GPU dedicated servers. The full GPU. Nobody else on it.

A physical server rented to one customer, with a graphics card installed and passed through to that customer’s own operating system. Nothing is virtualised or shared: you get every core, all the memory, the disks and the full GPU, with root access and a remote console.

A complete machine with a GPU installed runs from $91.72 to $1,682.67 a month. The price is two lines added together: the machine, and the GPU at $20.79 to $1,599 a month. Processor, memory, two disks, an unmetered port and one IPv4 address are included. Setup is a one-time $29 on a build with a GPU installed to order; a ready machine ordered as it stands has none.

No. One physical machine, one account, and the GPU passed through to the operating system you installed. There is no hypervisor, no MIG partition, no vGPU profile and no time-slicing. The numbers you benchmark on day one are the numbers you keep.

Yes. One month is the default term and there is nothing to sign. Three, six and twelve-month cycles save up to 15%, and the price does not step up at renewal. Pay by credit card or PayPal; for a larger order a bank wire can be arranged, and it usually clears fraud review faster than a card payment.

Yes. The list on this page is sorted cheapest first, and the cheapest complete build is Quadro 3000M in Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps at $91.72 a month. Older Tesla and Quadro GPUs still transcode video and drive remote desktops well, and they come in the same whole machine with the same unmetered port.

57 GPUs are on the list, from current NVIDIA compute accelerators and AMD Instinct down to older Tesla, Quadro and GeForce models. Each one is listed by name with its memory and the complete monthly price. If the GPU you want is missing, ask: we can often buy it in.

The ones you install. The server arrives with a clean operating system and root access, and you pin the driver, the CUDA or ROCm release and the framework build your code was tested against. If you want the driver in place at handover, put the exact version on the order. If a GPU has to support a particular CUDA release, ask before paying and we will check.

Often, and we quote it by hand: chassis clearance, the GPU’s width, the PCIe lanes and the power supply decide the count. Name the GPU and we reply with the chassis, the count and the lead time. The usual reason is memory rather than speed, since two 48 GB GPUs hold a model one 80 GB GPU cannot. Multi-GPU and multi-node servers covers the next step.

Yes, in the server you already have: no migration, no reinstall, no new IP addresses. It is a scheduled visit with a short outage, and the bill moves by exactly the difference between the two GPUs. RAM, drives and a faster port are added under the same rule.

We replace it within your support plan’s window, counted from when you report it: 24 hours on the included Basic plan, 4 on Bronze, 3 on Silver, 2 on Gold and Platinum. Past the window, service credit follows the schedule published on the support page.

Yes, a GPU you own, or any other part, can go into the server you rent here. It is not free, because the GPU draws power, and the charge depends on the hardware. We are not responsible if a part you brought fails; the replacement window covers only the hardware we supply. Book a call before you buy or ship it: we price it and check the chassis takes it, because a blade, for one, has no room for a full-size GPU.

Nothing is metered in either direction at any port speed we sell, so egress costs nothing. On GPU work that is often the largest line elsewhere: training pulls datasets in and pushes checkpoints out, and an inference endpoint answers real traffic. Unmetered bandwidth lists the faster ports.

Hourly wins while the work is bursty and really stops, up to a few hundred hours a month. Past about two thirds of the month, monthly is cheaper before you count egress, storage kept between runs and time spent waiting for capacity. Monthly also keeps the same machine, disks and addresses.

Within three days of activation, after troubleshooting, our refund policy applies as written. The memory table and the PCIe column above are there so you can pick the right GPU first, and we would rather talk you down to a smaller GPU in chat than refund a bigger one.