Cloud GPUs, from hourly cost to dev boxes

A cloud GPU's hourly cost spans more than an order of magnitude, from older consumer cards on marketplace hosts to data-center GPUs in eight-GPU nodes. For a developer box the card matters less than the habits around it: a box left on overnight can cost more than the gap between two providers' rates. rent.dev starts GPU pods on your own key at the GPU provider rent.dev connects to; a pod can expose SSH, and it runs until you terminate it.

rent.dev: one API for the sandboxes and GPU pods an agent needs.

The lowest hourly rate at each memory size

Rates as each vendor publishes them, normalised per GPU, from rent.dev's price board, generated 2026-11-14. See the whole board. Browse the full catalog.

MemoryGPULowest per GPU hourAtStockRate read
2 GBgtx1050 2Gi PCIe$0.02Akash2026-11-11
4 GBg2$0.65AWS2026-11-11
5 GBp2000 5Gi PCIe$0.24Akash2026-11-11
6 GBNVIDIA RTX A2000$0.12RunPod, communityNone2026-11-11
8 GBm4000 8Gi PCIe$0.04Akash2026-11-11
10 GBNVIDIA GeForce RTX 3080$0.17RunPod, communityNone2026-11-11
12 GBrtx3060 12Gi PCIe$0.06Akash2026-11-11
16 GBTesla V100 16GB$0.122Verda, spot2026-11-11
20 GBNVIDIA RTX A4500$0.19RunPod, communityLow2026-11-11
24 GBp40 24Gi PCIe$0.14Akash2026-11-11
32 GBv100 32Gi PCIe$0.20Akash2026-11-11
40 GBA100 SXM4 40GB$0.837Verda, spot2026-11-11
48 GBNVIDIA RTX A6000$0.33RunPod, communityNone2026-11-11
80 GBA100 SXM4 80GB$1.1695Verda, spot2026-11-11
94 GBNVIDIA H100 NVL$2.59RunPod, communityNone2026-11-11
96 GBRTX PRO 6000 machine$1.09Beam2026-11-11
141 GBH200 machine$2.09Beam2026-11-11
143 GBNVIDIA H200 NVL$3.79RunPod, secureNone2026-11-11
180 GBNvidia B200$3.59Daytona, preemptible2026-11-11
186 GBNVIDIA GB200 NVL72$10.50CoreWeave, on demand2026-11-11
192 GBAMD Instinct MI300X$1.91DigitalOcean, reserved2026-11-11
256 GBAMD Instinct MI325X$2.88DigitalOcean, reserved2026-11-11
268 GBB300 SXM6 268GB$6.025Verda, spot2026-11-11
288 GBb300 288Gi SXM6$0.23Akash2026-11-11

Questions

How much does a cloud GPU cost per hour?
It depends on the card, the tier and the provider, and rates change often, so read them live. Consumer cards on community hosts are the cheapest, data-center cards with large memory cost several times more, and the newest GPUs cost the most. rent.dev's price board at rent.dev/prices lists current hourly rates per GPU from published pricing. The board at rent.dev/prices, generated 2026-11-14, holds 448 priced GPU rows: the cheapest is the gtx1050 2Gi PCIe at $0.02 per GPU hour (Akash) and the dearest the GB300 at $18.00 per GPU hour (Oracle).
What are the best cloud GPUs for developer SSH boxes?
Pick a card with enough memory for the models you load and no more: a 24 GB card covers most single-model development, and a 48 GB or 80 GB card covers larger models and fine-tuning. Take on-demand rather than interruptible capacity, so the box is not reclaimed while you work, and keep code in a repository and data in object storage, so the GPU can be terminated between sessions. On rent.dev, rent gpu create --port 22/tcp exposes SSH on a pod whose image runs an SSH server. By memory on rent.dev's price board, generated 2026-11-14, cards with 2 GB start at $0.02 per GPU hour (gtx1050 2Gi PCIe, Akash) and cards with 288 GB at $0.23 per GPU hour (b300 288Gi SXM6, Akash).
How do I avoid hidden cloud GPU costs?
Most surprise bills come from things left running. Terminate GPUs when work stops rather than only stopping them, since a stopped instance can still bill for disks and addresses; delete volumes and snapshots you no longer need; keep large datasets in the same region as the GPU to avoid transfer fees; and check the minimum billing time before running many short jobs.
How do I shut down a cloud GPU automatically if a process crashes?
Run the job under a wrapper that terminates the machine when the process exits, whatever its exit code, and add an idle check that ends it when GPU utilisation in nvidia-smi stays at zero for a set time. Set a budget alert at the provider as a last line of defence. A rent.dev GPU pod has no timeout of its own, so the shutdown has to come from your script or your agent.