Multi-GPU nodes
8-GPU NVLink clusters, rented by the month
Eight H100, H200, B200 or B300 GPUs on one NVLink fabric, in one dedicated bare-metal node with 3.2 Tb/s of InfiniBand, up to 4 TB of RAM and 30 TB of NVMe. In stock, priced per node per month, and live in under 10 minutes — the same flat invoice as a single server, for the runs a single server cannot hold.
- 8× H100 SXM node$7,990per node, per month · $999 per GPU
- Eight H100s, hourly on-demand$19,568market median × 8 × 730 h
- You keepup to −59%across the four node configurations
Pricing
Cluster pricing per node, per month
Four configurations, each a whole machine. The node price is below eight single GPUs — and far below eight hourly meters running for a month.
| Node | GPU memory · host | Per node, per month | vs 8 single GPUs | vs on-demand | Order |
|---|---|---|---|---|---|
| 8× H100 SXMHopper · 14 left single GPUs | 640 GB HBM3 · NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe | $7,990 | −20% | −59% | Deploy |
| 8× H200Hopper · 9 left single GPUs | 1.1 TB HBM3e · NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe | $10,990 | −19% | −58% | Deploy |
| 8× B200Blackwell · 8 left single GPUs | 1.4 TB HBM3e · NVLink 5 · 3.2 Tb/s InfiniBand · 224 vCPU · 3 TB RAM · 30 TB NVMe | $18,900 | −12% | −55% | Deploy |
| 8× B300Blackwell Ultra · node sold out | 2.3 TB HBM3e · NVLink 5 · 3.2 Tb/s InfiniBand · 224 vCPU · 4 TB RAM · 30 TB NVMe | $27,900 | −8% | −39% | Ask for ETA |
Prices per node per month in USD, excl. VAT; launch rates fixed for your term for orders confirmed before Oct 31, 2026. “vs on-demand” compares eight GPUs at the market-median published hourly rate over 730 hours. Node availability follows the single-GPU counters, updated 18 September 2026. Terms of 3, 6 and 12 months take 5%, 10% and 15% off the node price.
Inside every node
One fabric, one host, nothing shared
A node is not eight servers with a switch between them. It is one HGX-class machine: the GPUs talk over NVLink, the host feeds them from local NVMe, and InfiniBand waits for the day you need a second node.
-
640 GB HBM3 GPU memory
NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe
$7,990per node / month
vs. $9,992 for 8 single GPUs · $19,568 hourly on-demand
Deploy this node -
1.1 TB HBM3e GPU memory
NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe
$10,990per node / month
vs. $13,520 for 8 single GPUs · $26,160 hourly on-demand
Deploy this node -
1.4 TB HBM3e GPU memory
NVLink 5 · 3.2 Tb/s InfiniBand · 224 vCPU · 3 TB RAM · 30 TB NVMe
$18,900per node / month
vs. $21,520 for 8 single GPUs · $41,760 hourly on-demand
Deploy this node -
2.3 TB HBM3e GPU memory
NVLink 5 · 3.2 Tb/s InfiniBand · 224 vCPU · 4 TB RAM · 30 TB NVMe
$27,900per node / month
vs. $30,320 for 8 single GPUs · $46,080 hourly on-demand
Node sold out · 5 single GPUs left
Workloads
What an 8-GPU node is for
Anything that does not fit one GPU — or that fits, but not this quarter.
-
Pre-training and long runs
Weeks of full-utilisation training with FSDP, DeepSpeed or Megatron over NVLink, checkpoints on 30 TB of local NVMe, and a flat price that does not move while the run does.
-
Full fine-tunes of large models
A 70B full fine-tune needs about 1.1 TB across the node: 8× H200 (1.1 TB), 8× B200 (1.4 TB) or 8× B300 (2.3 TB) hold it without optimizer offloading.
-
Serving 400B-class models
Tensor-parallel inference of the largest open models at FP8 or 16-bit, with the KV cache for long contexts spread across eight GPUs on one fabric.
Scaling past one node
Every node ships with 3.2 Tb/s of InfiniBand — eight 400 Gb/s ports — so two, four or more nodes join into one training fabric with NCCL, in the same facility. Multi-node clusters are placed by hand to keep them on the same switch: tell us the node count and the term, and you get a price and a go-live date before you pay.
Single nodes are self-serve. Pick the node on the deploy page, choose the image, settle the first invoice from your balance, and the node is yours in under 10 minutes.
Cluster rental: questions answered
How much does an 8× H100 cluster cost per month?
$7,990 a month for a dedicated 8× H100 SXM node — NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe — against $9,992 for eight single H100 servers here and roughly $19,568 for eight H100s at the market-median on-demand rate over 730 hours. H200, B200 and B300 nodes are priced on this page; every price includes the host, 30 TB of NVMe, the network and 24/7 support.
Is a node one machine or eight servers?
One machine: eight GPUs on a single NVLink fabric inside one dedicated bare-metal host, with its own CPUs, RAM and NVMe. Tensor and pipeline parallelism run over NVLink; the 3.2 Tb/s InfiniBand ports are for joining several nodes.
Can I connect several nodes?
Yes. Each node carries 3.2 Tb/s of InfiniBand for multi-node training with NCCL. For more than one node, open a sales ticket before ordering and we place them on the same fabric, in the same facility.
How quickly is a node live?
The same as a single server: once your payment confirms on-chain, the node is imaged with the OS and CUDA stack you chose, secured with your SSH key and handed over in under 10 minutes. Nodes are allocated from the units shown as in stock, updated by hand as orders are confirmed.
Which images and drivers come on a node?
Ubuntu 24.04 with the NVIDIA 570 driver, CUDA 12.6, cuDNN, NCCL, Docker and the NVIDIA Container Toolkit; or PyTorch 2.7, TensorFlow, JAX, the Hugging Face stack, vLLM, Triton or single-node Kubernetes (K3s) with the GPU Operator, chosen at checkout.
Eight GPUs, one invoice, ten minutes
Every node on this page is allocated from stock. Fund your balance, pick the node, and start the run tonight.