Innovative Nebius data centers and hardware
Nebius’ data centers and hardware reflect our dedication to green energy. With servers and racks designed in-house, as well as modern solutions of entire facilities, we continuously scale our GPU-equipped fleet.
Efficient power supply
We connect directly to high-voltage lines maintained by multiple operators, with own Nebius cable lines and substations. Our diesel rotary uninterruptible supply accumulates energy for use in case of external outage.
With environment in mind
All three facilities are eco-friendly. In Finland, we utilize a solar panel farm and use excess heat generated from the servers to keep nearby houses warm. In France, waste heat will be used to warm the urban farm on the data center’s roof.
Own design
Nebius’ hardware R&D team designs and assembles servers and racks in-house. To ensure fast data transfer, we equip nodes with 3.2 Tbit/s per host InfiniBand connection.
Keeping it cool
Our servers are designed to operate with inlet air of up to +40°C. They are cooled just with outdoor air under a free-cooling concept, which saves lots of energy and brings the efficiency up.
Own data center in Finland
Own data center in Finland
We filmed this video 60 kilometers from Helsinki, the home of the first Nebius data center. The facility is packed with modern systems in every aspect of compute, storage, and data processing. Here, we constructed our supercomputer and a supercluster of thousands GPUs.
Second region just launched in France
Second region just launched in France
Nebius’ Paris data center is a colocation based at Equinix’s PA10 campus in the Saint-Denis district of Paris. The facility will be one of the first anywhere to adopt NVIDIA H200 GPUs.
In the photo, you can see an urban farm on the data center’s roof, heated by waste heat from the servers.
Third region in the US
Third region in the US
Our colocation in the Kansas City data center owned by Patmos is due to go live in Q1 2025. Patmos recently repurposed the facility, converting the iconic Kansas City Star printing press to a modern AI DC. The installation can be expanded from an initial 5 MW up to 40 MW, or about 35 thousand GPUs — NVIDIA Blackwells and Hopper H200s — at full potential capacity.
Dedicated server nodes for training and inference, all designed in-house by Nebius
Training AI models is a data-intensive process, with significant input and output. Our dedicated server node features the SXM5 GPU board with eight NVIDIA H100s allowing the massive throughput, in which speedy data transfer is key. The aforementioned 3.2 Tbit/s are achieved through using eight InfiniBand network cards, 400 GB/s each.
After the training, comes inference. Once a model has been stored in the GPU’s memory, it’s all set for smaller tasks like generating text or images based on prompts. Our inference-ready solution accommodates up to four dual-slot air-cooled PCI Express 5.0 GPUs.
Using these solutions, we built ISEG, a top-19 supercomputer in the world
Using these solutions, we built ISEG, a top-19 supercomputer in the world
We weren’t aiming to create a supercomputer. Yet our R&D team decided to test a part of the platform which was free of customers’ workloads at that moment. For that, they used a benchmark from the Top 500.
ISEG is now 19th in the world’s ranking — and 4th in Europe.
