GPU Servers Overview

GPU servers are single-tenant bare-metal machines fitted with NVIDIA accelerators, sized for AI training, inference, and rendering. This guide covers the four tiers, how ordering and delivery actually work, what the setup fee covers, and how to manage a GPU machine once it is live.

What a GPU Server Is

A GPU server is a dedicated physical machine with one or more NVIDIA GPUs, plus the CPU, system memory, and fast storage needed to keep them fed. The whole machine is yours: nobody else is sharing the accelerators, the PCIe bandwidth, or the disks.

You install your own stack and run it however you like. CUDA workloads, PyTorch, TensorFlow, JAX, and the usual serving frameworks all run the way they would on any bare-metal Linux box, because that is what it is.

The reason to want single tenancy here is not only performance. GPU workloads tend to run flat out for hours, and shared accelerators behave very differently under that kind of sustained load than they do in a benchmark.

The Four Tiers

Spark is the entry point: enough GPU memory for smaller models, fine-tuning, inference, and rendering work. Forge, Foundry, and Reactor share the same accelerator and step up the system memory around it, which is what matters when you are staging large data sets or holding several models in host memory.

Every tier is sold both unmanaged and managed. Unmanaged hands you root and leaves the operating system to you. Managed keeps the same hardware with our engineers on the software layer. See Managed vs Unmanaged Cloud Servers.

For live pricing in your own currency, open Store in KPanel, then the GPU Compute card.

Choosing a Tier

Start from the model, not the price.

• How much GPU memory does your model need in the shape you are running it? That is the hard constraint. If weights plus activations do not fit, no amount of system RAM helps.
• Are you training or serving? Training is memory-hungry and long-running. Inference is often bounded by how fast you can get data to the GPU, which is where system RAM and disk matter more.
• How large is your working set? If you are streaming a data set that does not fit in host memory, the Foundry and Reactor configurations exist for exactly that reason.

If you are unsure, tell us what you are running, the model size, the framework, and how much GPU memory you think you need. We will point you at the right tier rather than let you overbuy. GPU and AI Compute Servers has more on the workload side.

How Ordering Actually Works

GPU hardware is scarce and expensive, so the process is deliberately not one-click.

1. Sign in to KPanel.
2. Click Store in the left sidebar.
3. Under Servers, click the GPU Compute card.
4. Switch between Unmanaged and Managed, and pick a tier.
5. Complete checkout.

After checkout you receive an order acknowledgement email, and our team confirms current capacity and the build before hardware is allocated to you.

You will hear from us either way. If you have paid and have not had a follow-up you were expecting, reply to the acknowledgement email or open a ticket from Support and we will tell you exactly where the order sits.

Setup Fee and Billing

GPU servers carry a one-time setup fee on top of the monthly price, in common with dedicated and storage servers, because a physical machine is built and handed over. Cloud servers carry no setup fee.

The setup fee is charged once, never recurs, and is shown separately from the monthly price at checkout.

Because a GPU tier is a monthly commitment on expensive hardware rather than an hourly rental, treat the decision accordingly. If you only need a GPU for a short burst, say so when you talk to us.

Reserving a Tier

When a tier has no capacity, the plan card offers a reservation rather than a configure action. Joining the queue records your interest in that tier.

Managing a GPU Server

Once a GPU machine is live it is managed exactly like any other bare-metal server, from Dedicated Servers in the KPanel sidebar. There is no separate GPU item in the sidebar.

The toolset, documented in full in Managing a Dedicated Server, covers:

• Power: graceful restart, hardware reset, forced power cycle, and Wake-on-LAN.
• Rescue system: a temporary Linux environment, or a browser-based virtual console.
• Reverse DNS: PTR records for each routed IP.
• IPs and subnets: what is routed to the machine, plus additional IPv4.
• Traffic: inbound and outbound volume by day, month, or year.
• Firewall: an upstream filter applied before traffic reaches the machine.
• Advanced: a full operating system reinstall, behind a typed confirmation.

Console access for a machine that will not boot is in Dedicated Server Console Access.

Practical Notes Before You Start

Drivers are yours on an unmanaged machine. Install the NVIDIA driver and CUDA stack that matches the framework versions you intend to run, and pin them. Mismatched driver and toolkit versions are the most common cause of a GPU that is present but unusable.

Watch your disk. Model checkpoints and data sets fill a disk faster than anything else on a server. If capacity is your real problem rather than compute, a storage server alongside the GPU machine is usually cheaper than a larger GPU tier.

Firewall it properly. A GPU box running an inference endpoint is an attractive target. Use the upstream firewall to allow only what needs to reach it, and keep Allow platform services switched on so you do not lock yourself out.

Reinstalling erases everything. The reinstall action under Advanced wipes every disk with no undo and takes no backup for you. Copy your checkpoints off first.

Getting Help

GPU is the product where talking to us first saves the most money. Ask Kora inside KPanel, open a ticket from Support, or email support@kapsulehost.com with what you are running and how much GPU memory you need.

For the rest of the compute range, see Cloud Servers Overview and Dedicated Servers Overview.