Bare Metal GPU servers — NVIDIA Blackwell in a Polish data center
A dedicated GPU server reserved entirely for you: NVIDIA RTX 5090 and RTX PRO 6000 Blackwell cards, AMD EPYC processors and NVMe Gen4 drives in RAID. Full root access and an IPMI console, any operating system or hypervisor, data and inference kept in Poland — TIER III facilities in Gdańsk and Gdynia connected by a 100Gbps ring.
What will you run on a GPU server?
AI inference
Serve LLM and multimodal models (Llama, Qwen, Mistral) with vLLM or TensorRT-LLM. 96GB of VRAM on a single RTX PRO 6000 fits 70B-class models in FP8 on one card, and <5 ms latency across Poland makes the difference for real-time APIs.
LLM fine-tuning and training
LoRA/QLoRA and full fine-tuning on multi-GPU machines — up to 384GB of VRAM in the 4x RTX PRO 6000 configuration. Full PCIe x16 lanes to every card thanks to AMD EPYC platforms, with no resources shared with other tenants.
Rendering and CUDA workloads
3D rendering (Blender, Octane, Redshift), video transcoding and scientific simulations. GDDR7 memory and NVMe Gen4 RAID arrays cut the load times of scenes, datasets and checkpoints.
Why Bare Metal GPU at nodea?
Data and inference in Poland
GDPR-compliant processing and full data sovereignty — your models and data never leave the EEA. TIER III data centers in Gdańsk and Gdynia.
<5 ms latency across Poland
Data centers in Gdańsk and Gdynia instead of Falkenstein or Amsterdam (25-40 ms). A decisive difference for real-time inference and responsive LLM APIs.
Three availability zones (3-AZ)
Three independent data centers in the Tricity — MP100 and GAC133 (TIER III) and IQ Data Center — linked by a redundant 100Gbps fiber-optic ring. The loss of an entire facility does not stop your services: you keep replicas and checkpoints in physically separated buildings, and private transfer between your servers comes at no extra charge.
24/7 engineer support
A Linux engineer answers your ticket, not a bot. We react proactively to monitoring alerts.
Bare metal, no virtualization
Full GPU and CPU performance reserved for you — no neighbours, no provider hypervisor. Running your own hypervisor or Kubernetes is allowed.
VAT invoice in PLN
Billing in Polish złoty with no exchange-rate risk — foreign providers invoice in EUR.
2026-grade hardware
NVIDIA Blackwell with GDDR7, AMD EPYC Genoa and NVMe Gen4 in RAID — no compromises on older card generations.
Bare metal or a cloud instance?
You can take a GPU server either as dedicated hardware reserved exclusively for you, or as a cloud instance billed by the hour. These are two different tools — the choice depends on whether you need maximum, predictable performance or flexibility and a fast start.
- Choose bare metal GPU when you run production inference or long training jobs, need full PCIe x16 lanes to every card and steady performance with no neighbours on the same machine. You get hardware isolation, full control over the hypervisor, Proxmox or a Kubernetes cluster, and no virtualization overhead — all GPU power goes to your workload.
- Choose a cloud instance when you are building a prototype, have variable or seasonal load, want to start in minutes and pay by the hour rather than for a whole month. It is a good fit for experiments, model testing and workloads that do not need to run around the clock — see our cloud instances.
Not sure which side your project falls on? Describe your workload and budget — we will help you pick the billing model and configuration.
Bare Metal GPU pricing
GPU.5090_1
2199 zł net/ per month (2704.77 zł incl. VAT)
- 1x NVIDIA RTX 5090 32GB GDDR7
- AMD Ryzen 9 9950X (16C/32T, 5.7GHz boost)
- 128GB DDR5 ECC
- 2x 2TB NVMe Gen4 (RAID1)
- 1Gbps unmetered (burst 10Gbps)
- DC Gdańsk or Gdynia
- Setup fee 499 zł net
GPU.5090_2
3899 zł net/ per month (4795.77 zł incl. VAT)
- 2x NVIDIA RTX 5090 32GB GDDR7 (64GB VRAM total)
- AMD EPYC 9354P (32C/64T)
- 256GB DDR5 ECC
- 2x 3.84TB NVMe Gen4 (RAID1)
- 2Gbps unmetered (burst 10Gbps)
- DC Gdańsk or Gdynia
- Setup fee 999 zł net
GPU.6000_1
4599 zł net/ per month (5656.77 zł incl. VAT)
- 1x NVIDIA RTX PRO 6000 Blackwell 96GB GDDR7 ECC
- AMD EPYC 9454P (48C/96T)
- 256GB DDR5 ECC
- 2x 3.84TB NVMe Gen4 (RAID1)
- 2Gbps unmetered (burst 10Gbps)
- DC Gdańsk or Gdynia
- Setup fee 999 zł net
GPU.6000_2
8499 zł net/ per month (10453.77 zł incl. VAT)
- 2x NVIDIA RTX PRO 6000 Blackwell 96GB (192GB VRAM total)
- AMD EPYC 9554P (64C/128T)
- 512GB DDR5 ECC
- 2x 7.68TB NVMe Gen4 (RAID1)
- 5Gbps unmetered (burst 25Gbps)
- DC Gdańsk or Gdynia
- Setup fee 1499 zł net
GPU.6000_4
15999 zł net/ per month (19678.77 zł incl. VAT)
- 4x NVIDIA RTX PRO 6000 Blackwell 96GB (384GB VRAM total)
- 2x AMD EPYC 9454 (96C/192T)
- 768GB DDR5 ECC
- 4x 7.68TB NVMe Gen4 (RAID10)
- 10Gbps unmetered
- Access to the 100Gbps 3-AZ ring
- DC Gdańsk + configuration replica in Gdynia on request
- Setup fee 2999 zł net
Ready to work from day one
Preinstalled CUDA/cuDNN drivers
The server arrives with the NVIDIA stack already configured — CUDA and cuDNN drivers ready to go, on Ubuntu or Debian images. You start by running your model, not by installing drivers.
Full root access and IPMI console
No limits, no middlemen — root on the machine plus a remote IPMI/KVM console for reinstalls, BIOS access and rescue access, even when the OS won't boot.
Any OS and hypervisor
Install whatever you like: your own Linux, Proxmox, plain KVM or a Kubernetes cluster. The hardware is yours, so the choice of system and orchestration layer is up to you.
NVMe Gen4 RAID + backup
Fast NVMe Gen4 arrays in RAID for datasets and checkpoints, with an off-server backup option. Your data stays close to the GPU and under your control.
1-10Gbps unmetered network + VLAN
Uplinks from 1 to 10Gbps with unmetered transfer, and the option of a private VLAN between your servers and replicas on the 100Gbps ring.
Integration with the app.nodea.io platform
You connect the server to our platform at app.nodea.io/app: Web Monitoring watches service availability and performance 24/7, nodeaD delivers anti-DDoS protection with automatic self-healing, and the Connection Manager gives you structured access to your servers. An engineer responds to alerts — proactively, before an issue grows.
99.99% availability SLA
We guarantee 99.99% infrastructure availability, backed by redundant power and network in TIER III facilities. If the SLA is not met, you receive compensation under the contract — with no hidden exclusions.
Your data stays in Poland
GDPR-compliant processing within the EEA — no data transfer outside the EU. Billing and VAT invoices in PLN, with none of the exchange-rate risk of providers invoicing in EUR.
Frequently asked questions
Are the listed prices net prices?
Yes, all prices in the table are net — 23% Polish VAT is added on top. We issue VAT invoices in PLN, so you avoid the exchange-rate risk of EUR billing.
How long does GPU server provisioning take?
Standard delivery time for a Bare Metal GPU server is 3-5 business days from the payment being booked. In that time we assemble and burn-in test the platform, configure the network and install the operating system of your choice.
Can I order a configuration other than the listed ones?
Yes — custom configurations (a different number of cards, more RAM, larger drives, additional uplinks) are quoted individually on request. Reach out through the contact form, describe your requirements and we will prepare an offer.
Is there a discount for a longer commitment?
Yes — with a 12-month commitment you get a -10% discount on the monthly price and the setup fee is waived.
Do I get full root access and an IPMI console?
Yes. You get full root access to the machine plus a remote IPMI/KVM console, so you can reinstall the OS yourself, enter the BIOS and keep rescue access even when the operating system won't boot.
Which operating systems and drivers are available?
We install Ubuntu and Debian images with a preinstalled NVIDIA stack — CUDA and cuDNN drivers ready to go. You can also deploy your own system, Proxmox, plain KVM or a Kubernetes cluster; the choice of system layer and hypervisor is yours.
How does bare metal differ from a cloud GPU instance?
Bare metal is a dedicated server for your exclusive use — full PCIe x16 lanes to every card, no neighbours on the machine, no virtualization overhead and maximum, predictable performance for training and production inference. A cloud instance is billed by the hour and starts in minutes, so it fits prototypes and variable workloads better. For that second scenario we offer cloud instances.
Let's talk about your AI project
Not sure which configuration will carry your model? Describe your use case — we will help you pick the number of cards, the amount of VRAM and the network topology.