BARE METAL GPU

Dedicated GPU power on demand

Single-tenant servers with NVIDIA professional and data center GPUs — built for AI training and inference, 3D rendering, video transcoding, scientific compute and virtual workstations. No virtualization overhead, no noisy neighbours.

See configurations →

Why bare metal for GPU workloads

Processing Unit
Cloud GPUs share PCIe lanes, memory bandwidth and host CPU with other tenants. For sustained training runs, real-time inference and rendering pipelines, that contention shows up as inconsistent throughput and longer job times. A bare metal GPU server hands you the entire box: the PCIe topology, the CPU pinning, the NVMe array, the network interface.
You install your own OS image, your own CUDA stack, your own scheduler — or we provision Ubuntu LTS with NVIDIA drivers and CUDA ready out of the box. Run Docker, Kubernetes, Slurm, Proxmox or anything else without restrictions. Every server includes IPMI/iLO access so you can reinstall, reboot, console in and monitor hardware health 24/7.

GPU Configurations

Standard configurations — custom builds available on request.

RTX A4000 / 16 GB

Entry tier — ML inference, light training, virtual workstations

  • GPU: 1× NVIDIA RTX A4000 (16 GB GDDR6, 6144 CUDA cores)
  • CPU: Intel Xeon / AMD EPYC, 8 cores
  • RAM: 64 GB DDR4 ECC
  • Storage: 2× 960 GB NVMe SSD
  • Bandwidth: 1 Gbps unmetered
  • OS: Ubuntu 22.04 LTS / Windows Server / custom

Contact for pricing →

RTX A5000 / 24 GB

Professional tier — mid-size model training, rendering farms

  • GPU: 1× NVIDIA RTX A5000 (24 GB GDDR6, 8192 CUDA cores)
  • CPU: Intel Xeon / AMD EPYC, 16 cores
  • RAM: 128 GB DDR4 ECC
  • Storage: 2× 1.92 TB NVMe SSD
  • Bandwidth: 1 Gbps unmetered
  • OS: Ubuntu 22.04 LTS / Windows Server / custom

Contact for pricing →

RTX A6000 / 48 GB

High-end tier — large model fine-tuning, multi-stream rendering

  • GPU: 1× NVIDIA RTX A6000 (48 GB GDDR6 ECC, 10752 CUDA cores)
  • CPU: AMD EPYC, 32 cores
  • RAM: 256 GB DDR4 ECC
  • Storage: 2× 3.84 TB NVMe SSD
  • Bandwidth: 10 Gbps
  • OS: Ubuntu 22.04 LTS / Windows Server / custom

Contact for pricing →

NVIDIA A100 / 80 GB

Data center tier — LLM training and high-throughput inference

  • GPU: 1–8× NVIDIA A100 80 GB HBM2e (PCIe or SXM4 NVLink)
  • CPU: Dual AMD EPYC, up to 128 cores
  • RAM: 512 GB – 2 TB DDR4 ECC
  • Storage: 4× 7.68 TB NVMe SSD
  • Bandwidth: 10–25 Gbps
  • OS: Ubuntu 22.04 LTS with CUDA + NCCL preinstalled

Contact for pricing →

NVIDIA H100 / 80 GB

Flagship tier — transformer training, frontier-scale inference

  • GPU: 1–8× NVIDIA H100 80 GB HBM3 (PCIe or SXM5 NVLink)
  • CPU: Dual AMD EPYC Genoa, up to 192 cores
  • RAM: 1–2 TB DDR5 ECC
  • Storage: 8× 7.68 TB NVMe SSD (RAID)
  • Bandwidth: 25–100 Gbps
  • OS: Ubuntu 22.04 LTS with CUDA + NCCL + NeMo stack

Contact for pricing →

Custom build

Tell us the workload — we spec the box

  • Multi-GPU NVLink topologies
  • L4, L40 / L40S, RTX 4090 also available
  • Private network interconnect (10/25/100 GbE)
  • NVMe-oF or Ceph storage backends
  • Colocation of customer-owned hardware
  • Reserved capacity and long-term contracts

Request a quote →

Common workloads

AI / ML training

Train, fine-tune and evaluate transformer, diffusion and vision models. Use PyTorch, JAX, TensorFlow or any framework you want. NVLink-connected A100/H100 nodes for multi-GPU jobs that need fast all-reduce.

Inference at scale

Serve LLMs, image generation, speech and embedding models with predictable latency. Triton Inference Server, vLLM, TGI, llama.cpp — you pick the runtime, we provide the hardware.

3D rendering & VFX

OptiX, RTX-accelerated path tracing, Blender Cycles, Redshift, V-Ray and Octane. Render farms scale linearly across our A5000 and A6000 nodes.

Video transcoding

NVENC/NVDEC hardware encoders for live and VOD pipelines — H.264, H.265/HEVC and AV1. Multiple concurrent streams per GPU at high quality.

Scientific compute / HPC

CUDA, OpenACC and HIP workloads — molecular dynamics, computational fluid dynamics, genomics, finance simulations. Bring your own MPI stack or Slurm cluster.

Virtual workstations

Remote GPU desktops via NICE DCV, Parsec or HP Anyware. Designers and engineers connect from anywhere; the heavy lifting stays on the server.

Hosted in Sweden

All GPU servers are physically located in our Swedish data centers, powered by Nordic hydro and wind — among the lowest carbon-intensity grids in Europe. EU data residency, GDPR-compliant operations, redundant power and cooling, on-site engineering response 24/7.

Network: multi-homed transit through Tier 1 carriers, IPv4 + IPv6, private interconnect to NetNod, STHIX and Equinix peering. Sub-millisecond RTT to Stockholm, low single-digit ms to Frankfurt, Amsterdam and Helsinki.

Frequently asked

Yes. You get full root / Administrator access from day one. Reinstall the OS yourself via IPMI/iLO at any time, pick any kernel, run any container runtime.

Dedicated. Every GPU listed is passed through directly to your OS — no MIG slicing, no vGPU partitioning, no other tenants on the same card. You get 100% of the VRAM, CUDA cores, tensor cores and PCIe bandwidth.

Stock configurations are typically live within a few hours. Custom builds with hardware we need to order in usually deploy within 3–7 business days.

Yes — Windows Server licensing is available, common for virtual workstations and certain rendering / GIS workflows. Linux (Ubuntu, Rocky, Debian, RHEL) is more common for AI / ML.

Standard billing is monthly. Reserved capacity (3, 6, 12 months) gets a discount. Short-term or hourly arrangements for burst training workloads can be discussed — contact us with the use case.

Yes — colocation in our data centers is available. We rack, power and network your hardware; you keep ownership and full control. See the colocation page.

Have a workload in mind?

Tell us the model size, the framework and the throughput you need — we’ll come back with a configuration and a price.

Get in touch →