An H100 cloud Asia search is worth doing properly before you give up and rent one on another continent. Teams in Japan, Korea, Singapore, Taiwan, India and Vietnam often want the H100 close to home for data, latency or simply convenience. On Cloudj, location is a country filter over a live list of independent host machines, so the honest answer to "is there an H100 in Asia?" is: search now and see. This article gives you a search order, explains why datacenter cards can be harder to find in some countries, and shows a split that often works better than waiting: train wherever an H100 is listed, serve from Asia.
The GPU cloud in Asia guide covers the Asian codes in general; this one is only about the H100.
Which H100 to look for
Two H100 variants matter here:
- H100 SXM, 80 GB. The usual choice for training and high-throughput serving, especially in multi-GPU machines.
- H100 NVL, 94 GB. A PCIe-based variant with a little more memory per card, often used for large-model inference.
Both are datacenter GPUs, so they need a higher account level than a new account has. If you try to deploy one too early, the console says "This GPU class is not available at your account level"; contact Cloudj support from the console if you need it. Do this before a deadline, not on the day of the run.
A search order for Asian teams
Open the deploy page with the H100 SXM and work down this list. Stop at the first step that gives you a machine you are happy with.
- Your own country code. For example
JP,KR,SG,TWorIN. - All five together. Enter
JP, KR, SG, TW, IN. Each row still shows its country. - Switch the variant. Repeat the search with the H100 NVL selected.
- Relax the filters. Lower minimum reliability slightly, raise the maximum price per hour, or choose a GPU count you had not planned (1x instead of 8x, or the reverse).
- Add nearby codes.
HKandAUare reasonable additions for many Asian teams. - Check at another time of day. Stock changes during the day, and datacenter machines can be rented for long blocks and then return.
If you are in Vietnam, you can include VN too, but machines in Vietnam are not guaranteed to be listed. SG and HK are the usual nearby codes.
Why datacenter GPUs may be scarcer in some countries
Every listing is an independent host. Datacenter cards such as the H100 sit in fewer machines than consumer cards, and in some countries they may be rented out for weeks at a time. That means an Asian search can come back empty even when the same search in another region shows several rows. We do not publish which countries have H100s, because the answer changes; we only suggest that you search more than once and plan for an alternative.
That alternative is usually not "give up on Asia". It is "split the work".
The split: train elsewhere, serve from Asia
Most H100 work falls into two parts with different needs.
| Part | Needs an H100? | Needs to be near users? |
|---|---|---|
| Training or fine-tuning | Often yes | No |
| Evaluation sweeps, batch inference | Sometimes | No |
| Live inference for Asian users | Depends on the model | Yes |
So the pattern is: run training wherever an H100 is listed, regardless of country, then move the resulting weights to a GPU in Asia for serving. The serving GPU can be a different model entirely if a quantized version of your model fits. The H100 cloud in the USA guide covers the training side on US hosts.
Moving the weights is a plain file transfer. From the training instance, push the checkpoint to storage you control, or straight to the serving instance:
rsync -avP -e "ssh -p <port>" ./checkpoints/final/ root@<serving-host>:/workspace/model/Use the host and port from each instance's Connection tab. Bandwidth is charged per GB at each machine's listed price, so compress where you can and avoid moving intermediate checkpoints you do not need. The guide to moving models and data between GPU regions covers this step in more depth, including resuming interrupted transfers.
Serving from Asia without an H100
Once training is done, ask whether serving really needs an H100. Many models, once quantized, serve comfortably on 48 GB cards or even 24 GB consumer cards. An A100 listed in an Asian country can also be a strong serving option for models that need more memory than a consumer card.
Put a stable Singapore VPS in front for the domain, TLS, authentication and queue, and point it at whichever GPU is serving. When you switch GPUs, the public endpoint stays the same.
What to check before a long H100 run
- Max duration and "available until". When that date comes, the instance and its disk are removed. Make sure the run, plus a margin, fits.
- Disk size. It cannot be resized after deploy. Plan for datasets, checkpoints and logs.
- Checkpoints off the machine. Machines are independent hosts with no hardware SLA, and there is no automatic move if a host goes offline. Copy checkpoints out regularly.
- Deposit. Deploying holds a deposit from your balance; the unused part returns after you destroy the instance, possibly up to 72 hours later while bandwidth charges settle.
FAQ
Is there an H100 in Japan right now?
Search with JP and the H100 selected to find out. If a machine is listed you can deploy it; availability changes during the day.
Why can't I select the H100?
Datacenter GPUs need a higher account level. The console shows a message; contact Cloudj support from there.
Is it fine to train in the US and serve in Asia?
Yes, for most teams. Training is not latency-sensitive. Plan the transfer of weights and keep the serving GPU near users.
Where do I see H100 prices?
Live on the H100 SXM page, the H100 NVL page and GPU cloud.
Next step
Open the deploy page with the H100 SXM, try JP, KR, SG, TW, IN, then the NVL variant. If nothing fits today, train where an H100 is listed, move the weights, and serve from Asia behind a Singapore VPS.