Skip to content

Cloud VPS · GPU

A GPU card, and an honest description of how much of it is yours.

17 sizes, from 2 GB slice of one A16 at $77.90 a month to 8× whole cards of NVIDIA B300 SXM with 2304 GB of VRAM. Every size prints its card, its VRAM and whether it is a slice or a card of its own — and every size lists only the cities where that card is actually in stock.

Configure a GPU serverSee the sizes
Rendering of a graphics card, fans facing upThe card, or a sliceOnly where in stock

Sizes

17 sizes, 7 cards, and where each one is built.

Slices are priced as slices and whole cards as whole cards — the 'What you get' column says which, in the supplier's own terms. The site column is not a line-wide footprint — it is where that exact size has stock.

Start here

Cheapest way in

NVIDIA A16 2 GB · 2 vCPU · 8 GB

G2-8

The least expensive on this line — and $9.74 per GB a month.

$77.90/mo

ConfigureDetails

The middle of the line

AMD Instinct MI325X 256 GB · 20 vCPU · 160 GB

G20-160

Half the sizes on this line cost less and half cost more — and $31.81 per GB a month.

$5089.90/mo

ConfigureDetails

Most RAM

NVIDIA B300 SXM 2304 GB · 224 vCPU · 3840 GB

G224-3840

The largest on this line: 3840 GB — and $24.64 per GB a month.

$94608.90/mo

ConfigureDetails

Or narrow it down

RAM
Disk
Cores

17 machines on this line, priced from their cheapest city.

  • NVIDIA A16 2 GB · 2 vCPU · 8 GB

    G2-8 · 1 TB transfer · 7 cities

    $77.90/mo

    ConfigureDetails

  • NVIDIA A40 2 GB · 1 vCPU · 5 GB

    G1-5 · 3 TB transfer · New York

    $99.90/mo

    ConfigureDetails

  • NVIDIA A16 4 GB · 2 vCPU · 16 GB

    G2-16 · 2 TB transfer · 5 cities

    $154.90/mo

    ConfigureDetails

  • NVIDIA A16 8 GB · 3 vCPU · 32 GB

    G3-32 · 3 TB transfer · 4 cities

    $309.90/mo

    ConfigureDetails

  • NVIDIA A16 32 GB · 12 vCPU · 128 GB

    G12-128 · 10 TB transfer · Bangalore

    $1238.90/mo

    ConfigureDetails

  • NVIDIA L40S 48 GB · 8 vCPU · 96 GB

    G8-96 · Traffic not counted · Warsaw

    $2317.90/mo

    ConfigureDetails
Every GPU plan with its card, VRAM, how much of a physical card it is, the machine around it, where it is built and its monthly price
PlanCardVRAMWhat you getMachineBuilt inMonthly
G2-8NVIDIA A162 GB2 GB slice of one A16 (shared card)2 vCPU · 8 GB RAM · 50 GB NVMe · 1 TB🇺🇸 Atlanta, 🇮🇳 Bangalore, 🇩🇪 Frankfurt, 🇺🇸 New York, 🇺🇸 Chicago, 🇺🇸 Silicon Valley, 🇸🇬 Singapore$77.90
G1-5NVIDIA A402 GB2 GB slice of one A40 (shared card)1 vCPU · 5 GB RAM · 90 GB NVMe · 3 TB🇺🇸 New York$99.90
G2-16NVIDIA A164 GB4 GB slice of one A16 (shared card)2 vCPU · 16 GB RAM · 80 GB NVMe · 2 TB🇺🇸 Atlanta, 🇮🇳 Bangalore, 🇯🇵 Tokyo, 🇺🇸 Silicon Valley, 🇸🇬 Singapore$154.90
G3-32NVIDIA A168 GB8 GB slice of one A16 (shared card)3 vCPU · 32 GB RAM · 170 GB NVMe · 3 TB🇺🇸 Atlanta, 🇮🇳 Bangalore, 🇺🇸 Silicon Valley, 🇸🇬 Singapore$309.90
G12-128NVIDIA A1632 GB2× whole cards12 vCPU · 128 GB RAM · 700 GB NVMe · 10 TB🇮🇳 Bangalore$1238.90
G8-96NVIDIA L40S48 GB1× whole card8 vCPU · 96 GB RAM · 1490 GB NVMe · unmetered traffic🇵🇱 Warsaw$2317.90
G24-240NVIDIA H10080 GB1× whole card24 vCPU · 240 GB RAM · 2794 GB NVMe · unmetered traffic🇵🇱 Warsaw$4519.90
G16-192NVIDIA L40S96 GB2× whole cards16 vCPU · 192 GB RAM · 2980 GB NVMe · unmetered traffic🇵🇱 Warsaw$4635.90
G20-160AMD Instinct MI325X256 GB1× whole card20 vCPU · 160 GB RAM · 720 GB NVMe · 15 TB🇨🇦 Toronto$5089.90
G20-240NVIDIA H10080 GB1× whole card20 vCPU · 240 GB RAM · 720 GB NVMe · 15 TB🇳🇱 Amsterdam, 🇺🇸 New York, 🇨🇦 Toronto$5905.90
G24-240-720NVIDIA H200141 GB1× whole card24 vCPU · 240 GB RAM · 720 GB NVMe · 15 TB🇺🇸 Atlanta, 🇺🇸 New York$5986.90
G48-480NVIDIA H100160 GB2× whole cards48 vCPU · 480 GB RAM · 5588 GB NVMe · unmetered traffic🇵🇱 Warsaw$9039.90
G32-384NVIDIA L40S192 GB4× whole cards32 vCPU · 384 GB RAM · 5960 GB NVMe · unmetered traffic🇵🇱 Warsaw$9271.90
G64-768NVIDIA L40S384 GB8× whole cards64 vCPU · 768 GB RAM · 11921 GB NVMe · unmetered traffic🇵🇱 Warsaw$18542.90
G56-960NVIDIA B300 SXM576 GB2× whole cards56 vCPU · 960 GB RAM · 5588 GB NVMe · unmetered traffic🇫🇷 Paris$29896.90
G112-1920NVIDIA B300 SXM1152 GB4× whole cards112 vCPU · 1920 GB RAM · 11176 GB NVMe · unmetered traffic🇫🇷 Paris$53737.90
G224-3840NVIDIA B300 SXM2304 GB8× whole cards224 vCPU · 3840 GB RAM · 22352 GB NVMe · unmetered traffic🇫🇷 Paris$94608.90

Prices are in USD, exclude tax and renew at what you first paid. Annual billing takes 10% off the same rate. There is no setup fee and no minimum term beyond the one you choose.

What you are buying

A Linux server with a card attached — and the driver stack is your job.

These machines are resold supplier stock. That buys you cards we could not otherwise offer, and it costs you the conveniences of our classic platform. Both halves are below.

The card, stated as a fraction where it is one

Where a size is a fraction of a physical card, it says so — “2 GB slice of one A16 (shared card)” — and where it is whole cards, it says how many. The fraction comes from the supplier’s own plan definition, not from us. Nothing here rounds a slice up to a card.

Plain Linux, no preinstalled stack

You get root on one of the same six Linux images the rest of the catalogue ships. NVIDIA drivers, CUDA and your framework are yours to install and yours to keep patched. There is no machine-learning image and no agent of ours on the machine.

No snapshots, resize or managed backups on this line

Those are features of our classic virtual platform, and this is a different platform underneath. Take your own backups off the box — restic, borg or rsync to storage you control — because there is no nightly backup add-on to buy here.

No benchmarks, because ours would not be yours

Throughput depends on your model, batch size, precision and framework version. We publish the card, the VRAM and the share; the only measurement worth trusting is the one you take on your own workload, and a monthly term is short enough to take it.

Ordering

Stock is confirmed before the machine is built.

This is the one place where a GPU order behaves differently from a CPU one, so it is worth reading before you pay rather than after.

Built to order, not in a minute
Allow one business day. The order is checked against the supplier's real stock for that card in that city, and the machine is built once it is confirmed.
If the card has gone, you hear first
Nothing is charged for a machine we cannot build. We tell you which other size or site can be built instead, and you decide.
Availability is per size, not per line
Footprints differ by size — G2-8 is in 7 cities, G224-3840 in 1. The configurator only offers the sites where the size you picked has stock, and the cart refuses the rest.
Monthly or annual, your choice per server
Annual is 10% off the same rate for paying twelve months up front. Prepaid terms are not refunded if you destroy the server early, which is why monthly is the default.
Self-managed, like the rest of the catalogue
Root is yours, patching is yours, and the driver stack is yours. Support answers questions about the machine and the network, not about your training run.
Bigger cards are quoted, not listed
H100, H200 and MI300-class capacity moves week to week. Email [email protected] with the card, the duration and the region for an answer about what is actually orderable.

Questions

The seven we are asked before every GPU order.

Including the two answers that lose us sales, which is usually a sign they are the true ones.

Is the GPU shared?
It depends on the size, and every size says which it is. The smaller sizes are slices of one physical card — the 'What you get' column prints the fraction, for example a 2 GB slice of one A16 with other tenants on the rest. The larger sizes are whole cards, one or several, with nobody else on them. We label it that way because it is the fact that decides whether a plan suits you, not a detail to discover after the first invoice.
Can I get an H100?
Yes, where it is listed. The whole-card sizes on this page right now are G12-128 (2× whole cards of NVIDIA A16), G8-96 (1× whole card of NVIDIA L40S), G24-240 (1× whole card of NVIDIA H100), G16-192 (2× whole cards of NVIDIA L40S), G20-160 (1× whole card of AMD Instinct MI325X), G20-240 (1× whole card of NVIDIA H100), G24-240-720 (1× whole card of NVIDIA H200), G48-480 (2× whole cards of NVIDIA H100), G32-384 (4× whole cards of NVIDIA L40S), G64-768 (8× whole cards of NVIDIA L40S), G56-960 (2× whole cards of NVIDIA B300 SXM), G112-1920 (4× whole cards of NVIDIA B300 SXM), G224-3840 (8× whole cards of NVIDIA B300 SXM). Each one is listed only in the cities where the supplier has confirmed the card, and that moves week to week — so the table is the answer, not a promise. For a card or a region that is not on it, write to [email protected] with the card, the duration and the region, and you get a straight answer about what is orderable that week.
Why is provisioning not instant?
Because there is a physical card at the other end of the order. A CPU instance comes from a pool that is effectively always there; a GPU is a scarce item that our supplier may have sold to somebody else this morning. So we confirm the card against real stock before the machine is built rather than after — allow one business day. If the card has gone, you hear that before anything is charged, and we say which of the other sizes or sites can be built instead.
Do you bill hourly?
No. Web orders are monthly or annual, and annual takes 10% off the same rate. The hourly figure printed beside each price is the monthly price divided across a 730-hour month — the number you need when comparing us against a metered provider, not a rate we charge. On GPUs the distinction matters more than it does on CPU plans: if your job runs for six hours a week, a metered provider will cost you less than any monthly plan, ours included, and we would rather write that here than have you work it out later.
Which size should I start on?
G3-32 for most work — 8 GB slice of one A16 (shared card) with 8 GB of VRAM at $309.90 a month. Eight gigabytes of VRAM is the threshold where a 7-billion-parameter model at 8-bit fits with room for context and where SDXL runs without fighting for memory. Start below it only if you want a CUDA machine to develop against; move above it when you hit a specific wall — a model that does not fit, or a queue that is deep because the card is shared. Both are obvious when they happen.
What is installed on the machine?
A plain Linux image and nothing else. NVIDIA drivers, CUDA, cuDNN and whatever framework you use are yours to install, on the same self-managed terms as every other server we sell. We do not ship a preconfigured machine-learning image, and we would rather say so on this page than have you find an empty driver directory at 2am.
Do you publish benchmarks?
No, and we do not intend to. Throughput on a GPU depends on your model, batch size, precision and framework version, so any figure we printed would be a number from a workload that is not yours. The specification is the card, the VRAM and how much of the card you get; those are facts. What you do with them is measurable only on your own workload, and the monthly term is short enough to measure it.

Keep reading

Where to go from here.

Not sure a GPU is the answer

Two cheaper places to look first.