Offers and slices
When you open Rent GPUs, every card on the page is an offer. Picking one is the first half of a deployment — the second half is picking a template.
What an offer is
An offer is a rentable shape on one physical machine in the Superheat fleet: 1, 2, 4 or 8 GPUs of a single model, published at one hourly price.
A machine is not the unit you rent, and an offer is not a single item on a shelf. One machine publishes every shape it can serve, all drawn from the same pool of free GPUs. An idle 2-GPU host is simultaneously two 1× offers and one 2× offer — the same two cards, described two ways.
So a card in the marketplace carries a count, not just a yes/no. Renting one 1× on that host leaves the other 1× on sale and takes the 2× off it, because there is no longer a pair to sell.
What you rent is the slice. The GPUs in it are yours alone — a card is never split between two renters — but the machine around them is not: another renter may hold the neighbouring card. See What you share on a machine.
Your slice stays yours until the instance on it is destroyed, including while that instance is stopped. This is why stopping still holds your slot.
Why the price is per slice
The hourly price on the card covers the whole slice, not one GPU. A 4× offer at $3.20/hr costs $3.20/hr, not $12.80/hr.
Two consequences worth internalizing:
- Comparing a 1× card to an 8× card by their headline prices compares different amounts of hardware. Compare offers of the same shape, or divide by the GPU count yourself.
- The GPU price is not the whole bill. Disk is quoted separately, per GB per hour, and it keeps accruing while the instance is stopped. The deploy page adds both into one estimate.
Superheat sets prices. Machine owners do not bid, and there is no auction or spot market — the number on the card is the number you pay for as long as you hold the offer.
What the fields on a card mean
| Field | What it tells you |
|---|---|
| GPU model | The NVIDIA model in the slice, for example RTX 4090 or H100 PCIe. |
NVIDIA · ×N | How many GPUs are in this slice: 1, 2, 4 or 8. |
| Price | The hourly rate for the whole slice. GPU only — disk is extra. |
| VRAM per card | Memory on one GPU. Multiply by the GPU count for the total the slice gives you. |
| TFLOPS | The throughput figure published for this offer. Useful for ranking offers against each other, not as a benchmark result. |
| Storage | The physical machine's total disk. This is the ceiling on the disk you can attach at deploy time. |
| Location | The machine's region, with its country flag. See Regions. |
| Reliability | The host's measured reliability, as a percentage. A shield marks a verified host. |
| Available | How many of this shape the machine can serve right now, out of its ceiling — for example 1 of 2. Shown only where the machine can serve more than one, so a single-GPU host stays uncluttered. |
N left | The same number beside the selection dot, as a reminder while you scan. |
| All rented | Shown instead of the selection dot when every unit of this shape is taken. Those cards cannot be selected. |
More details opens a panel with the rest of the machine: total VRAM across the slice, the maximum CUDA version the host supports, PCIe generation and width, CPU model and core count, RAM, disk size and type, network up and down, the hostname, and the storage rate quoted per GB per month. Offers that carry a maximum rental duration show it here as well.
These figures are what the deploy page checks a template's constraints against: the machine's maximum CUDA version against the template's CUDA floor, VRAM per card against its VRAM floor, plus the architecture it needs and whether the host can run a virtual machine. An incompatible template is disabled in the picker with the reason attached, so a mismatch shows up before you deploy rather than after.
Availability and staleness
An offer is listed or retired, and separately it has a count of how many are free. "Sold out" is a count of zero, not a status — a shape can be rented by someone else and still have units left for you.
The marketplace hides sold-out offers by default; the Show sold out toggle brings them back, grayed out, which is useful for judging what the fleet holds even when you cannot rent it right now. Retired offers — a machine that left the fleet, or a model no longer sold — are never shown.
The catalog is refreshed on a poll. When that poll falls behind, the marketplace shows a banner saying how old the listings are. A failed refresh leaves the last known catalog in place — the page goes stale, never empty — so treat the banner as a warning that a card you select may already be gone.
The count on a card is a snapshot; it is already slightly out of date by the time you read it. If two people take the last unit at the same time, the first deploy wins and the second is sent back to the marketplace with OFFER_UNAVAILABLE. Nothing is charged for a deploy that loses this race.