GPUaS.site
Rows of Frontier supercomputer cabinets with dense blue and red liquid-cooling hoses under yellow cable trays
Frontier (HPE Cray EX, AMD Instinct GPUs), Oak Ridge National Laboratory, USAPhoto: OLCF at ORNL · CC BY 2.0 · source

GPU infrastructure exchange

Every GPU listing, with the country it really runs in.

Compare source-backed GPU cloud offers by price per GPU-hour, memory and the physical data-center country. Each row shows its source and when GPUaS last fetched it.

Open the GPU exchangeEstimate a job's total cost
  • Free search and comparison
  • Complete job cost, or what is missing
  • Source, fetch time and host country on every row
Lowest price per GPU-hour, by modelLoading listings…

Real hardware

Modern AI runs on rows of liquid-cooled GPU cabinets like these. GPUaS helps you find where comparable capacity is offered, and in which country it physically runs.

End of a Frontier supercomputer row with the Frontier compass logo and blue-cabled GPU cabinets
Frontier, Oak Ridge National Laboratory, USAPhoto: OLCF at ORNL · CC BY 2.0 · source
Rows of Summit supercomputer cabinets at Oak Ridge
Summit (IBM, NVIDIA V100 GPUs), Oak Ridge National Laboratory, USAPhoto: OLCF at ORNL · CC BY 2.0 · source
NVIDIA HGX B200 eight-GPU baseboard
NVIDIA DGX B200 8-way GPU board (HGX)Photo: Pokiiri · CC BY-SA 4.0 · source
JUWELS supercomputer cabinets at Forschungszentrum Jülich
JUWELS (Booster module with NVIDIA A100 GPUs), Forschungszentrum Jülich, GermanyPhoto: Forschungszentrum Jülich · CC BY-SA 4.0 · source

Photos of public supercomputers (Frontier and Summit at Oak Ridge National Laboratory, JUWELS at Forschungszentrum Jülich) and an NVIDIA HGX board, each credited under its Creative Commons licence. They illustrate GPU infrastructure and are not facilities of any provider listed on GPUaS.

Estimate the complete cost of a job

Describe the work and GPUaS checks fit, adds storage and transfer, and shows what is still unknown.

Describe your workload

Workload type

Memory each GPU allocation must provide. A MIG slice counts as its own size, not the full card.

More requirements

Optional. These refine the cost estimate and the eligibility checks.

GPUs in one allocation (1–1024).

Leave empty for no limit.

Language never changes currency. Offers in another currency are compared only with an approved exchange-rate snapshot.

You do not need to know a GPU model to start.

How comparisons work

Every result goes through the same three checks, in this order.

  1. Fit

    Usable VRAM per allocation, GPU count, region and interruptibility are checked against your hard requirements. Unknown values are never treated as a match.

  2. Complete cost

    Compute, storage, data transfer and required fees are added for your runtime. If a component is unknown, the estimate is marked incomplete with a known lower bound.

  3. Freshness

    Each offer shows its source and when it was last checked. Stale or expired evidence needs confirmation before you rely on it.

Ranking is organic: providers cannot pay for position, and a paid software license is not a verification badge.

Already providing GPU capacity?

Use the GPUaS provider portal for your own customers: offers, quote requests, versioned quotes and direct-pay order tracking.

See the provider portal