Browsing the catalog

GET /offers is the endpoint you will spend most of your requests on. It returns what is rentable right now, filtered however you like.

curl -s -G https://api.gpuoutlet.ai/v1/offers \
  -H "Authorization: Bearer $GPUOUTLET_API_KEY" \
  -d gpu_family=h100-80gb \
  -d region=us-east-1 \
  -d available=true \
  -d sort=price_asc

Requires catalog:read. Cursor-paginated — see Pagination, and note in particular that a short page is not the last page.

Filters

All are optional and combine with AND.

By GPU

string

Exact SKU code, e.g. h100-sxm-80gb. The narrowest filter — one specific card in one specific form factor. Values come from /gpu-models.

string

Family id, e.g. h100-80gb. Broader than gpu_model: covers every variant of a card, so h100-80gb matches both SXM and PCIe. This is usually the filter you want — most workloads care about “an H100”, not about the interconnect. Values come from /gpu-families.

integer

Minimum VRAM per GPU, in gibibytes. Filter by what your model needs (min_vram_gb=48) rather than by name, and you will match cards you had not thought to ask for.

string

Free-text over GPU name and code. For a human-typed search box. Prefer the structured filters in code — q=h100 also matches things you did not mean.

By location

string

Exact region code, e.g. us-east-1. From /regions.

string

Coarse bucket: us, eu, asia, other. Use it when the requirement is “keep the data on this continent” rather than a specific datacentre.

By price and shape

integer

Ceiling on the hourly price, in cents, per GPU. max_price_per_hour_cents=300 is “no more than $3.00/hr”.

'standard' | 'dedicated'

standard — the shared-host pool, cheapest and most plentiful. dedicated — the whole host is yours, for workloads that cannot tolerate a neighbour.

'on_demand' | 'spot'

Only offers rentable under this tier. spot is cheaper and interruptible — see Spot rentals.

By availability

'true' | 'false'

true returns only offers with capacity right now.This is a snapshot, not a reservation. An offer can be taken between your read and your POST /rentals, which is why creation can still answer offer_unavailable. Handle that; do not assume available=true guarantees a machine.

Sorting

string

default:"price_asc"

price_asc · price_desc · vram_asc · vram_desc · availability_desc

availability_desc puts the deepest capacity first — the right sort when you need eight GPUs and would rather not discover the shortage at creation time.

Paging

string

Opaque; from meta.next_cursor. Valid only for the exact query that produced it.

integer

default:"50"

1–200.

What an offer looks like

{
  "id": "off_3kd91mz7",
  "gpu_model": "h100-sxm-80gb",
  "gpu_name": "NVIDIA H100 SXM",
  "gpu_family": "h100-80gb",
  "gpu_arch": "Hopper",
  "gpu_vendor": "NVIDIA",
  "vram_gb": 80,
  "ram_gb": 192,
  "vcpu": 24,
  "max_gpus": 8,
  "instance_class": "standard",
  "region": "us-east-1",
  "region_name": "US East",
  "geo": "us",
  "currency": "usd",
  "price_per_hour_cents": 249,
  "spot_price_per_hour_cents": 137,
  "availability": { "kind": "count", "count": 6 },
  "service_tiers": ["on_demand", "spot"]
}

string

required

Opaque, and stable for as long as the offer exists — but offers are transient. Do not store one and expect it to be rentable tomorrow; re-read the catalog.

integer

required

Per GPU, per hour, in cents. A two-GPU rental of this offer costs twice this. (Note that the quote reports the price for the whole rental — the two are different numbers on purpose, and the quote is the one to show a user.)

integer | null

The interruptible price, when this offer sells one. null means it does not.

integer | null

The largest gpu_count this offer can serve. null when the upstream does not say.

string | null

Join key to /gpu-families and /prices. null for models outside a known family.

string[]

required

Which tiers this offer can be rented under right now. This array is the authority on what you can buy; the price fields only say how much.It cuts both ways. An offer with a spot_price_per_hour_cents but no spot here cannot currently be taken as spot. An offer with no on_demand here cannot be taken on-demand either, even though price_per_hour_cents is still populated — some machines are sold at their spot rate only, and the on-demand figure is then the reference that rate undercuts rather than an offer.Send the tier you want on POST /rentals. Omitting it means on-demand, so an integration that never sets it will be refused on a spot-only offer rather than quietly charged a different rate.

Reading `availability`

This field is a discriminated union. Check kind before reading anything else:

{ "kind": "count", "count": 6 }

{ "kind": "signal", "level": "high" }

av = offer["availability"]
if av["kind"] == "count":
    free = av["count"]              # exact
else:
    free = {"high": 5, "medium": 2, "low": 1}[av["level"]]   # your own heuristic

Some upstreams report an exact number of free units; others only report a rough level. Rather than invent a number for the second case — which would be a fabricated integer indistinguishable from a real one — we tell you which kind of answer you are holding.

{"kind": "count", "count": 0} and available=false mean the same thing. Do not write if offer["availability"]["count"]: without handling the signal case — a KeyError on a perfectly valid response is the most common first-integration bug on this endpoint.

One offer by id

curl -s https://api.gpuoutlet.ai/v1/offers/off_3kd91mz7 \
  -H "Authorization: Bearer $GPUOUTLET_API_KEY"

Returns the offer object directly, not wrapped in data. 404 offer_not_found covers both “no such offer” and “no longer visible” — reported identically on purpose.

Worked examples

Cheapest H100 available right now, anywhere

/offers?gpu_family=h100-80gb&available=true&sort=price_asc&limit=1

Anything with at least 48 GB, in Europe, under $2/hr

/offers?min_vram_gb=48&geo=eu&max_price_per_hour_cents=200&available=true

Filtering by VRAM rather than by model is what makes this useful: it will surface an L40S or an A6000 you would not have thought to name.

Room for an 8-GPU job

/offers?gpu_family=h100-80gb&available=true&sort=availability_desc

Then check max_gpus >= 8 and, where availability.kind == "count", count >= 8. Neither field alone is sufficient: max_gpus is what the offer can serve, count is what is free.

Cheap interruptible capacity for a checkpointed job

/offers?service_tier=spot&available=true&sort=price_asc

Only for work that can be killed and resumed — see Spot and preemption.

A whole host to yourself

/offers?instance_class=dedicated&gpu_family=h100-80gb&available=true

Get started

API keys

Conventions

Catalog

Rentals

Radar

Browsing the catalog

Filters

By GPU

By location

By price and shape

By availability

Sorting

Paging

What an offer looks like

Reading `availability`

One offer by id

Worked examples

​Filters

​By GPU

​By location

​By price and shape

​By availability

​Sorting

​Paging

​What an offer looks like

​Reading availability

​One offer by id

​Worked examples

Filters

By GPU

By location

By price and shape

By availability

Sorting

Paging

What an offer looks like

Reading `availability`

One offer by id

Worked examples