Guide4 sections2 min read
The short answer
GPU rental means paying for a cloud GPU by the hour instead of buying the hardware. You choose a GPU model and configuration, launch an instance, connect over SSH or an API, and pay for the hours used.
On EverXCloud, GPU hours trade on a spot exchange: you rent at the best live offer, or set the price you are willing to pay and are matched when an offer meets it.
Choose the GPU from the workload
Start with memory: the model, data and batch must fit on the GPU. Then compare throughput and bandwidth, and finally cost per completed job. The table below maps GPU families to the workloads they are usually rented for.
How hourly prices are set
Prices are per GPU-hour, excluding tax, and move with supply and demand on the exchange. Region, GPU count, storage and network change the price of a configuration. Accounts are prepaid: rentals draw on your balance by the hour.
During pre-launch, prices shown on EverXCloud are indicative and become rentable on launch day.
Rent a GPU in four steps
Create an account and add an SSH key. Fund your balance in the currency of the offer. Pick an offer or place a bid, then launch the instance and connect when it is ready.
Release the instance when the job ends, and review persistent storage separately.
GPU families and the workloads they are usually rented for
| GPU family | Typical workloads | Example |
|---|---|---|
| Data center, latest generation (B200, B300, H200, H100) | Training and serving large language models, multi-GPU jobs | NVIDIA H200 141GB |
| Data center, inference and graphics (L40S, L4, A10, A30) | Inference, image and video generation, smaller models | NVIDIA L40S 48GB |
| Data center, previous generation (A100, V100) | Fine-tuning, mid-size training, scientific computing | NVIDIA A100 SXM 80GB |
| Workstation (RTX PRO 6000, RTX 6000 Ada) | 3D rendering, visualization, large-memory inference | NVIDIA RTX PRO 6000 96GB |
| GeForce (RTX 5090, RTX 4090, RTX 3090) | Image generation, LoRA fine-tuning, development and tests | NVIDIA GeForce RTX 4090 24GB |
| AMD Instinct (MI300X, MI325X) | Large-memory inference with ROCm | AMD Instinct MI300X 192GB |
Frequently asked questions
- How does GPU rental work?
- You rent a cloud GPU by the hour: choose a configuration, launch an instance, connect over SSH or the API, and pay for the hours used from a prepaid balance.
- Which GPU should I rent?
- The smallest GPU whose memory fits your model and data, with the best cost per completed job. Large language models usually need data center GPUs; image generation and small models often run on 24 to 48 GB cards.
- Can I choose my own price?
- Yes. On EverXCloud you can place a bid at the price you are willing to pay per GPU-hour; it is matched when an offer meets it.
- Are the prices shown final?
- No. During pre-launch, EverXCloud prices are indicative. They are shown per GPU-hour, excluding tax, and become rentable on launch day.