Compute for data

A GPU hour costs money.An hour at the arm earns it.

If you already own a robot arm, your recordings are the one thing on this platform that cannot be rented. Sell them, or record against somebody else’s open request, and pay for your cloud training and inference out of what you made. This page does the arithmetic in both directions and says plainly where the loop is still manual.

10 %
Platform share on an hour of Collection work
10 %
Platform share on a dataset sale
€0.30
Estimated cloud inference, per hour
two
Ledgers on one account: earnings and credit

One account, two ledgers

Earnings

Dataset sales and released Collection work land here. This ledger pays out. It does not pay for compute.

Credit

Cloud training, cloud inference and the budget of a request you post yourself are billed from here. Top-ups and subscription credit land here.

Same account, same page, separate books. Moving a balance from the left column to the right one is a payout followed by a top-up, and you do both yourself.

The loop

Record, sell, compute. And the one step that is not a button.

The loop works. It is just not automatic, and a page that claimed otherwise would send you looking for a switch that does not exist.

  1. 01

    Record on the arm you already own

    Desktop client, your arm, your cameras, your table. One run from start position to end position is one episode.

  2. 02

    Sell it, or serve an open request

    Two routes out of the same recordings. Publish a dataset in the marketplace at your own price, or record against an open Collection where somebody else already set the task and the hourly rate.

  3. 03

    It lands in your earnings ledger

    A sale is credited net, after the 10 percent platform share. Collection work is held first and released once the requester has accepted your episodes.

  4. 04

    You request a payout

    From €10 upwards. The transfer is executed by hand, so treat it as days, not seconds.

  5. 05

    You top up the credit ledger

    Also from €10 upwards. Or by subscription, where the monthly fee comes back in full as training credit.

  6. 06

    You train and run inference from that credit

    A training run rents a card from the pool and is billed by its runtime. Cloud inference is billed the same way. Both refuse to start on an empty credit ledger.

What does not happen by itself

Earnings do not turn into GPU credit automatically. There is no path in the platform that moves a cent from the earnings ledger to the credit ledger, and this page will not pretend there is. The route is payout, then top-up, both with a €10 floor, and the payout leg is carried out by a person.

What is genuinely closed is the other direction. The credit ledger pays for compute and for data out of the same balance: the money that funds a training run also funds the budget of a request you post yourself. Buying somebody else’s recordings and buying GPU time come out of one pot.

Post such a request yourself

The shortcut, if you would rather not wait for a first sale

A subscription is the one place where money you pay comes back as compute at full value: the monthly fee is credited to the credit ledger at the same amount. Starter costs €49.99 and credits €49.99. Pro costs €299.99 and credits €299.99.

That credit does not expire. Nothing in the billing code reduces a balance because of its age, so an unused month is still there the month after.

See plans and top-ups
Before you start

What you need, and what you can skip.

The expensive half of robot learning is not the card. It is the hour somebody spent demonstrating the task.

You do need

  • An arm and the desktop client

    Recording happens on real hardware: the arm, the cameras and the client from the download page. This is the part nobody can lend you.

    If the arm already stands in a university lab
  • A task worth repeating

    One task, many runs, objects moved between them, failed attempts deleted. Ten clean minutes of one task are worth more than an hour across five.

  • An account

    Selling, serving a Collection, training and inference all sit behind the same login.

  • A starting balance for the very first run

    Training and inference read the credit ledger before they start. Either you top up once, or you earn first and move it across yourself.

You do not need

  • Your own GPU

    Training and inference rent cards from a pool. The model decides which tier is requested, and the run is billed by the runtime of the machine that was actually rented.

    Where the credit for a rented card comes from
  • A Python setup

    Nothing on the training path asks you to install CUDA, torch or a trainer, or to keep them in step with each other.

  • A dataset of your own to begin with

    A training run accepts a public Hugging Face dataset repository, a free marketplace listing or one you bought. Your own recordings make the policy fit your table; they are not the entry ticket.

  • Rare hardware

    An SO-100 is enough to record and to publish. Rarer arms answer questions nobody else can answer, which is a different page and a different argument.

    The page for the rarer arm

Three things that look possible here and are not

The browser simulator does not record datasets.

The 3D arm on this site really is steered by your keyboard, but its recording panel is a mock-up and carries a sample-data label. Real episodes come off real hardware.

The public live arm does not run your policy.

That arm is teleoperated. It is there to show latency and control from a browser, not to serve inference.

Local inference is free, but only on your own card.

If you own a GPU, running a checkpoint locally costs nothing and skips the cold start entirely. Renting is what you pay for when you do not.

Both sides in one table

Turn an hour at the arm into GPU hours.

Left is what the work earns you. Right is what that buys back. Move the sliders until it looks like your week.

€8€40

This is what the requester pays per hour of recorded material. The country rates on the platform run from 8 to 22 euro per hour and that band is marked on the track. Above it you are looking at your own assumption, not a rate anyone can show you.

20 sec180 sec

Anything from 5 to 1,800 seconds is accepted. Shorter episodes mean more of them per hour, each worth less.

1 hr20 hr
What the work earns
Price of one episode
€0.28
Episodes per hour
60
Platform share, 10 percent
-€0.03
Your net per episode
€0.25
Your net per hour
€15.00

€90.00

per week of recording

What that buys back
Hours of cloud inference
300 hr
Sessions, at the 4 hour billing cap
75

At the estimate shown before a start, 30 cents per hour. The booking is the runtime of the pod that was actually rented, which is usually less.

Training is deliberately not converted here. A run is billed by its own runtime and the pool buys in dollars, so the honest form of that number is the range in the table above, not a euro figure invented on this page.

Clears both the payout floor and the top-up floor, so this amount can be moved to the credit ledger.

The fee arithmetic is the same code the platform bills with, not a copy of it. The sliders are your assumptions, and nothing here is a forecast.

Amounts are net of the platform share and before whatever tax applies where you live.

Recordings you already have on disk can be a listing today, free or priced, your call.

Open the marketplace
The hardware question

Buy a card, or rent one per run?

For somebody recording at home the answer turns on how often they train, not on how much they could spend.

What it decidesYour own cardRented from the pool
Before the first runOne purchase, paid whether you train or notNothing
Cost of a training runPower, and a card that is busy for hoursRuntime times the hourly price of the machine that was rented, with no markup on top
InferenceFree, and it answers immediatelyBilled per hour, with a cold start of roughly 7.5 min before the first action
Large modelsCapped by the memory you boughtThe pool requests a tier per model, so a large model gets a large card
What it does not solveYou still need an arm to recordYou still need an arm to record

Both columns end on the same sentence, and that is the argument of this page. The card is the half of the problem you can rent. The recordings are the half you cannot, and that is also the half you can sell.

What a run costs today

Two tiers, because the model decides which card is requested.

ACT and SmolVLA

Typical runtime
2 to 5 hours
Hourly price of the card
0.30 to 0.60 USD
Estimated for the run
about 1 to 3 USD

GR00T and Pi0.5

Typical runtime
3 to 6 hours
Hourly price of the card
1.20 to 2.00 USD
Estimated for the run
about 4 to 12 USD

Spot market prices, quoted in US dollars because that is how the pool buys them. Your credit ledger is in euro and the conversion happens at booking, at a rate that is deliberately rounded in your favour.

The card is rentable. Your recordings are not.

Anyone can rent a GPU by the hour. Nobody can rent the hour you spent demonstrating your task on your arm. That is exactly why that hour is the one worth selling.