This is the hands-on companion to the location articles on this blog. It shows exactly how to choose GPU server country filters on the Cloudj deploy page, what each column in the results means, and what to do when the list comes back empty. If you want the reasoning behind choosing a country in the first place, start with GPU cloud by location; this page is about the buttons.
Before you start
- Sign in to the console and make sure your account balance can cover the deposit. Deploying holds about one hour of GPU, some hours of storage and a bandwidth reserve; the unused part returns after you destroy the instance.
- Add at least one SSH key to your account. GPU instances accept SSH keys only; password login is disabled.
- Know roughly how much VRAM your workload needs.
Step 1: open the deploy page and pick a model tile
Open the GPU deploy page. The page shows a tile for each GPU model. Click the one you want, for example RTX 4090. You can also open a model directly: the RTX 4090 deploy link preselects it. The RTX 4090 model page on the website shows its specs and the current "from" price.
Some models need a higher account level. If you pick one that is not available to you, the console says "This GPU class is not available at your account level".
Step 2: choose the GPU count
Select 1×, 2×, 4× or 8×. All GPUs in a result are in the same machine. Higher counts narrow the list, since fewer hosts have many GPUs of one model free at the same moment.
Step 3: fill in the Country field
The Country field takes two-letter ISO country codes.
- One country:
JP - Several countries:
JP, SGorJP SG; commas and spaces both work. - A wider set:
DE, NL, FR, PL, SE
Common mistakes:
| You typed | What happens | Use instead |
|---|---|---|
UK | Not the ISO code for the United Kingdom | GB |
Japan | Full names are not codes | JP |
USA | Three letters | US |
KO | Not Korea's code | KR |
Leave the field empty to see machines in every country. That is often the right choice for batch work.
Each machine shows its country only: no city, data centre or region. For Vietnam, you can try VN, but machines there are not guaranteed to be listed; SG and HK are the usual alternatives.
Step 4: decide on Datacenter only
The Datacenter only switch keeps only hosts located in data centres, which carry a Datacenter badge in the results. It narrows the list. Turn it on when a client or your own rules require it. Note that it describes where the host is, not whether the card is a datacenter model, and it does not add any SLA. Datacenter vs independent GPU hosts explains the trade-off.
Step 5: optional filters
Use these only when you need them; each one shortens the list.
- Minimum VRAM, useful when searching across several models.
- Maximum price per hour.
- Minimum reliability. See the GPU host reliability score for how to read it.
- Maximum bandwidth price per TB, download and upload.
- Minimum download speed in Mbps, useful when you will pull large models.
- Disk size used for prices, so the totals reflect the disk you plan to choose.
Step 6: read the results
The list shows up to 50 machines, sorted by total price, lowest first. For each one you see:
| Column | What it tells you |
|---|---|
| GPU and count | Model and how many in the machine |
| VRAM, vCPU, RAM | Resources available to your instance |
| Download speed | How quickly the host can pull models and data |
| Reliability | How dependable the host has been, as a percentage |
| Country | The host's country, the only location detail |
| Max duration | Days until the machine's end date; the instance and disk are removed then |
| Storage price | Per GB-month, charged while running and while stopped |
| Bandwidth price | Per TB down and up |
| Datacenter badge | Present when the host is in a data centre |
A practical reading order: country, then reliability, then max duration, then price. For a live service, also measure the network from your users before you commit; measuring latency to a GPU host explains how.
When the list is empty
Stock changes during the day, so empty results are common with narrow searches. Loosen one thing at a time:
| Likely cause | Try |
|---|---|
| Country too narrow | Add neighbouring codes, for example JP to JP, KR, TW, or DE to DE, NL, FR |
| Datacenter only is on | Turn it off and judge hosts by reliability and max duration |
| Minimum reliability too high | Lower it a little and keep backups current |
| GPU count too high | Try fewer GPUs, or several single-GPU machines |
| Price or bandwidth caps too tight | Relax them and compare totals |
| Model is scarce right now | Try a model with similar VRAM, or check again later |
| The job is batch work | Clear the Country field entirely |
Step 7: deploy
Pick a machine, then:
- Choose a template: Ubuntu 22.04 + CUDA (SSH), PyTorch + JupyterLab, or Custom image (a public Docker image; needs a higher account level).
- Choose disk size within the template minimum and the machine maximum. It cannot be resized later.
- Select your SSH key.
- Check the price shown; it is locked for this instance.
- Deploy. If the deployment fails to start, nothing is charged and the deposit is returned.
When it is running, the Connection tab shows the host, port and full SSH command.
FAQ
Can I filter by city?
No. GPU machines show country only. City choice exists for Cloudj VPS.
Why does UK return nothing?
The ISO code for the United Kingdom is GB.
Is the cheapest result the best one?
Not always. Check reliability and max duration first, then pick the cheapest machine that passes.
Does a country search guarantee the same country next time?
No. Each deployment is on whatever machine you pick from the live list at that moment.
Try it now
Open the RTX 4090 deploy page, type JP, SG or DE, US, GB in the Country field, and read the results with the table above. Live "from" prices for every model are on GPU cloud.