How it works
From request to payout
A two-sided market for GPU compute: renters bring workloads, providers bring silicon, and the marketplace handles matching, isolation, and settlement.
Request
A renter asks for capacity: GPU class, region, and a price ceiling. From the console, CLI, or API.
Match
The marketplace matches the request to verified hosts above their floor price. Typical match is seconds, not tickets.
Run
The workload runs in a hardware-isolated sandbox on the host. Usage is metered per second while it runs.
Settle
The renter's prepaid balance pays for metered seconds. The provider keeps 85%, accrued to a weekly payout.
If you need GPUs
- 1kracht launch --gpu h100
Request capacity with one command, an API call, or a click. No quota forms, no sales call.
- 2matched · attested · 18s
You get a host that passed driver, thermal, and benchmark checks. The device is wiped and re-attested before you land on it.
- 3running · $/s metered
Per-second billing from your prepaid balance while the job runs. Stop any time; billing stops with it.
- 4checkpoint · resume
Long jobs checkpoint to your persistent volume, so a stop or reschedule never loses your work.
If you have GPUs
- 1verify identity + hardware
KYC plus an automated hardware pass: drivers, thermals, and a benchmark. You control when you're listed.
- 2set floor price + windows
Your hardware, your terms. The marketplace only matches demand above your floor, inside your availability.
- 3sandboxed workloads run
Renters never get host access. Jobs run isolated, and the GPU is wiped between renters. Your machine stays yours.
- 4earn 85% · weekly payout
Earnings accrue per rented hour and pay out weekly once a payout account is connected. Idle and unrented costs you nothing.
$
Built for the day something breaks
A marketplace between strangers only works if the failure cases are boring. These are the three that matter.
A host degrades mid-job?
The job is rescheduled onto healthy capacity and resumes from its last checkpoint. The renter doesn't pay for the failed host's dead time, and the incident counts against the host's reliability score.
A renter misbehaves?
Workloads are sandboxed and hardware-isolated, so there's nothing on the host to reach. Renters are verified before they can launch, and abuse ends the session and the account.
A payment dispute?
Billing is metered per second against a prepaid balance, so there's no invoice surprise to dispute. Usage logs for every session are visible to both sides.
The isolation and attestation details live on the Security page. Reliability scoring and checkpoint behavior are described as designed; figures across this page are illustrative until the marketplace is live.
See it from your side of the market
Spin up a GPU in seconds, or find out what your idle hardware could earn.