Skip to main content

How earnings work

You are paid per hour, not per token​

Your GPU is rented for time, and the meter does not care how many tokens move through it. A client sending one request an hour and a client saturating your card pay the same for the same wall-clock time — and you earn the same.

This is deliberate. Metering by token would make your income depend on how efficiently someone else wrote their prompts.

The clock starts at first serving​

A deployment goes through several stages before it is useful:

  1. Queued and matched to your GPU
  2. The engine image is pulled
  3. The model is downloaded and loaded into VRAM
  4. It starts serving ← the clock starts here
  5. The client uses it
  6. It stops

Stages 1 to 3 are unpaid, and they can take ten minutes or more on a large model. That time is not billed to the client, so it is not paid to you either.

The trade favours you in the long run: a host that is billed from creation would have to be priced lower to stay competitive.

What decides your rate​

Your hourly payout comes from your card's tier, which is set mainly by VRAM — that is what decides which models fit, and larger models rent for more.

The client pays your payout plus a 20% platform commission. Those are two different numbers, and the console shows you the one you receive.

VRAM matters more than speed​

A 24 GB card that is slower than a 12 GB card still earns more, because it can hold models the smaller card cannot run at all. When choosing what to list, or what to buy, VRAM is the number to look at.

Staying eligible for work​

You are offered jobs when your GPU is online and has enough free VRAM. Two things quietly cost hosts money:

  • Using the card yourself. Gaming or rendering leaves less free VRAM, so you are offered smaller jobs, or none.
  • Sleeping. A suspended machine stops serving, and any job in flight fails.

Failed jobs​

A job that fails on your machine is not billed, so it does not earn. Repeated failures also make you less likely to be matched.

The usual causes are avoidable: recreating the container mid-job, letting the machine sleep, running out of disk, or a network change that breaks the container's DNS — see Troubleshooting.

Gaps in monitoring​

If we lose sight of your agent for a while but the job container never restarted, that time counts as served and is billable. A gap where your GPU genuinely stopped serving is not.

You are not penalised for a brief connectivity blip on our side or yours, provided the work actually continued.