Skip to main content
The part of your wallet you can spend on a new rental. Excludes reserved (held against running rentals). Shown in the top-right pill.
Automatic Stripe charge to refill your wallet when it dips below a threshold. Requires a saved card. See Topping up.
A GPU model line (e.g. h100-80gb). Distinct from a listing, which is one provider’s specific offering within that family.
The canonical string identifier for a family in the API: rtx-4090, rtx-5090, a100-80gb, h100-80gb, b200, b300.
Append-only history of every wallet movement. Read via GET /v1/billing/ledger.
One provider’s specific offering: “vast.ai, US-East, 1× H100 SXM 80GB, $2.80/hr”. The catalog returns many listings per family; the rental orchestrator picks the cheapest matching one.
Slang for an instance. We use them interchangeably across the docs.
Our billing model — you deposit credit, instances burn it per second. No invoices, no monthly bills.
An upstream marketplace or cloud we source GPUs from. We hide differences behind a unified API; you don’t usually interact with the provider name (it’s surfaced read-only on rental details).
The portion of your wallet held against currently-running rentals. Each rental reserves 1 hour at its rate. As the rental runs, the reserve drains.
A single instance lifecycle from launch to settlement. Identified by a rent_… ULID. Synonyms: instance, pod.
The starting image for a pod: ubuntu-24-04, pytorch, vllm, etc. See Instances.
The 26-character ID format used everywhere (01HZ3J9X8K2P5T7R4W1Q6Y8M0E). Sortable by creation time, URL-safe, no hyphens. Prefixed with type (usr_, rent_, etc.) for legibility.
GPU memory, in GB. The number you care about for whether your model fits.
The per-user credit balance. There’s exactly one per account. Three numbers: available, reserved, total.
Outbound HTTP call from Stripe to our server confirming an event (checkout.session.completed, setup_intent.succeeded, etc.). We require these to trust that Stripe really collected the money.