Renting compute from strangers only works if both sides know exactly what is guaranteed. Here is the honest version — including the parts we are still building.
When a host installs the agent, it generates a keypair on the machine and signs a report of its CPU, RAM, and GPU model. We verify that signature before the GPU can be listed, so a listing cannot be forged or altered in transit.
What this is not: it is not a hardware root of trust. A determined host could still report hardware it does not have. Reputation from completed jobs is the stronger signal, and it is shown on every listing.
Your workload runs in a Docker container on the host, with privileges dropped, a process limit, and a memory cap. Where the host has gVisor installed, we run the container under a user-space kernel for a stronger boundary between your job and their machine.
What this is not: containers are not a hardware boundary. Do not put data on a node that you could not tolerate the host seeing. For sensitive work, use a host that advertises confidential computing, or don't use shared infrastructure.
When you book, the money leaves your wallet and is held by Petabyte for the rental. The host is paid on completion; the platform takes 10% of the rental. This is an internal ledger, not an on-chain escrow — Petabyte is the custodian.
Stop early and you are billed only for the hours you held the machine (minimum one). The unused prepay returns to your wallet.
Hosts send a heartbeat. If one goes quiet, we move your machine to another eligible node — the address you connect to does not change — and restore from the most recent snapshot. If no node can take it, the rental is refunded.
Be aware: recovery is from the last checkpoint, not a live mirror. A failover means restarting from a snapshot, not zero data loss.
We would rather be trusted than impressive. These are on the roadmap and are not live today:
If a claim matters for your workload, ask us before you book — security@petabyte.market.