What you need to know: GPU VPS pairs a dedicated NVIDIA RTX PRO 6000 GPU with 18 vCPUs, 96 GB RAM, and 900 GB storage. It costs €999 / $1199 / £969 per month (excluding VAT), with 15% off on 12-month terms and 20% off on 24-month terms. It’s available in Europe and US Central and runs on an Ubuntu 24.04 CUDA image by default. It targets AI inference, rendering, and single-node simulation, not multi-node training clusters.
Today we’re launching GPU VPS, the first product in our range to pair a dedicated NVIDIA GPU with the flat monthly pricing and fast setup you already get from a Contabo Performance VPS. Offering dedicated GPU resources in a VPS format, it’s built to give you the advantages of both: a GPU reserved entirely for your instance, wrapped in the quick provisioning, transparent pricing, and flexibility of the VPS platform.
With AI driving more demand than ever for GPU processing power in user-friendly affordable formats, the GPU VPS gap was clearly waiting to be filled at Contabo. We already offer enterprise GPUs through the GPU Cloud, while elsewhere, the usual options are renting fractional computing time from a hyperscaler, taking your chances on a marketplace, or committing to an entire GPU server.
What’s been missing is a dedicated GPU that’s affordable, ready in minutes, and comes at a flat monthly price with no hidden fees. That’s exactly what GPU VPS offers. Now, you can run your own LLMs on your own infrastructure in an EU or US data center you choose. Self-hosted AI has never been simpler or more affordable.
So if you’ve been pushing inference, rendering, or simulation jobs onto a CPU-only VPS and hitting a wall, or watching a hyperscaler bill climb with every token, GPU VPS is for you. Read on for the full picture, including detailed information on the powerful NVIDIA card we chose and what it can do for your AI and ML projects.
A dedicated GPU, wired into a VPS
GPU VPS is a new tier alongside Core VPS, Performance VPS, and Max Performance VPS: a dedicated NVIDIA GPU attached to a Performance VPS, not a shared or fractional slice of one. The launch configuration, GPU VPS, comes with an NVIDIA RTX PRO 6000 GPU (96 GB of GDDR7 VRAM, built on NVIDIA’s Blackwell architecture), 18 vCPUs, 96 GB RAM, and 900 GB of NVMe storage.
The point of the format is to take the best of two setups that normally force a choice. The GPU is dedicated, so you get guaranteed parallel compute with no noisy neighbors on the card, the way a dedicated server gives you guaranteed hardware. The surrounding server is a VPS, so you get all the advantages of the Contabo VPS platform: customize it with your chosen apps and Add-Ons, have it running in minutes, pay one flat monthly price, and drop it when a project ends. Since the GPU carries the heavy lifting for these workloads, running it on the VPS platform rather than a full dedicated machine is exactly what keeps the price low and flexibility high.
GPU VPS comes standard with Ubuntu 24.04 and CUDA preinstalled. The usual VPS Add-Ons are also available, including Storage Extension, Private Networking, and a range of 1-click apps.
Pricing starts at €999 / $1199 / £969 per month (excluding VAT), with 15% off on 12-month terms and 20% off on 24-month terms. That’s one flat monthly fee, whether the GPU sits idle overnight or runs a rendering job at full load.
The card doing the work
The RTX PRO 6000 is a professional Blackwell-generation GPU with 96 GB of GDDR7 memory, 24,064 CUDA cores, and 120 TFLOPS of FP32 compute, on roughly 1.8 TB/s of memory bandwidth. We chose it because those numbers line up with what single-node AI, rendering, and simulation work actually needs, rather than with the biggest headline price tag.
The 96 GB of VRAM is the spec that matters most. It holds a 70-billion-parameter model on one card, and the 5th-generation Tensor Cores add FP4 support, so you can serve even larger or longer-context models at lower precision without splitting them across machines. For local LLM inference, RAG, and agent pipelines, that’s the difference between running on one GPU and fighting fragmentation across several instances.
For rendering, the 4th-generation RT Cores and 96 GB of memory keep large scenes resident on the GPU instead of paging them in and out, which is what drags out render times. The card is also ISV-certified for professional tools like Maya, Houdini, and Cinema 4D, so compatibility is guaranteed rather than hoped for. For scientific work, the strong FP32 throughput and high bandwidth move single-node CFD, molecular dynamics, and FEM runs along without stepping up to a multi-node cluster.
To put things in perspective: a card like this retails around €16,000 to buy outright, and GPU technology moves fast enough that owning one is a real bet on resale value and utilization. Renting it by the month at a flat price is a way to use that class of GPU without tying up the capital or carrying the depreciation. You get current hardware without having to guess where the GPU market goes next.
Where it fits into the GPU market
The GPU market splits into three lanes today, and each asks you to give something up. Hyperscalers give you managed access, but at a premium and with token-and-hour billing that can turn into runaway cost the moment a workload grows. Marketplace and spot GPU providers are cheaper, but availability and price move around, which is hard to plan a real project on. Dedicated GPU servers give you stable capacity, but with slower procurement, a higher entry cost, and a full-server commitment even for a single-GPU job.
GPU VPS is built to sit in the middle of all three. It gives you a dedicated GPU, reserved for your instance the way a Dedicated Server reserves hardware, so there is no oversubscription and no competing for GPU resources. It gives you a flat monthly price instead of metered billing, so cost is something you plan for once rather than worry about every month. And it provisions in minutes, so you get a GPU without the procurement wait or the full-server commitment a dedicated machine asks for. The workload runs on your own infrastructure, in an EU or US data center you pick, which keeps models, prompts, and data inside your own stack and jurisdiction rather than a hyperscaler’s.
Who it’s for
GPU VPS is built for teams and individuals whose work needs one dedicated GPU, not a cluster. The three groups below are likely to get the most out of it, but it has many other potential use cases as well.
For AI and ML developers: run inference, fine-tuning, or LoRA training on your own models, and keep local LLMs, RAG, and agent pipelines close to your app and data instead of behind someone else’s API. This is the group data sovereignty matters to most, since the model and its inputs never leave your instance.
For rendering and FX studios: 3D rendering, motion design, and compositing, with burst capacity for deadline spikes that a single workstation cannot absorb. The card’s professional tool certification means Maya, Houdini, and Cinema 4D behave as expected.
For scientific and engineering teams: single-node CFD, molecular dynamics, FEM, and data-science workflows that are too slow on a CPU but do not need a multi-node, NVLink-connected cluster.
GPU VPS is not aimed at large-scale, multi-node model training. For that class of workload, H100 or A100-based infrastructure with NVLink is the better fit, and we’ve designed GPU VPS to play to its strengths rather than competing there.
How to spin one up
Order GPU VPS as you would any other Contabo product, with setup complete in minutes. As mentioned earlier, the launch image is Ubuntu 24.04 with CUDA preinstalled, so there is no driver setup before you can use the GPU.
We’re launching GPU VPS in our Europe and US Central locations only for now, with plans to expand based on demand.
Finding the right balance
We think the honest answer to “why not just add GPUs to every location and every plan” is that GPU capacity does not scale the way CPU and storage capacity does. Sticking to two locations and a fixed configuration at launch is a practical decision that allows us to offer you this product at the right price and with the rock-solid reliability you’ve come to expect from Contabo.
What we do believe is that the gap between “rent a fraction of a hyperscaler GPU by the token” and “buy or lease an entire dedicated GPU server” was underserved, and that a dedicated GPU on top of VPS-style provisioning and VPS-style pricing fills it better than either extreme.
Where GPU VPS goes from here
We’re starting with a limited run of GPU VPS instances across our EU and US locations, with plans to expand based on demand. This means two things: if you’re eager to get your hands on a GPU VPS, don’t wait too long! And of course, if you have feedback or want to see this product in other configurations and locations, tell us. Your feedback is part of how we decide what comes next.
We built GPU VPS because we don’t believe self-hosted AI and GPU compute should be locked behind hyperscaler pricing or a full-server purchase. If you put it to work on something interesting, we’d genuinely like to hear about it, so feel free to get in touch and show us what you built.
FAQ
Why put a GPU in a VPS instead of a Dedicated Server?
The VPS format gives you the advantages of both setups. The GPU is dedicated, so you get guaranteed parallel compute with no sharing, while the VPS gives you fast self-service provisioning, flat pricing, and flexibility. It is not an upgrade from Performance VPS but a separate tier, for inference, rendering, and simulation work that CPU-only infrastructure cannot handle efficiently.
In which Contabo Locations can I order a GPU VPS?
GPU VPS launches in Europe and US Central so availability and performance stay consistent. We’ll be adding additional locations to expand quickly based on demand, so watch this space!
Is this a dedicated GPU, or shared access to a pool?
GPU VPS gives you a dedicated GPU, not best-effort access to shared accelerator capacity. The GPU is reserved for your instance alone, the same way a Dedicated Server reserves physical hardware for one customer. That is why GPU sits as its own tier: you choose it when a workload needs guaranteed parallel compute, not general-purpose CPU performance.
Why the RTX PRO 6000 instead of an H100 or H200?
GPU VPS is built for single-node AI inference, rendering, and simulation, where an RTX PRO 6000 offers a stronger price-to-usefulness balance than H100-class accelerators. Those cards are built for large-scale, multi-node training across many connected GPUs, which is a different job than running your own model or rendering a scene on one machine.
Can I migrate my GPU VPS instance to another region later?
No. Because GPU VPS supply is limited, cross-region migration is not available. This ensures that we can keep availability consistent.
Is GPU VPS suitable for training, or only inference?
GPU VPS is built mainly for inference and other single-node GPU workloads: rendering, video processing, and data science. Selected fine-tuning and LoRA training workloads fit well within a single GPU’s memory and compute budget. Large-scale, multi-node training is a different class of infrastructure, and GPU VPS is not positioned to replace it.
Does GPU VPS support Windows?
Not at this time. If your workflow depends on Windows, this is not the tier for you yet.