Single GPU server rental: a whole GPU in a dedicated VM

Single GPU server rental at FullGPU means one whole physical GPU, passed through to a virtual machine that only you use. The vCPU, RAM and local NVMe that come with it are dedicated to that VM, and you log in as root over SSH. You pay one flat price per month, from $449 for a GeForce RTX 5090 to $5,190 for an NVIDIA B300 with 288 GB of HBM3e.

How single GPU server rental works

Each plan is a dedicated GPU VM. The hypervisor passes the entire GPU through to your virtual machine, so the guest operating system loads the NVIDIA driver and drives the card directly. Run nvidia-smi inside the VM and you see the full memory of the GPU. No other tenant has a slice of it: there is no vGPU, no MIG partition and no time-sharing scheduler between your code and the hardware.

Around the GPU, every single-GPU server comes with:

• A public IPv4 address and an IPv6 /64, with DDoS mitigation on the public IPs.
• SSH-key access and Ubuntu images with the GPU stack preinstalled, described on our OS images page.
• A 99.9% monthly availability SLA for power, network and hardware, with service credits if we fall short.
• Local NVMe storage that is securely erased between customers.

The seven single-GPU plans compared

Host resources grow with the plan: the higher up the table, the more vCPU, RAM, NVMe and network come with the GPU.

A month is about 730 hours. Used around the clock, the H100 works out to about $3.00 per GPU-hour and the RTX 5090 to about $0.62. If you need a GPU for a few hours a week, hourly billing usually fits better. Our monthly vs hourly comparison walks through the trade-off.

When one GPU is enough

Memory usually decides. As a rough rule, model weights take the parameter count times the bytes per parameter (2 in BF16, 1 in FP8, 0.5 in FP4). Add the KV cache and runtime overhead, then plan 10 to 30% headroom on top.

• Small models and image pipelines. An 8B model needs about 16 GB in BF16, which leaves room for the KV cache on a 32 GB RTX 5090 or a 48 GB L40S.
• 70B-class inference. About 70 GB in FP8: tight on one H100, comfortable on an H200. In BF16 it is about 140 GB, so an H200 is tight and a B200 or B300 is comfortable.
• Fine-tuning. LoRA and QLoRA need far less memory than full fine-tuning. Full fine-tuning with Adam takes roughly 16 to 20 bytes per parameter before activations, so even a 7B model needs about 112 to 140 GB. That points to a B200 or a B300.
• Rendering and generative media. Large scenes, video models and big batches benefit from the 96 GB of the RTX PRO 6000.

To run these numbers for your own model, precision and context length, use the VRAM calculator.

Limits of a dedicated GPU VM

A single-GPU plan is one GPU in one VM. Two plans are two separate machines that talk over the network, not over NVLink. When one model has to be split across GPUs, for example a model larger than the 288 GB of a B300 or a training job that needs tensor parallelism, choose an 8-GPU server instead. It is bare metal, with NVLink between the GPUs on HGX systems and 50 TB of monthly bandwidth instead of 20 TB.

The 8-GPU servers also cost slightly less per GPU. Eight single H100 plans add up to $17,520 a month, while an HGX H100 server is $16,900. Bare metal also gives you the whole machine: both CPUs, all system memory and every drive.

Upgrades and ordering

You are not locked into a size. The minimum term is one month, plans renew automatically with an email reminder before each renewal, and you can cancel any month with no fee. The server then runs until the end of the paid month.

To move up, say from an H100 to an H200 when your contexts grow, ask through the contact form. Changes depend on stock, and we confirm the new configuration, price and delivery date in writing before you pay. Our engineers can advise on moving your data as part of expert onboarding, and the upgrades and changes guide covers the details.

A first order follows the same path: request a server on the site, receive a written confirmation of the configuration, price, location and delivery date, pay by card or crypto, and receive your credentials by email. If we miss the confirmed delivery date, you can cancel for a full refund.