If you want to rent RTX 4090 USA capacity for training, fine-tuning or rendering, the useful questions are practical ones. Is a machine listed right now? Should you take one card or a 2x or 4x machine? And how do you avoid paying for hours you did not use? This guide answers those for Cloudj, where GPUs come from a live inventory of independent host machines and you find US machines by filtering the list by country.
For everything else about renting in the United States, including other models, see the GPU cloud USA guide.
Why a 4090 in the US for batch-style work
The RTX 4090 is a 24 GB Ada Lovelace consumer card. It is a popular choice for work that fits in 24 GB per GPU: LoRA and QLoRA fine-tuning of small and mid-size models, Stable Diffusion and SDXL training, computer-vision training, and GPU rendering with tools like Blender in headless mode.
For batch-style jobs, distance to users barely matters. You upload data once, run for hours, and download results. A US host works well for teams in the Americas, and for anyone whose data or storage already sits in the US. If your users are elsewhere and the job is interactive, search near them instead.
How to rent RTX 4090 USA listings
- Open the deploy page with the model preselected: deploy an RTX 4090.
- Enter
USin the Country field. - Choose the GPU count: 1x, 2x, 4x or 8x. All GPUs in a listing sit in the same machine.
- Set your disk size. It is fixed at deploy and cannot be resized later.
- Read the results, sorted by total price, lowest first.
If a machine is listed, you can deploy it. If not, try a different GPU count, loosen the reliability or price filters, or add CA for Canada. Stock changes during the day, so searching again later often helps. Each listing shows the country only, never the state or city.
One card or several?
Multi-GPU 4090 machines appear when hosts list them. Whether you need one depends on how your code uses GPUs.
| Setup | Good for | Watch out for |
|---|---|---|
| 1x 4090 | Most LoRA runs, SDXL training, single-GPU inference tests, rendering one scene | Model plus optimizer must fit in 24 GB |
| 2x 4090 | Data-parallel training to halve wall-clock time, two parallel render jobs | 4090s have no NVLink; GPUs talk over PCIe, so model-parallel scaling is limited |
| 4x 4090 | Hyperparameter sweeps (one run per GPU), larger data-parallel jobs, render farms for animation frames | Check vCPU, RAM and disk keep up with four GPUs feeding data |
A useful rule: if your job is many independent pieces (frames, sweep runs, dataset shards), multiple 4090s scale well. If it is one model too large for 24 GB, a 48 GB or 80 GB card may be simpler than splitting across consumer GPUs. The single vs multi-GPU guide goes deeper into that choice.
Check that all GPUs are visible after you connect:
nvidia-smi --query-gpu=index,name,memory.total --format=csvPer-minute billing in practice
Cloudj bills GPU instances per minute from your prepaid balance while they run. The running price is GPU plus disk. The price shown when you deploy is locked for that instance, and live "from" prices are on the RTX 4090 page.
Per-minute billing changes how you plan:
- Test short. Deploy, check that your code runs and the GPU is used, then decide. A failed test costs minutes, not a day.
- Prepare before you deploy. Have your dataset, scripts and environment ready so the GPU is not idle while you set up.
- Watch utilisation. The instance page shows GPU utilisation and the last lines of the container logs. A GPU sitting at low utilisation is money spent on waiting, usually for data loading.
- Mind bandwidth. Downloads and uploads are charged per GB at the machine's listed price, which matters for render outputs and large checkpoints.
When you deploy, a deposit is held from your balance and the unused part returns after you destroy the instance.
Stop or destroy?
This is where most avoidable cost comes from.
- Stop keeps the disk and charges storage only. It also releases the GPU, so someone else may rent it. When you start again, you may see "Waiting for GPU" until that card is free.
- Destroy deletes the instance and its disk permanently. All charges stop.
| Situation | Choose |
|---|---|
| Job finished, outputs copied off | Destroy |
| Overnight pause, same environment needed tomorrow, can wait for the GPU | Stop |
| Need a GPU right now and the stopped one is "Waiting for GPU" | Destroy and deploy a new instance |
| Machine's "available until" date is near | Copy data off, then destroy |
Copy results to your own machine or a VPS before destroying:
rsync -avP -e "ssh -p <port>" root@<host>:/workspace/outputs/ ./outputs/The stop vs destroy guide covers edge cases in detail.
A US VPS alongside the GPU
For pipelines that run every day, keep the orchestration on a small US VPS: a queue of jobs, the scripts that deploy and connect to GPUs, and a place to receive outputs. The GPU instance then becomes disposable. If a host goes offline, billing pauses but the instance is not moved, so having your state on the VPS means you simply start again on another machine. The same pattern is used for agents in the AI agent VPS in the United States article.
FAQ
Are RTX 4090s always available in the US?
No. Search with US and the 4090 preselected; if nothing is listed, change the GPU count, loosen filters or check back later.
Can I rent four 4090s in one machine?
When a host lists a 4x machine, yes. The GPU count selector offers 1x, 2x, 4x and 8x.
Is there a monthly plan?
No. GPU instances are billed per minute from your balance.
Does stopping save my GPU for me?
No. Stopping keeps the disk but releases the GPU, so it may be rented by someone else before you start again.
Find a US RTX 4090
Open the RTX 4090 deploy page, enter US and see what is listed now. You can compare it with other models on GPU cloud.