If you want to rent RTX 4090 Japan capacity, you probably have users or collaborators in Japan and a workload that fits on a single 24 GB consumer card: image generation, a quantized language model, or a small fine-tune. This article shows how to search for that exact combination on Cloudj, what a 4090 can and cannot hold, and what to do when Japan does not list one at the moment you look.
For Japan in general, including other models and the Tokyo VPS, see the GPU cloud Japan guide.
Searching for an RTX 4090 in Japan
Cloudj GPUs come from a live list of host machines in many countries, and you choose location with a country filter. The quickest way in is the console link with the model preselected: open the RTX 4090 deploy page and sign in.
Then:
- Set the Country field to
JP. - Keep the GPU count at 1× unless your job splits across cards.
- Optionally set a minimum reliability if you plan a long job.
- Read the results. Each machine shows its country (never a city), VRAM, vCPU, RAM, download speed, reliability, max duration, bandwidth price and a Datacenter badge where relevant.
If a machine is listed, you can deploy it, and the price shown at deploy is locked for that instance. If nothing is listed, that is normal for a single model in a single country; stock changes during the day, so check the console again later or use the fallbacks below. The live "from" price for the card is always on the RTX 4090 page.
What 24 GB of VRAM fits
The RTX 4090 has 24 GB of VRAM. As general guidance (exact needs depend on resolution, batch size, context length and the software you use):
| Workload | Fits on one RTX 4090? |
|---|---|
| SDXL and similar image models, including LoRAs and ControlNet | Yes, comfortably at common resolutions |
| Newer, larger image models | Often, sometimes with lower precision or offloading |
| 7B to 13B LLMs in 8-bit or 4-bit | Yes, with room for context |
| 30B to 34B LLMs in 4-bit | Usually, but tight; long contexts may not fit |
| 70B LLMs | No, not without heavy offloading; use more VRAM |
| LoRA fine-tuning of 7B to 13B models | Yes, with 4-bit or 8-bit base weights |
If you are setting up Stable Diffusion or ComfyUI for a Japanese audience, the Stable Diffusion GPU in Japan article covers that workflow end to end.
When no RTX 4090 is listed in Japan
You have two directions: keep the country and change the card, or keep the card and change the country. Which one is right depends on whether location or the specific GPU matters more to you.
Keep Japan, change the card. Clear the model selection, set the Country to JP and set a minimum VRAM of 24 GB. You will see whatever fits:
- An RTX 5090, with 32 GB, gives you more headroom for larger quantized models and video workloads. It is a Blackwell card, so check that your image's CUDA and PyTorch versions support it. The RTX 4090 vs RTX 5090 cloud comparison covers when the extra memory is worth it.
- An RTX 3090 also has 24 GB. It is an older generation and slower, but the same models fit, which makes it a reasonable stand-in for inference and image work.
Keep the RTX 4090, change the country. Add nearby codes:
JP, KR, TWKorea and Taiwan are close to Japan, so for many interactive uses a 4090 there is a workable substitute. Test from a Japanese connection before you put users on it.
For batch work, such as overnight image generation or a LoRA training run, location barely matters. Remove the country filter and pick on price, reliability and max duration.
After you deploy
Access is by SSH key only. The instance's Connection tab shows the host, the port (not always 22) and the full command:
ssh -p <port> root@<host>
nvidia-sminvidia-smi should show one RTX 4090 with 24 GB. If you are running a web UI such as ComfyUI, do not expose it publicly; forward the port over SSH instead:
ssh -p <port> -L 8188:localhost:8188 root@<host>Then open http://localhost:8188 in your browser. Software like ComfyUI, Stable Diffusion WebUI or Ollama is installed by you; the console provides an Ubuntu + CUDA template and a PyTorch + JupyterLab template.
Pairing with a Tokyo VPS
If you are serving users rather than working alone, keep the public part of the service on a stable cloud server. GPU machines are independent hosts with no hardware SLA, and each one has an "available until" date when the instance and disk are removed. A VPS in Tokyo can run your domain, TLS, API and queue, and forward model requests to whichever GPU is current. The best VPS in Japan article covers that server.
A few cautions
- Disk size is fixed at deploy. Image models and checkpoints add up quickly, so choose generously.
- Stop releases the GPU. Stopping keeps your disk but someone else may rent the card; starting again waits until it is free. If you need a GPU now, destroy and deploy a new one.
- Licences. Check the licence of each model and LoRA you use, especially for commercial work in Japan.
FAQ
Is an RTX 4090 always available in Japan?
No. Stock changes during the day. Set the Country to JP with the 4090 selected; if a machine is listed, you can deploy it.
Can I choose a Tokyo GPU?
No. GPU location is country level only. Tokyo applies to the Cloudj VPS.
Will a 34B model fit on one RTX 4090?
In 4-bit it usually fits, but long contexts can push it over. If you need more room, look at an RTX 5090 or a 48 GB card.
Is an RTX 3090 in Japan a good substitute?
For inference and image generation, often yes: it has the same 24 GB. It is slower, so jobs take longer.
Find your RTX 4090
Open the RTX 4090 deploy page, set the Country to JP and see what is listed now. If nothing fits, try JP, KR, TW or a 24 GB minimum in Japan. All live prices are on GPU cloud.