Browse the marketplace
Open Rent GPUs in the sidebar. Every card on that page is an offer: a rentable slice of one real machine, holding 1, 2, 4 or 8 GPUs. The price shown is for the whole slice, not per GPU, and Superheat sets it, not the machine's owner. One host can list more than one slice size, so two cards may live on the same physical box.
Selecting a card is the first of two steps; the second is Deploy.
What a card shows
| Field | What it means |
|---|---|
| GPU model | The card in the slice, with NVIDIA · ×N underneath |
×N | How many GPUs the slice contains. The price covers all of them |
| Price | Hourly rate for the whole slice, metered per second while the instance runs |
| VRAM per card | Memory on one GPU. Multiply by N for the total the slice gives you |
| TFLOPS | The offer's advertised throughput figure, and what the TFLOPS sort ranks on |
| Storage | Total disk on the host machine. This is the ceiling on the disk you can request at deploy |
| Location | Country flag plus the region code, such as US-CA or FI-HEL |
| Reliability | The host's reliability score as a percentage. A shield icon means a verified host |
| Rented | Badge on an offer that is already rented. Only visible with Show unavailable on |
Clicking a card selects it and raises a bar across the bottom of the page with the model, price, VRAM, disk and location, and a Continue button. Clicking the same card again deselects it.
More details
More details on any card opens a drawer with the full machine spec. The drawer works on rented offers too, so you can inspect hardware you cannot currently take.
| Field | What it means |
|---|---|
| Total VRAM | VRAM per card × number of GPUs |
| Max CUDA | The highest CUDA version this machine supports. The deploy page checks template CUDA floors against it |
| PCIe | Generation and lane width, for example 4.0 ×16 |
| CPU, CPU cores, RAM | The host machine's CPU model, core count and RAM |
| Disk | Host disk size and type (NVME, SSD or HDD) |
| Network | Advertised uplink and downlink |
| Host | The machine's hostname |
| Storage rate | What the disk costs per GB per month, quoted from the per-hour rate |
| Max duration | Shown only when the offer declares one — the longest rental the machine advertises |
The drawer's button reads Rent this GPU and takes you straight to the deploy page. On a rented offer it reads Currently rented and is disabled.
Filters
The filter row sits above the grid. Filters combine — an offer has to satisfy all of them — and the current filter state lives in the page URL, so a filtered view can be bookmarked or sent to a teammate.
| Control | Values | What it matches |
|---|---|---|
| Search | Free text | GPU model, hostname or region, case-insensitive substring |
| GPU family | Families present in the catalog, such as Hopper, Ampere, Ada, Blackwell | The architecture family derived from the model name |
| Region | Countries present in the catalog | The machine's country |
| # GPUs | 1×, 2×, 4×, 8× | Slices with exactly that many GPUs |
| Min VRAM per card | 24, 32, 48, 80, 141 GB+ | Offers at or above that VRAM per card |
| Max price | ≤ $0.5, $1, $2, $5, $15, $30 /hr | Slice price at or below the cap |
# GPUs, Min VRAM per card and Max price live behind the More filters button, which switches to a filled style while any of the three is set.
If nothing matches you get an empty state rather than a partial list. Widen the price cap or clear the family and region filters first — those three cut the most.
Most of these have an equivalent on GET /v1/offers — num_gpus, min_vram_gb, region and a maximum hourly price — but three differences are worth knowing. The API filters on an exact gpu_model, while the console's family filter groups models by architecture. The API's region accepts a region code or a country code, while the console's Region control lists countries. And the free-text search box has no API equivalent; it runs in the browser over the offers already loaded. See Offers.
Sorting
| Option | Order |
|---|---|
| Price: low → high | Cheapest slice first. This is the default |
| Price: high → low | Most expensive slice first |
| TFLOPS | Highest advertised throughput first |
| VRAM | Highest VRAM per card first |
Sorting applies after filtering, and price sorts are on the slice price — an 8× slice is not cheap because its per-GPU cost is low.
Show unavailable
The Show unavailable switch at the right end of the filter row brings rented offers into the grid. They render dimmed, carry a Rented badge, and cannot be selected; More details still opens.
This is the include_rented flag on GET /v1/offers, which defaults to false. Offers the fleet has taken out of service entirely are never returned either way.
Turn it on when you want to know whether the hardware you are after exists at all. An empty grid with the switch off and a full grid with it on means the machines are there and busy, which is a different problem from having filtered too hard.
Catalog freshness banner
A banner above the filters reads "Catalog last updated Nm ago — listings may be out of date." when the console's catalog refresh has fallen behind, or "Catalog has not synced yet — listings may be out of date." before the first refresh lands.
The console re-checks catalog freshness every 30 seconds and shows the banner once the last successful refresh is more than five minutes old. A refresh that fails leaves the previous listing in place rather than blanking the page, which is why you see a stale marketplace with a warning instead of an empty one.
While the banner is up, treat prices and availability as approximate. Deploying still works; the offer is checked again server-side at that moment.
From offer to deploy
Three routes reach the same place:
- Select a card, then Continue on the bottom bar.
- Open More details, then Rent this GPU.
- Start from a template's Launch button, which drops you into the marketplace with that template carried along, then follows either of the routes above.
Availability is decided when you press deploy, not when you select the card. If someone reaches the same machine first, the deploy fails with an offer-unavailable error and you land back here. See Deploy an instance.