NVIDIA H200 SXM cloud GPU rental

The NVIDIA H200 SXM is a datacenter Hopper GPU with 141 GB of HBM3e at 4.8 TB/s of memory bandwidth. On FastInfra it is a good fit for frontier-scale training, 200B+ inference, long-context serving. As a rule of thumb, one H200 SXM holds the weights of a model of about 56B parameters at 16-bit precision, or about 220B with 4-bit quantization, after leaving room for the KV cache.

Two on-demand tiers are available. Secure runs in vetted datacenters at $5.71 per GPU-hour; Value uses lower-cost community hosts at $3.88 per GPU-hour for the same GPU model, with more variable capacity.

Best for

Frontier-scale training, 200B+ inference, long-context serving.

About the Hopper architecture

Hopper introduced the Transformer Engine with FP8 precision and fourth-generation NVLink at up to 900 GB/s per GPU. It is the most widely deployed architecture for LLM training and serving, and vLLM, SGLang, TensorRT-LLM, FlashAttention 3 and PyTorch are all tuned for it, so most open-source training and inference code runs at full speed without changes.

Only need inference?

If you want to call an open model such as Llama, Qwen, DeepSeek or GPT-OSS rather than manage a server, the pay-per-token inference API is often cheaper than keeping a GPU busy around the clock. It is OpenAI-compatible and uses the same account and wallet as your pods. Browse every model with live prices.

How much does it cost to rent an NVIDIA H200 SXM per hour?

$5.71 per GPU-hour on the Secure tier and $3.88 on the Value tier. Billing is per minute while the pod exists, so 24/7 use for a 730-hour month costs about $2,831 per GPU at the lowest on-demand rate. Prices sync from live capacity and are shown before you deploy.

How much VRAM does the NVIDIA H200 SXM have?

141 GB of HBM3e at 4.8 TB/s of memory bandwidth. The chip is rated at up to 700 W. A pod with 8 GPUs gives 1,128 GB of combined VRAM for tensor- or pipeline-parallel workloads.

What size LLM can I run on an NVIDIA H200 SXM?

Roughly 56B parameters at FP16 or BF16, 110B at 8-bit and 220B at 4-bit on a single GPU, reserving about 20% of VRAM for the KV cache and activations. Mixture-of-experts models need memory for all experts, not only the active ones. Long contexts and large batches need more headroom.

Is the NVIDIA H200 SXM billed per hour or per minute?

Per minute. The hourly rate is divided by 60 and charged for every minute the pod exists. Attached volumes bill per GB-month and continue while a pod is stopped; delete the pod to stop all charges. There is no minimum term or reservation fee.

How do I rent an NVIDIA H200 SXM on FastInfra?

Create an account, add prepaid credit, open the GPU configurator and choose the H200 SXM, the tier, the GPU count, a container template and storage. The configurator shows the hourly and monthly price, then deploys a pod you reach over SSH or JupyterLab within minutes.

Can I rent multiple H200 SXM GPUs in one pod?

Yes, up to 8 per pod. The price scales linearly: 8 GPUs cost $45.71 per hour on Secure.

Is there a cheaper GPU than the H200 SXM with the same VRAM?

Yes. The AMD Instinct MI300X has 192 GB and starts at $2.58 per hour. It differs in architecture (CDNA 3), memory bandwidth and multi-GPU interconnect, so compare throughput for your workload before switching.