Cloud GPU quotas, and working around a zero limit
On the large clouds a GPU is gated twice: by stock in the zone, and by a quota on the project. New projects often start with a GPU quota of zero, both per region and in a global cap across all regions, so the first create fails with a quota error before stock is even checked. An increase is a request filed from the console's quota page or the quota API, and the answer can take minutes or days. While it waits, the work can run somewhere that does not put GPUs behind a quota ticket. rent.dev starts GPU pods through one API, on the agent's own key at the GPU provider rent.dev connects to, and that provider bills the key directly.
rent.dev: one API for the sandboxes and GPU pods an agent needs.
GPU stock as the vendors read it
| Stock | Rows | GPU types | Lowest per GPU hour | GPU | At | Rate read |
|---|---|---|---|---|---|---|
| High | 6 | 2 | $0.34 | NVIDIA GeForce RTX 4090 | RunPod, community | 2026-11-11 |
| Medium | 10 | 4 | $0.59 | NVIDIA A40 | RunPod, secure | 2026-11-11 |
| Low | 57 | 26 | $0.13 | NVIDIA GeForce RTX 3070 | RunPod, community | 2026-11-11 |
| None | 24 | 14 | $0.12 | NVIDIA RTX A2000 | RunPod, community | 2026-11-11 |
Questions
- What to do when cloud GPU quota is zero?
- Read the error first: it names the quota metric, the region and the limit, and a limit of 0 means the project was never granted that GPU, not that the zone is out of stock. Request an increase for both the regional quota on that GPU model and the global GPU quota (one large cloud calls it GPUS_ALL_REGIONS), because raising one leaves the other at zero. Give a short use case and a modest count, start with one or two GPUs, and make sure the billing account is in good standing: new accounts with no payment history are refused more often.
- Are there GPU alternatives without a quota increase?
- Yes. Preemptible or spot GPUs sometimes carry a quota of their own that is granted separately, and a serverless GPU product on the same cloud may have its own limits. GPU-focused clouds and marketplaces usually let an account start a GPU once it has a payment method or credit on file, subject to their stock and their own spending limits. rent.dev's GPU pods run on your own key at the GPU provider rent.dev connects to, so that provider's account limits apply instead of a cloud project's quota. Of the GPU rows on rent.dev's price board whose vendor reads its stock as High, the cheapest is the NVIDIA GeForce RTX 4090 at $0.34 per GPU hour (RunPod, community), read 2026-11-11.
- How does per-second GPU billing work on cloud platforms?
- The meter starts when the instance or container starts and stops when it is deleted, counted in seconds, often after a one-minute minimum. The hourly rate on a pricing page is that per-second rate scaled to an hour. Boot time and image pulls are usually on the meter, and a stopped VM can still bill for its disks and reserved addresses.
- Why does a GPU request fail with quota exceeded when the region has capacity?
- Quota and capacity are separate checks. Quota is a ceiling the cloud sets on your project before it looks for hardware; capacity is whether a machine is free in that zone. A quota of zero fails every request everywhere, while a stock-out fails in one zone and may succeed in another. On rent.dev's price board, generated 2026-11-14, 97 GPU rows carry their vendor's stock reading: 6 High, 10 Medium, 57 Low and 24 None.