Browse GPUs

GET /gpus returns the nodes you can rent right now — online, accepting rentals, priced, and not already taken.

curl https://paraloncloud.com/api/v1/gpus \
  -H "Authorization: Bearer prlc_your_key_here"

Any active key can call this — it doesn't need the rental scope (that's only for actually starting a rental).

Response

{
  "gpus": [
    {
      "node_id": "ab728006-0282-40bc-bc67-70b40a0a36b8",
      "gpus": [
        { "model": "NVIDIA GeForce RTX 4090", "vram_mb": 24564, "compute_cap": "8.9" }
      ],
      "price_per_hour": 0.39,
      "verified": true,
      "country": "Germany",
      "country_code": "DE"
    }
  ],
  "total": 1
}

Fields

Field	Meaning
node_id	The unique node identifier. This is what you rent with — pass it to POST /rentals.
gpus	The GPUs in the node (a node may have several). Each has model, vram_mb, and compute_cap.
price_per_hour	What the rental bills per hour. The rate is locked at creation for the life of the rental.
verified	Whether the node passed hardware verification.
country / country_code	The node's country (full name + ISO code). Filter on country_code — it's stable across geo lookups.

There is no node name and no city in the response — a node's name and city are the provider's private information and aren't needed to rent. You always rent by node_id.

Reading compute_cap

compute_cap is the GPU's CUDA compute capability — the hardware/architecture version, not raw speed. It decides which frameworks and precisions will run:

compute_cap	Architecture	Notable support
6.x	Pascal	No tensor cores — FP16 is emulated and slow
7.0–7.5	Volta / Turing	Tensor cores; FP16
8.0–8.6	Ampere	+ BF16
8.9	Ada Lovelace	+ FP8
9.0	Hopper	+ FP8 (data-center)
12.0	Blackwell	Newest consumer

For modern LLM inference (FP8 weights / KV cache) you want 8.9 or higher. A cheap node with lots of VRAM but a low compute capability may not run your stack at all — check compute_cap before you rent, not just VRAM and price.

Two nodes can look identical by GPU model — tell them apart by node_id, price, VRAM and country.