Put your idle GPUs to work
Workstations, lab machines and on-prem clusters spend most of their lives waiting. Roundhouse routes real inference work to that spare capacity — jobs from organizations that need compute and, often, own no AI hardware at all. The marketplace is running now, in a small invitation-only beta, on non-sensitive workloads only.
Invitation only while we are small — tell us what hardware you have and we will be in touch. What the beta pays, and when.
Install one client. Complete jobs. Get paid per token.
Install the client
The same lightweight, pip-installed worker agent that powers corporate on-prem routing. Point it at whatever OpenAI-compatible server your machine already runs.
Take on real AI jobs
The router dispatches inference jobs your models can handle, at the capability tier they qualify for — from bulk classification to standard language work.
Earn per token completed
Every execution is metered to the token in the routing core. When the marketplace opens, that metering becomes your earnings statement.
Want to work out whether it would be worth it? The worksheet does the arithmetic on your own assumptions — your rate, your utilization guess, your electricity price. It won't pre-fill a number, because we haven't set a rate and the demand doesn't exist yet. Open the capacity worksheet →
What the beta is, and what it pays
The marketplace is running now, in a small invitation-only beta, on non-sensitive workloads only. Real jobs from real organizations are already completing on hardware their owners don’t operate.
We are not publishing a per-token rate yet. Rates set during a beta with a handful of machines would not survive contact with a real market, and we would rather set one we can keep than one we have to walk back. So we are telling you the terms instead of the number.
Your earnings accumulate from the first job your machine completes — every execution is already metered to the token, and that record is what you get paid from. Once your balance reaches $100 we pay it out by ACH, free. Below that it keeps accruing; nothing expires.
You can withdraw earlier, for a flat $2.50. That is roughly what it costs us to send an ACH payment, and it is charged only when you ask to be paid ahead of the threshold — waiting for $100 costs you nothing. On a small balance that fee is a large share of it: withdraw at $10 and you receive $7.50. We will show you the exact figure before you confirm, because a fee you discover afterwards is a different thing from one you agreed to.
What this means in practice: you can plug in a machine, watch what it actually serves, and decide later whether the rate we set is worth your electricity. If it isn’t, take what you earned and go. One honest caveat: while your account is open we won’t send a payment below $2.50, because the fee would consume the whole of it and handing you nothing is not a service — it keeps accruing instead. Close the account and we pay whatever is left anyway, absorbing the transfer cost. There is no floor below which your earnings stop being yours.
Already true of the technology
- Pull-based, outbound-only client. Outbound HTTPS requests only, zero inbound ports — nothing on the internet can connect in to your machine.
- Works with common local stacks. Ollama, vLLM, LM Studio, llama.cpp — any OpenAI-compatible server.
- Per-token metering in the core. Every execution is already metered to the token; compensation plugs into accounting that exists today.
- Deadline-aware dispatch. Jobs are routed and re-checked every 30 seconds, so an offline machine never strands work.
Provider questions, answered straight
What hardware qualifies?
Anything that can serve a model through an OpenAI-compatible server: a single GPU workstation, a lab machine that sits idle overnight, or a full on-prem cluster. If it runs Ollama, vLLM, LM Studio, or llama.cpp, it can take jobs.
How will I get paid?
Compensation is per token of completed work, powered by the same metering that already tracks every execution in the router. Rates and payout details are being finalized — the waitlist gets you early access and first word.
Is it safe to run on my network?
The client is pull-based: it makes outbound HTTPS requests only and needs zero inbound ports, so nothing on the internet can connect in to your machine. It works behind NAT and firewalls without network changes.
Is the capacity marketplace live today?
Yes — in a small invitation-only beta, on non-sensitive workloads only. The routing core, the client and per-token metering are all in production. Apply and tell us what hardware you have.
What kind of jobs would my hardware run?
AI inference jobs submitted by organizations, each declaring a capability tier — basic, standard, or frontier. Your hardware is offered jobs the models it serves can handle.
Do I have to commit to uptime?
You choose when the client runs. The router plans around real availability: queued work is re-checked every 30 seconds, so if your machine goes offline, its jobs are simply re-planned and routed elsewhere.
Be first in line when capacity opens
Tell us what hardware you have — even roughly — and we’ll let you know the moment provider onboarding begins.
Run AI workloads instead? See how Roundhouse routes every job to the right model in the right place.