How it works
The marketplace, end to end.
Everything that happens between a GPU you need and a provider who got paid, as searchable topics. Open any one for the full explanation.
Launch. One command matches a verified host and opens a shell on the card.
Want the whole path in order instead? Follow one rental end to end, from the command to the payout.
Knowledge base
How the marketplace works
Search the mechanics or browse by topic. Open a card for the detail in full; the link to where the claim is proven stays inside the panel, so you decide whether to leave.
Renting
Ask for cards, a region, and a price ceiling from the console, CLI, or API. No quota form and no sales call.
You get the card to yourself for the whole rental; the container and scratch volume are destroyed when it ends.
Per-second billing from a prepaid balance. Stop the job and billing stops with it, to the second.
A stop can snapshot your workspace first, and the next rental restores it. Nothing is kept automatically.
Pick a region and the work runs on a dedicated host there. The region is snapshotted with the rental.
You top a balance up and spend it down by the second. Nothing runs once it is empty.
Launch from the console, the CLI, or the API. Same rental, same meter, whichever door you come in.
Providing
KYC plus an automated hardware pass: drivers, thermals, and a benchmark. You control when you are listed.
Your hardware, your terms. The market only matches demand above your floor, inside your availability.
Renters never get host access. Jobs run sandboxed and the workspace is wiped when a rental ends.
One script brings up the agent that talks to the control plane. Nothing to configure per rental.
Whenever you like, with one command. There is no term to see out, and unrented time costs you nothing.
If your box drops mid-rental, the worker rescues the stranded rentals and settles at the last heartbeat.
Billing
Usage is metered per second while the job runs, paid from the renter's prepaid balance, never a rounded-up hour.
The provider keeps 85%. The split is exact-integer, so no cent leaks in the rounding.
Balances are sums of ledger entries, never edited columns, and every posting is idempotent.
The ledger is in EUR, stored to six decimals. Money is never a float, so per-second rounding cannot drift.
Account spend caps are enforced by the worker, so a runaway job stops even when nobody is watching.
Marketplace
The request is matched to verified hosts above their floor price. A typical match is seconds, not a ticket queue.
The agreed price is frozen when a rental starts, so a provider re-pricing cannot change a running job.
Every host clears identity and an automated hardware check before it can take a rental.
GPU models and regions are catalog entries keyed by code, not typed-in strings.
Security
One tenant per card, workloads sandboxed, and neighbours completely walled off.
The container and scratch volume are destroyed at the end of every rental. Nothing of the workload survives.
TLS on every call, encrypted at rest by the database provider, credentials stored only as hashes.
Outbound network is off unless you enable it, so a workload cannot phone home on its own.
Payouts
Earnings accrue to the provider's balance for every second rented, settled to the microeuro.
Payout runs are still manual, so there is no cadence to promise yet, and this page will not invent one.
Payments run on EU rails, iDEAL and SEPA, through an EU provider rather than a US processor.
One market, three products
Compute gets the hardware. Studio builds and runs the workload on it. Autopilot watches what that costs.
Compute
Rent GPUs by the second from idle hardware, or earn from the cards you own. Priced at match, settled in euro.
Manage
plannedDraw the workflow and attach hardware to each step, then read a finished run back step by step.
Optimise
plannedSee what the work costs and where it's leaking. Simulates a change against what you already paid, applies nothing until you approve.
03Proven, dated
Exercised against the live stack.
The mechanics above, in the order they were proven to work. Each row names the issue or the document that records it.
A GPU rental ran end to end on the 7900 XT: passthrough scoped to one card, the iGPU left out, failed sessions charged nothing, every hold released.
Money conserved to the microeuro across the first settlements: renter −0.009019, provider +0.007665, platform 15.01%.
Launch, meter and settle against the live stack, then the connect hop: a renter peered with the hub and read files out of the running container.
#37The worker settled a 27-tick rental under real load, and the provider agent survived a reboot and two control-plane redeploys, re-registering each time.
A renter SSHed into a GPU rental on the default image: one card scoped (card1, renderD128), /dev/kfd openable, the provider's own sshd blocked from inside.
#199Snapshot and restore end to end: a rental stopped with a snapshot, and a second rental restored from it served back the file the first one wrote.
#186A fresh Ubuntu 24.04 box provisioned through the documented one-liner installed the agent, built the instance image itself and kept a session off the host's sshd.
#155
Spin up a GPU, or earn from your own
Rent on-demand compute by the second, or list idle hardware and let it pay for itself. Same marketplace, both directions.
No card required
- 85%
- revenue to providers
- Per-second
- billing, live
- EUR
- metered to the microeuro