Public Nodes Feed

Every GPU node currently rentable on Lium, as one unauthenticated JSON call. No key, no signup, no rate-limit negotiation — point your ingester at it and start listing us today.

Resource	URL
Live feed	`https://lium.io/api/public/v1/nodes`
OpenAPI schema	`https://lium.io/api/openapi.json`
Per-model price table (min/max live ask, availability, page URL — one row per GPU model, hourly)	`https://lium.io/pricing.json` — fields on Pricing Pages and Price Feed

curl -sSL https://lium.io/api/public/v1/nodes

Response

{
  "generated_at": "2026-08-17T09:14:03.117Z",
  "count": 95,
  "nodes": [
    {
      "id": "ad83de56-8a04-4ab1-ac5a-4ab007b4e2b3",
      "gpu_model": "RTX 5090",
      "gpu_count": 8,
      "available_gpu_count": 8,
      "price_per_gpu_hour": 0.55,
      "price_per_node_hour": 4.4,
      "min_rentable_gpu_count": 8,
      "gpu_memory_gb": 32,
      "cpu_count": 384,
      "ram_gb": 283,
      "disk_gb": 2460,
      "cpu_model": "AMD EPYC 9654 96-Core Processor",
      "network_upload_mbps": 1057.7,
      "network_download_mbps": 2535.7,
      "uptime_minutes": 4815,
      "reliability_score": 97.5,
      "disk_type": "nvme",
      "tier": "secure",
      "country": "Japan",
      "country_code": "JP",
      "region": "13",
      "city": "Tokyo",
      "rent_url": "https://lium.io/browse-pods/ad83de56-8a04-4ab1-ac5a-4ab007b4e2b3",
      "disk_free_gb": 1435,
      "gpu_driver": "570.133.20",
      "cuda_driver": "12.8",
      "gpu_interconnect": "nvlink",
      "region_name": "Tokyo",
      "timezone": "Asia/Tokyo"
    }
  ]
}

Fields

Field	Type	Notes
`generated_at`	ISO 8601	When this snapshot was built. Use it to detect a stale source.
`count`	integer	Number of nodes in this response.
`id`	UUID	Stable identifier for the node. Safe to use as your primary key.
`gpu_model`	string, nullable	Normalized name, vendor prefixes stripped — `RTX 5090`, not `NVIDIA GeForce RTX 5090`.
`gpu_count`	integer, nullable	GPUs on the node.
`available_gpu_count`	integer, nullable	GPUs free right now. Always at least 1 — a fully rented node leaves the feed.
`price_per_gpu_hour`	float, nullable	USD per GPU per hour.
`price_per_node_hour`	float, nullable	USD for the whole node per hour, `null` when the price is unknown. Published so you never have to derive it.
`min_rentable_gpu_count`	integer, nullable	Smallest order we accept on this node. Quoting a single GPU below this number sells an order our checkout refuses.
`gpu_memory_gb`	integer, nullable	VRAM per GPU.
`cpu_count`	integer, nullable	CPU cores on the whole node; a partial rental gets a proportional share.
`ram_gb`	integer, nullable	System RAM of the whole node; a partial rental gets a proportional share.
`disk_gb`	integer, nullable	Whole-node disk; a partial rental gets a proportional share.
`cpu_model`	string, nullable	CPU model string as the host reports it, at most 128 characters, for example `AMD EPYC 9654 96-Core Processor`.
`network_upload_mbps`	float, nullable	Measured upload speed of the node in Mbps, one decimal. `null` until Lium has measured it.
`network_download_mbps`	float, nullable	Measured download speed of the node in Mbps, one decimal. `null` until Lium has measured it.
`uptime_minutes`	integer, nullable	Minutes Lium has seen this node online, counted 15 per validation cycle since the node was first listed. This is cumulative online time, not the current streak.
`reliability_score`	float, nullable	Provider reliability over a rolling 14-day window of rentals other than voluntary spot, 0 to 100 with one decimal, recomputed every 12 hours. It counts penalized rentals and rentals that failed to start because of the provider's node. A failed start does not count when it failed on the renter's own image (building, pulling or running it, or the registry login); when the renter deleted the pod before it started; when it failed on Lium's side before the request reached the node; or when it was a reboot of a pod that had started. It is scored per provider, so every node of one provider carries the same value. The Lium browse page shows the same score rounded to a whole percent. `null` means the provider had neither a billed rental nor a counted failed start in the window, so there is no score; rank an unscored node as unknown, not as 0.
`disk_type`	string, nullable	Kind of disk under the node's container storage, as Lium reads it off the kernel: `nvme`, `ssd`, `hdd`, or `unknown` when it looked and could not tell (a virtual or stacked device, or an unreadable reading). `null` until Lium has reported it. Rank an `unknown` or `null` node as unknown, not as `hdd`.
`tier`	string, nullable	`secure` or `spot`. A spot node may be reclaimed by its provider at any time.
`country`, `country_code`, `region`, `city`	string, nullable	Node location; `region` is the subdivision code.
`rent_url`	string	Deep link straight to the rent page for this node.
`disk_free_gb`	integer, nullable	Free disk on the whole node at the last scrape; a partial rental gets a proportional share. `0` is a full disk, `null` means not yet reported.
`gpu_driver`	string, nullable	NVIDIA driver build as the host reports it, at most 32 characters, for example `570.133.20`.
`cuda_driver`	string, nullable	Highest CUDA version the installed driver supports, as `major.minor`, for example `12.8`. Compare as a version, not as a number.
`gpu_interconnect`	string, nullable	`nvlink` when every GPU pair on the node is joined by NVLink, `pcie` when at least one pair is not. `null` on a single-GPU node, and until our validator has scraped the topology.
`region_name`	string, nullable	Name of the subdivision `region` is the code of, for example `Tokyo` for `13`.
`timezone`	string, nullable	IANA time zone of the node's city, for example `Asia/Tokyo`.

Only id and rent_url are guaranteed non-null; treat every other field as optional in your ingester.

Location is published to city level and no finer: postal code and coordinates are not in the feed and will not be added to /public/v1.

Freshness

Lium prices are dynamic — providers set them and they move. The feed is rebuilt at most once per minute, so polling faster than that returns the same snapshot. Once a minute is the right cadence.

Rate limit

60 requests per minute per IP. If you need more, get in touch and we will raise it for your ingester.

Compatibility promise

This feed is a published contract, not an internal endpoint that happens to be reachable. Concretely:

Fields are only ever added. No field under /public/v1 will be removed or renamed.
A field's meaning never changes. If price_per_gpu_hour ever needed to mean something different, it would become a new field, not a redefinition of this one.
Breaking changes ship as /public/v2. They never land on /public/v1.
A retired version keeps serving for six months after we publish the deprecation notice, so you have half a year to migrate without an outage.

Fields added so far, newest first:

21 Sep 2026: disk_free_gb, gpu_driver, cuda_driver, gpu_interconnect, region_name, timezone — all nullable; a consumer built against the 12 Sep shape keeps working unchanged.
19 Sep 2026: disk_type.
12 Sep 2026: cpu_model, network_upload_mbps, network_download_mbps, uptime_minutes, reliability_score.

Build against /public/v1/nodes and it will keep working.

Listing Lium on your site

If you run a GPU price comparison site, marketplace, or agent that provisions compute, this feed is everything you need to list us. Ingest it directly — you get correct prices, real availability, and a deploy link that converts.

Questions about a partnership, a higher rate limit, or a custom field: reach us at lium.io.

Response​

Fields​

Freshness​

Rate limit​

Compatibility promise​

Listing Lium on your site​