# H200 cloud rental: where H200 GPUs are listed and how to rent URL: https://cloudj.net/blog/h200-cloud-rental Published: 2026-11-12 · Updated: 2026-11-12 > A practical H200 cloud rental guide: what H200 NVL memory buys you, searching any country then narrowing, account level, and disk sizing for big weights. An H200 cloud rental is usually a decision about memory. The H200 NVL carries a 141 GB class of HBM, enough to hold models or contexts that do not fit on an 80 GB card. On Cloudj, H200 machines appear in the same live list of independent hosts as every other GPU, so renting one is a matter of searching well, being ready on the account side, and planning disk and data before the meter starts. This guide covers all three. It makes no promises about stock or price, because both change during the day. ## What the H200 NVL is for The [H200 NVL](https://cloudj.net/gpu-cloud/h200-nvl) is a datacenter GPU with a 141 GB class of memory per card. In practice that memory helps when: - **A large model fits on fewer GPUs.** A 70B model at 16-bit precision needs roughly 140 GB for weights alone, so it is tight on one card but comfortable on two; at 8-bit it fits on one with room for cache. These are rough figures; check your model's actual size. - **Long contexts need a big KV cache.** Serving many concurrent users with long prompts eats memory quickly. - **Training or fine-tuning needs headroom.** Larger batches and full fine-tunes of mid-sized models fit more easily. If your model fits comfortably in 80 GB, the [H100 SXM](https://cloudj.net/gpu-cloud/h100-sxm) may be the more practical search. [A100 vs H100 in the cloud](https://cloudj.net/blog/a100-vs-h100-cloud) compares the older generations. ## Before you search: account readiness The H200 is a datacenter GPU. What an account can rent depends on its account level, set by Cloudj. New accounts start at a level for consumer GPUs; datacenter models such as A100, H100 and H200 need a higher level. If you select the H200 and see "This GPU class is not available at your account level", contact Cloudj support from the console before you plan around it. Also check that your prepaid balance covers the deposit. Deploying holds about an hour of GPU, some hours of storage and a bandwidth reserve; the unused part returns after you destroy the instance, though it can stay held up to 72 hours while final bandwidth charges settle. ## Search any country first, then narrow For a scarce card, starting with a single country is a quick way to see nothing. Reverse the usual order: 1. Open the [deploy page with the H200 NVL preselected](https://app.cloudj.net/deploy/gpu?model=h200-nvl) and leave the Country field empty. 2. Choose the GPU count you need (1×, 2×, 4× or 8×; all GPUs in one machine). 3. Look at what is listed: country, reliability, max duration, download speed, bandwidth price and the Datacenter badge. 4. Now narrow if you need to. For US users, try `US`. For a Europe-only rule, try `DE, NL` and other EU codes you accept. 5. If the narrowed search is empty, decide whether location really matters for this job. For most H200 work, which is training, fine-tuning or batch inference, it does not. The [GPU cloud by location guide](https://cloudj.net/blog/gpu-cloud-by-location-guide) explains when location matters and when it does not. ### Datacenter only Datacenter cards are usually, though not always, in data centres. Turning on "Datacenter only" filters to hosts carrying the Datacenter badge. How much it narrows an H200 search depends on what is listed at the moment; compare the list with the switch on and off. It does not add an SLA: every host is independent, with no hardware SLA and an end date. ## Plan disk before you deploy Disk size is set at deploy and **cannot be resized later**. With large models this is where people get caught. Add up: | Item | Rough guide | |---|---| | Model weights | Size of the checkpoint you download, in the precision you will load | | Second copy during conversion or quantization | Same order as the weights, if you convert on the machine | | Framework cache (pip, Hugging Face, compiled kernels) | Tens of GB is common | | Datasets and outputs | Your own numbers | | Headroom | Leave a comfortable margin | Storage is charged per GB-month, so oversizing has a cost, but running out halfway through a download is worse. [Choosing GPU disk size](https://cloudj.net/blog/choose-gpu-disk-size) walks through this in detail. Two related checks: - **Download speed.** The results show each machine's download speed in Mbps. Pulling a very large checkpoint over a slow link wastes paid GPU time. - **Bandwidth price.** Each machine lists a price per TB down and up. Large uploads of datasets or downloads of checkpoints show up here. ## Multi-GPU H200 machines If a 2× or 4× H200 machine is listed, all GPUs are in the same host. Frameworks such as vLLM can split a model across them with tensor parallelism: ```bash vllm serve MODEL_ID --tensor-parallel-size 2 ``` You install vLLM yourself over SSH; Cloudj does not ship it as a ready-made template. [Running a 70B LLM on multiple GPUs in Europe](https://cloudj.net/blog/llm-70b-multi-gpu-europe) covers multi-GPU serving in more depth. ## Protect the job An H200 hour is valuable, so plan for interruptions: - Note the "available until" date; the instance and disk are removed when it arrives. - Save checkpoints to storage you control at regular intervals. - Run long jobs inside `tmux` so an SSH drop does not kill them. - If the host goes offline, billing pauses, but there is no automatic move to another machine. Have a script that rebuilds the environment. Stopping an instance keeps the disk but releases the GPU; someone else may rent it, and restarting waits until it is free. For a scarce card, destroying and redeploying later may be more practical than stopping. ## FAQ ### Is the H200 always in stock on Cloudj? No. Stock changes during the day. The model page shows whether it is currently in stock, and the console shows what is listed. ### Why can't I select the H200? Datacenter GPUs need a higher account level. Contact Cloudj support from the console. ### Can I resize the disk after I download the weights? No. Disk size is fixed at deploy, so plan it before. ### Should I search my own country first? For a scarce card, search all countries first and narrow only if location matters for the job. ## Next step Open the [H200 NVL deploy page](https://app.cloudj.net/deploy/gpu?model=h200-nvl) with the Country field empty and see what is listed now. Live "from" prices for every model are on [GPU cloud](https://cloudj.net/gpu-cloud).