Every card on the marketplace, with the rate it is listing at today and the range it has traded in. On-demand, interruptible or reserved — pick the one that matches how much interruption your job can take.
Rates move with supply and demand and are shown for reference. The observed range covers listings we have seen for that model, outliers included — a single mispriced machine can stretch the top end, so read the median column first.
The hardware is identical. What differs is how much interruption you will accept, and how much that willingness is worth.
The machine is yours until you destroy it. Nothing reclaims it mid-run.
Cheaper capacity that can be reclaimed when someone bids higher. Best for work that checkpoints.
A block held for a fixed term at an agreed rate. For steady load that never really stops.
Providers set their own prices and compete for your job. We publish what they are asking rather than setting a number ourselves.
No minimum hours, no rounding up to the next block. A four-minute job bills four minutes.
From a GTX 1080 for cheap batch work to a B300 for frontier training — the whole spectrum behind one key and one bill.
On-demand needs no contract. Scale up, scale down, switch GPU types, or leave — nothing to unwind.
Everything on this page is available from the CLI and the SDK, so procurement can be scripted rather than screenshotted.
These rates are what hosts set. If you have idle cards, the same market works in your favour — see the earnings calculator.
No contract, no minimum. Rent one card for an hour and see whether the numbers hold.
