# Put an idle GPU on the route

> Closed beta: run the UserBus worker on an idle NVIDIA GPU, set your price per delivered second of video, and take the jobs whose cap covers it.
>
> HTML version: https://userbus.com/providers/

You run the UserBus worker on your own machine and name your price per delivered second of video. Buyers name the most they will pay. When a job’s cap covers your price and no cheaper machine is idle, the job is yours.

**Closed beta, for Linux with an NVIDIA card.** Tell us about your machine. A person reads every request, and invites the Google account of those we can take to Fleet, the provider console.

[Send your request](https://studio.userbus.com/providers/request)

## What you need today

These requirements are those of the beta’s one experimental model integration. The model for a commercial launch has not been selected, and its requirements may differ.

**An invitation**: Invite-only access: send a request about your machine, and a person decides on each one. Fleet, the provider console, uses Google sign-in and refuses accounts that were not invited.

**Linux and an NVIDIA card**: The only admitted runtime is CUDA on Linux. (Measured) Our recorded run used an RTX 4090 with 24 GB and peaked near 22 GiB of video memory, with the runtime using about 48 GB of system memory.

**Disk for the model**: The beta model’s weights are about 41 GiB across five files. (Measured)

**Not admitted**: Apple Silicon: the Metal and MLX runtimes are disabled because no run of the model exists on them. Nothing is recorded for Windows or for other NVIDIA cards, so we make no claim about them.

## How the money is meant to work

- You set a rate per delivered second of video, in whole cents, for each worker and model. A worker with no rate gets no jobs.
- A job can reach you only if its cap is at or above your rate, and a claim is refused while a cheaper machine is idle. Price is how you compete.
- The buyer is charged your rate, not their cap. Your earning for a job is your rate times the seconds billed.
- Pay is for delivered video, not for time online. An idle hour earns nothing, and a failed job is not charged to the buyer.
- Each worker runs one job at a time and holds a short queue – two jobs by default.

**Payouts are not available yet.** Fleet records earnings, but this release moves no money: its payout actions are switched off. UserBus has not announced a platform fee.

[How provider pay works](https://userbus.com/providers/earnings/)

## From request to first job

1. Send your request about your machine. A person reads it, writes back by email and invites your Google account.
2. Sign in to Fleet with Google.
3. Create a worker. Fleet shows its secret once; only a hash is stored, so it cannot be recovered later.
4. Install the worker and the model’s runtime on the machine, and start it with that secret.
5. Set your rate for the model.
6. Stay online. Jobs that fit arrive on their own.

[Send your request](https://studio.userbus.com/providers/request)

[Already invited? Open Fleet](https://fleet.userbus.com/)

## What Fleet shows you

Per worker: jobs delivered, attempts per delivery, failures by reason, time from claim to finish, when the worker was online, and earnings. A figure that cannot be measured yet is shown as unknown with the reason, never as zero.

Fleet does not show prompts, job settings or videos. It also makes no earnings projection and does not rank providers.

## What you can and cannot see of a job

The prompt reaches the runtime as data, stays out of the worker’s logs, and is not embedded in the delivered file. A worker’s credential cannot download buyers’ results. This is housekeeping, not a hardware guarantee: buyers are told to treat prompts as not confidential.