Rent NVIDIA H100 SXM servers from $2,190 per month

If you want to rent NVIDIA H100 GPUs on monthly terms, we offer the SXM version in stock: a single H100 in a dedicated virtual machine for $2,190 per month, or a bare-metal NVIDIA HGX H100 server with eight GPUs for $16,900 per month. Every GPU is a whole physical H100 assigned only to you, and the price stays flat whatever your utilization.

The H100 SXM in 2026: what you get

The H100 is NVIDIA's first Hopper GPU. It brought FP8 Tensor Cores and the Transformer Engine to NVIDIA's data-center line, and years of production use mean that PyTorch, vLLM, SGLang and TensorRT-LLM all support it well. NVIDIA lists 1,979 TFLOPS of dense FP8 and 989 TFLOPS of dense BF16 for the SXM version, with 80 GB of HBM3 at 3.35 TB/s. If you are coming from the previous generation, our page on how the H100 compares with the Ampere generation shows what changed.

The SXM detail matters when you compare offers. SXM is the module format used on NVIDIA HGX boards, with NVLink 4 at 900 GB/s per GPU and a 700 W power budget, so check that any price you compare refers to the same version. Against its sibling, the H200, the H100 has the same vendor-listed compute but less memory (80 GB instead of 141 GB) and less bandwidth (3.35 TB/s instead of 4.8 TB/s). That makes it the value pick for workloads that fit in 80 GB.

What fits in 80 GB

Use the usual rule: weights take roughly parameters × 2 bytes in BF16, 1 byte in FP8 and half a byte in 4-bit, plus KV cache and 10–30% headroom.

• Up to 32B for inference: an 8B model in BF16 takes about 16 GB and a 32B model in FP8 about 32 GB, both leaving tens of gigabytes for KV cache. A 32B model in BF16 (about 64 GB) fits on one GPU at moderate contexts; long contexts or many concurrent requests need FP8 or a second GPU.
• 70B for inference: about 70 GB in FP8, which fits one H100 only tightly, with little room for KV cache. For production, split it across two H100s with tensor parallelism, quantize to 4-bit (about 35 GB), or use one H200.
• Fine-tuning on one GPU: QLoRA on a 70B model (about 35 GB of 4-bit weights plus adapters and activations) and LoRA on 8B–14B models in BF16 fit on a single H100.
• Full fine-tuning on the HGX H100: Adam needs roughly 16–20 bytes per parameter before activations, so an 8B model needs about 128–160 GB, an easy fit across eight GPUs and 640 GB. A 70B model needs about 1.1–1.4 TB, beyond the server's HBM3: you would offload optimizer state to its 2 TB of DDR5, or move to an HGX B300.
• Very large models: a 405B-class model in FP8 (about 405 GB) fits on the HGX H100 with roughly 200 GB left for KV cache. A 671B mixture-of-experts model in FP8 does not fit in 640 GB at all.

Our guide to fine-tuning a 70B model compares these routes in more detail.

One H100 or an 8-GPU HGX H100 server?

The single H100 plan includes 24 vCPUs, 225 GB of RAM, 2 TB of NVMe and a 25 Gbps port. It is the right size for one endpoint, one fine-tuning job at a time, a course or a proof of concept.

The HGX H100 is a bare-metal server with eight H100 SXM GPUs in one NVLink domain at 900 GB/s per GPU, two Xeon 8480+ processors with 112 cores in total, 2 TB of DDR5, eight 3.84 TB NVMe drives and eight 400G ConnectX-7 adapters. Take it when a job needs several GPUs working on the same model: tensor parallelism for 70B serving, sharded training, or multi-GPU research code. Eight single H100s would cost $17,520 per month against $16,900 for the server, which also adds NVLink; for one to seven GPUs of independent work, single plans cost less.

Software stack on an H100 server

Each server comes with root access over SSH and an Ubuntu 22.04 or 24.04 image with NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit preinstalled; Debian or Rocky Linux are available on request. You can pull your usual containers and start. Two commands to confirm the GPU and keep an eye on it:

docker run --rm --gpus all pytorch/pytorch:latest \
  python -c "import torch; print(torch.cuda.get_device_name(0))"
nvidia-smi --query-gpu=memory.used,memory.total,utilization.gpu \
  --format=csv -l 5

The second command prints memory and utilization every five seconds, a quick way to see whether your job is limited by memory or leaves the GPU idle. Our GPU monitoring guide goes further.

H100 rental price per month and per GPU-hour

The single H100 costs $2,190 per month, exactly $3.00 per GPU-hour if it runs all 730 hours. The HGX H100 costs $16,900 per month, about $2.89 per GPU-hour. Because the price is flat, your real cost per used hour depends on utilization:

If you can buy H100 time by the hour for less than the figure in your row, hourly billing costs less; if not, monthly does. Put another way, the break-even point is $2,190 divided by the hourly rate you would pay. OverGPU's October 2026 survey of public price lists found the median on-demand H100 price per GPU-hour up roughly 13% over the past year, so fixing the price for each paid month also removes a budgeting risk. Our analysis of H100 rental costs in 2026 goes deeper.

The price includes 20 TB of outbound traffic per month on the single GPU, or 50 TB on the 8-GPU server, with no egress fees within that allowance, plus a public IPv4 address and an IPv6 /64, DDoS mitigation, 24/7 support from engineers and the 99.9% availability SLA. There is no setup fee.

How to rent NVIDIA H100 servers

1. You choose a single H100 or an HGX H100 on the deployment page, set up the OS image, optional preinstalled software and your login (root password or SSH key), and create a free account or log in in the same form.
2. You check the monthly price, shown before you pay.
3. You pay the first month from your prepaid balance, which you top up in BTC, ETH, XMR, USDT on TRON or SOL, with up to 50% bonus credit. Card payments are coming soon.
4. We deliver root SSH access within 10 minutes of payment, with the IP and SSH details in your dashboard.

The plan renews monthly from your balance with a reminder before each renewal, and you can cancel any month with no fee. For sizing, the VRAM calculator and the sections above show what fits; customers with an active server can ask for volume pricing by ticket from the dashboard.