrative. Sample fleet, sample numbers.

Onboarding

Four steps, and the first one is the only one you type

No image to build, no Kubernetes to stand up first, and no firewall change. A node goes from a clean Ubuntu install to scheduling work in a single pass.

  1. 1

    Run one command

    On a clean Ubuntu install, with a provisioning token. It installs the agent, the runtime, and the GPU passthrough setup.

    curl …/enroll.sh | sudo bash
  2. 2

    The node dials out

    It opens a QUIC tunnel outbound and reports its GPUs, NVLink topology, IOMMU groups, RAM, and disk. You open no ports.

    orl nodes list
  3. 3

    Your teams start scheduling

    A VM, a pod, a cluster, or an inference endpoint, against the real topology we detected. Same API, same dashboard, same CLI as our own fleet.

    orl pods create --gpu h100 -n 8
  4. 4

    Idle hours start earning

    Nothing to configure. A chip nobody is using picks up inference, and your team gets it back near-immediately when they want it.

Control

Your hardware stays yours

Your work always outranks ours. A chip your team is using is never shared, never co-scheduled, and never queued behind a guest. Guests are isolated at the hypervisor level with whole GPUs handed through by VFIO, so a device only ever belongs to one tenant at a time.

Handing a chip back is near-immediate, because what we put on it is inference. In-flight requests finish in seconds, the GPU is released, and your job starts. There is nothing for you to configure and nothing to wait out.

ONE NODE, ONE DAYInference, interruptibleYour training jobInference resumesyou submitdrain · secondsGuest work is inference, measured in seconds. Your job does not queue behind a training run.

Building it yourself vs onboarding it

Same hardware either way. Different amount of your team's time, and a different bill at the end.

ConcernRoll your ownBring your own capacity
Time to first self-serve GPU jobWeeks of Kubernetes, device plugins, and image plumbingOne command per node
Scheduling surfacesPick one and maintain it foreverVMs, pods, clusters, and endpoints on the same fleet
Inbound network exposureBastion plus public IPs or a VPN to maintainNone, nodes dial out over QUIC
GPU passthrough and NVLink topologyHand-managed VFIO binding and IOMMU groupsDetected and scheduled against automatically
Per-team usage and chargebackBuild it from Prometheus and a spreadsheetMetered per user and project out of the box
Patching, drivers, and fleet healthSomeone's weekend, foreverOperated for you, included
What the platform costsSalariesNothing. The idle hours fund it.
What idle capacity earnsNothing. It just depreciates.Inference revenue share on the unused hours

Frequently asked questions

What is bring your own capacity?
You keep the hardware and the contracts. OpenRelay supplies the control plane that turns a fleet into a cloud: enrollment, scheduling, GPU passthrough, VMs, pods, clusters, inference endpoints, SSH, metering, and fleet health. When your teams are not using the chips, that capacity joins the OpenRelay inference network, serves paid traffic, and earns you a share of the revenue.
Where can the capacity live?
Anywhere with outbound internet. On-prem racks, a colo cage, a university cluster, or reserved and committed instances inside your own cloud account. Nodes need outbound HTTPS and outbound UDP 443, which is the QUIC tunnel the data plane rides on. Nothing inbound, so they can sit behind NAT with no public IP.
How do you make sure I can get my own chips back?
Near-immediately, and without you doing anything. There is no releasable flag to set and no policy to write. Everything we place on your fleet is interruptible inference, measured in seconds, so when your team wants a chip the in-flight requests finish and it is theirs again. Your workloads always take priority on your own hardware.
What does the platform cost?
Nothing. There is no platform fee, no per-node charge, and no support tier. We make our money serving inference on the hours your fleet is idle, which means the orchestration, the patching, the driver and firmware work, the health checks, and the scheduler are all included. If your fleet were somehow busy every hour of every week, we would make nothing, and you would still pay nothing.
How much do idle chips earn?
It depends on the GPU model and how much of the time the capacity is actually free, since you are only paid for hours served. High-demand parts on a fleet that sits idle overnight and on weekends earn the most. We model the number against your real utilization before you sign anything.
How is this different from becoming a provider?
Becoming a provider is renting your hardware out as the whole business: you are supplying capacity to other people and the revenue share is the point. Bring your own capacity is platform-first. Your teams are the primary tenant, and the payout exists to stop the remaining hours from being wasted. Same enrollment path, different center of gravity.
Is anyone else's workload on my machine while mine is running?
Not on a chip your team is using. Guest inference only lands on capacity that is otherwise sitting idle, and guests are isolated at the hypervisor and GPU level, with VFIO passthrough handing whole devices to a single tenant rather than sharing them. Nothing is co-scheduled alongside your job on the same GPU.
What hardware can I enroll?
x86_64 machines with IOMMU (Intel VT-d or AMD-Vi), 32 GB of RAM or more, an NVMe SSD, and a clean Ubuntu install. NVIDIA RTX 3090 through 5090, A100, H100, H200, and B200, plus AMD Instinct parts. CPU-only nodes are welcome too.
How do I get started?
Book a call. We size the fleet, confirm the hardware is eligible, model what the idle hours are worth, and walk through enrollment on one node before you commit the rest. Most customers have a first workload running on their own hardware the same week.

Find out what your idle hours are worth

Tell us the GPU models, the node count, and roughly how busy they are. We will come back with what the fleet earns and what it takes to onboard it.