GPU cluster rental: multi-node builds on 8-GPU servers

GPU cluster rental at FullGPU means several of our bare-metal 8-GPU servers, joined by a high-speed fabric and a private network, and rented together. We design each multi-node GPU cluster around the job it has to run, confirm the fabric and every other detail in a written quote, and hand you root access on every node.

How GPU cluster rental works

A cluster starts from the 8-GPU servers you can also rent one at a time. Inside each node, the eight GPUs talk over NVLink on HGX systems or Infinity Fabric on MI355X. Between nodes, traffic runs over the fabric, attached to each node's adapters: eight 800G ConnectX-8 on HGX B300, eight 400G ConnectX-7 on HGX B200, H200 and H100, and eight 400G RoCE on MI355X.

The fabric is designed per order. We choose InfiniBand or RoCE for your build and state it in the quote before you pay. Alongside it, a private network links your nodes for data movement, orchestration and services that should not face the internet. Each node keeps its own 50 TB of outbound bandwidth per month under fair use, its public IPv4 address and IPv6 /64, and DDoS mitigation.

Choosing nodes for a multi-node GPU cluster

As a sizing example, four HGX H200 nodes give you 32 GPUs and 4,512 GB of HBM3e for $95,600 a month at list price, before any term or volume pricing. Full fine-tuning of a 70B model with Adam needs roughly 1.1 to 1.4 TB before activations, so two H200 nodes already hold it with sharding, and more nodes shorten the run.

Keep one GPU type per job, since the slowest node sets the pace of synchronous training. MI355X nodes run ROCm, and their collectives go through RCCL instead of NCCL. The 8× RTX PRO 6000 server, with two 100 GbE ports and no NVLink, is better suited to independent jobs than to a tightly coupled cluster. If your parallelism has to span more than eight GPUs on NVLink, compare with our rack-scale systems.

InfiniBand or RoCE

Both carry RDMA traffic, which NCCL and RCCL use for collective operations between nodes. InfiniBand is a dedicated fabric with its own switches and is lossless by design. RoCE runs RDMA over Ethernet and relies on careful network configuration to behave the same way. We pick the one that fits your nodes and your build, and explain the choice in the quote.

For your jobs, the difference mostly shows up in configuration: environment variables such as NCCL_IB_HCA and NCCL_SOCKET_IFNAME, and the bus bandwidth to expect in tests. We tell you which fabric your cluster uses and how to point NCCL at it. The multi-GPU and NCCL guide shows how to run all-reduce tests across nodes.

What you manage and what we handle

We deliver and maintain the hardware, the fabric and the private network. Each server carries the 99.9% monthly availability SLA for power, network and hardware, with service credits. You run everything from the operating system up: the scheduler, such as Slurm, your containers, your training code and your checkpoints. We do not operate a managed Kubernetes or Slurm service.

Storage is local NVMe on each node, about 30.7 TB raw on HGX and MI355X servers, and it is not backed up by us unless agreed in writing. Plan how data reaches every node and where checkpoints are copied. Our engineers help at no extra fee, from sizing before you order to multi-node NCCL tests after handover; see expert onboarding.

Pricing and the quote process

1. Tell us through the contact form what the cluster has to run: model size, GPU count, start date, term and any fabric preference.
2. Our engineers propose the nodes, the fabric and the private network.
3. You receive a written quote with the configuration, price, location and delivery date. Term pricing for 6 to 12 months and volume pricing, usually from three servers, are part of that quote when they apply.
4. Large orders may require verification. We tell you exactly what is needed and why.
5. You pay by card or crypto and receive credentials by email. If we miss the confirmed delivery date, you can cancel for a full refund.

GPU capacity is tight in 2026, and many providers prefer one-year contracts. Sharing your start date early gives us the best chance to reserve the nodes you want.