✨ NEW · A100 + A40 inference GPUs now available in Amsterdam — read more → ✨ NEW · A100 GPUs in Amsterdam → STATUS
Contact Sales Get Started

AI & ML

GPU compute for inference, training & fine-tuning

Dedicated NVIDIA accelerators with NVLink, low-latency InfiniBand, and the bandwidth to move datasets between regions in minutes — not hours. From single-GPU experimentation to 8× A100/H100 NVLink nodes for multi-day training runs and parameter-efficient fine-tuning of open-source LLMs.

Configure your GPU server Talk to engineering

A100 / H100 in stock · NVLink + InfiniBand · 24–72h deploy

Frontier accelerators

From a single L4 to 8× H100 NVLink nodes

Skip the hyperscaler waitlist. Our standard catalogue spans L4 / L40S / A10 / A40 / A100 / H100 / H200 — single-GPU instances for prototyping and experimentation, all the way up to 8× SXM nodes wired with NVLink and NVSwitch for distributed training. PCIe and SXM both available, real stock and real lead times listed on every plan card. Custom builds (4× A100 PCIe + InfiniBand fabric, 8× H200 with shared-storage Lustre) ship in 5–14 days, not quarters.

Browse GPU plans

GPU fleet — single-GPU experimentation up to 8× SXM nodes with NVLink

Why ML teams pick us

Built for real training workloads

  • iconcolor performance Bare-metal GPUs Full physical GPU, no shared timeslicing, no MIG partitioning unless you opt in. NVLink and NVSwitch wired across SXM nodes for full inter-GPU bandwidth.
  • iconcolor fastpresence Same-day to 2-week deploy Standard SKUs ship in 24–72h. Custom 8× A100/H100 builds in two weeks. No “contact sales” loop for the first server — just configure and go.
  • iconcolor scalability Train where your data lives Datacentres in NL, DE, SE, RO, US, HK. Place training next to your data lake; inference next to your users. Cross-region transfer is unmetered between our PoPs.
  • iconcolor security Engineers, not ticket-routers 15-minute SLA on hardware faults. The first reply on a Slack-shared channel is from someone who can run nvidia-smi and read your training logs.

Reserved capacity

Lock in capacity for a training run

3 / 6 / 12-month reserved pricing with a discount that scales with commitment. Guaranteed topology — NVLink, NVSwitch, InfiniBand fabric — and dedicated power. Right for production training fleets and reproducible multi-month research runs where the hyperscalers will quote you a year out.

Request a reservation quote

Storage & data

Bring your own weights, checkpoints, datasets

High-throughput object storage in the same DC as your GPUs, with pre-warmed networking to upload from your existing S3-compatible buckets. Typical 1 TB transfer over the 100 GbE uplink completes in 10–20 minutes. NFS / Lustre available on request for shared training datasets.

About cloud storage

Network performance

Move datasets in minutes, not hours

Up to 100 GbE per server, with 25 / 40 / 100 GbE intra-rack interconnects on InfiniBand-capable nodes. Tier-1 transit + IXP peering at every site (AMS-IX, DE-CIX, NL-IX, NIX.CZ, MIX-IT) keeps egress to your edge users sub-15 ms p95 across Europe and sub-10 ms within the same region. No per-GB egress charges — your transfer quota is included.

About our network

Network performance — sustained 100 GbE per server with sub-15ms p95 across Europe

Preconfigured servers · AI & ML 3 configurations

Best Value GPU

Accelerator RTX6000

AMD EPYC 7702p

128 GB RAM · NVMe · 10G uplink

Starting at 424/mo

Global footprint

Deployed in 5+ datacenters across three continents

Pick the location closest to your users — or to your existing infrastructure. Every location runs the same hardware, same network, same NOC.

Built for production

Why teams stay with Netrouting

Network engineers, not order takers. Hardware that lasts, support that picks up, infrastructure you can grow into.

  • Expert-Level Support Our staff is available 24 hours a day, 7 days a week to handle network administration and systems management issues as they occur.
  • Scalable Solutions You can build whatever level of depth or brevity you require of your network and compute infrastructure and then expand as necessary to meet future needs.
  • Enhanced Security Enable 2-factor authentication and restrict by IP address from the control panel to protect your account.
  • Cost-Efficient Infrastructure You will always receive the best value from your investment as you will be optimized for budget without any compromise on Quality.

Common questions

Frequently Asked Questions

  • Do you support PyTorch / TensorFlow / JAX out of the box?
    Yes. Our GPU images ship Ubuntu 22.04 LTS with the matching CUDA + cuDNN + NCCL stack pre-installed. Container runtimes (Docker, containerd, NVIDIA Container Toolkit) are one click in the deploy form, and you can drop your own image via Packer or a custom ISO if you prefer.
  • Can I get reserved capacity for a multi-month training run?
    Yes — 3 / 6 / 12-month reserved pricing is available with a discount that scales with commitment. Talk to engineering before booking large A100/H100 capacity so we can guarantee the topology you want (NVLink, NVSwitch, InfiniBand fabric).
  • Are GPUs bare-metal or virtualised?
    Bare metal. You get the full physical GPU with no shared timeslicing and no MIG partitioning unless you opt in. PCIe-attached and SXM nodes both available; NVLink/NVSwitch is wired across SXM systems for full inter-GPU bandwidth.
  • What GPUs are in stock right now?
    Standard catalogue covers L4 / L40S / A10 / A40 / A100 / H100 / H200 across single-GPU, 2× and 8×-GPU nodes. Real-time stock + lead times on each plan card. Custom builds (4× A100 PCIe with InfiniBand, 8× H200 NVLink, etc.) ship in 5–14 days.
  • Can I bring my own model weights and checkpoints?
    Yes. We provide high-throughput object storage in the same datacentre and pre-warmed networking to upload from your existing buckets — typical 1 TB transfer from S3-compatible storage takes 10–20 minutes on the 100 GbE uplink. NFS / Lustre available on request for shared training datasets.
  • How do you bill GPU usage?
    Hourly with no minimum (start a 4-hour fine-tune run and pay for 4 hours), or monthly / quarterly / annual reserved pricing for production training and inference fleets. No per-GB egress charges, no surprise data-transfer bills — egress is included up to the listed monthly transfer quota.

$ netr deploy --ready

Ready to deploy?

200+ teams trust Netrouting for bare metal, cloud, GPU and colocation. Configure a server in minutes — or talk to an engineer for custom builds.

Browse plans Talk to engineering