Regions & availability
How region and hardware availability affect which GPU offers show up in the catalog, and how to widen a search.
GPU capacity is pooled across a global network, so what shows up in the catalog — and at what price — depends on region and current demand for that hardware class.
How offers are matched
Your account region is applied to the catalog automatically, so results default to the hardware closest to you. Widen what you see by relaxing filters like GPU model or max price — see Browse offers.
When a GPU class is scarce
Popular GPUs (H100s especially) can be scarce at peak times. If the catalog comes back empty or thin, try a nearby GPU class, relax the max price filter, or check back shortly — offers refresh continuously as capacity turns over.
What scarcity looks like in the catalog
| Signal | What it means | What to try |
|---|---|---|
| Catalog returns few or no offers for a GPU model | Demand for that hardware class currently exceeds available capacity | Relax the GPU model filter, or check back shortly — availability shifts as instances are returned to the pool. |
| Offers exist but prices look unusually high | Scarce hardware still has some supply, priced at a premium | Compare against a nearby GPU model's price for the same job before paying the premium. |
| A larger count (e.g. multi-GPU) is unavailable but a single GPU is not | Multi-GPU hosts are rarer than single-GPU hosts for that class | Consider whether the job can start smaller, or split across more, smaller launches. |
Practical tips
- Search with a slightly higher max price first to see the full spread of what's available, then tighten it once you know the going rate for that GPU class.
- If your workload can tolerate a less powerful card, checking a secondary GPU model in the catalog widens your odds of an immediate match instead of waiting on a scarce class.
- For a launch that must happen at a specific time, browse and launch a little ahead of schedule rather than at the exact moment you need the instance.
- Availability tends to loosen outside peak hours — if the job isn't time-critical, checking back later can turn up capacity that wasn't there before.
Tip
For predictable long-running work, model servers and fine-tuning jobs both draw from the same pool, so the same availability guidance applies.