Nebius Group (Nasdaq: NBIS) signed a multi-billion dollar agreement to deliver dedicated GPU capacity to Microsoft from a new data center in Vineland, New Jersey, with services beginning later this year. A 6-K filing pegs the base value around $17.4 billion through 2031, expandable to ~$19.4 billion if Microsoft purchases additional services.
The Vineland facility forms part of Nebius’s U.S. build-out, which targets up to 300 MW at the New Jersey site and adds capacity in Kansas City and Iceland. Nebius has also been rolling out H200 and Blackwell-class GPU clusters in Europe, including Paris (Equinix PA10) and a forthcoming UK deployment.
Nebius plans to finance associated capex using cash flows from the contract and debt secured against the agreement; obligations commence upon confirming required financing. The announcement follows broader industry momentum for “neocloud” providers—specialist AI clouds that scale GPUaaS via contract- and asset-backed financing.
- Contract scope: Five-year agreement to deploy GPU services in tranches during 2025–2026; base value ~$17.4 B, with upside to ~$19.4 B.
- Location & start: Dedicated capacity from Nebius’s Vineland, New Jersey data center; service begins later in 2025.
- Scale pipeline: New Jersey design capacity up to 300 MW; additional ramp in Kansas City (up to ~40 MW) and Iceland; European clusters in Paris and Finland.
- GPU roadmap: H200 availability in Paris, UK Blackwell Ultra (B300) cluster targeted for Q4 2025.
- Sustainability: Finland site recovers ~20,000 MWh/year of heat for district heating; PUE as low as 1.1 at high IT loads.
“Nebius’s core AI cloud business, serving customers from AI startups to enterprises, is performing exceptionally well… I’m happy to announce the first of these contracts, and I believe there are more to come.” (Arkady Volozh, Founder & CEO)
🌐 Analysis
Nebius sits in the fast-growing “neocloud” segment alongside CoreWeave, Lambda, and Crusoe—providers that focus on GPU-dense clusters, bare-metal orchestration, and AI-centric services.
Nebius Group N.V., a leading provider of AI-centric cloud infrastructure, has rapidly evolved from its origins as Yandex’s international cloud business into a standalone entity focused on delivering high-performance computing (HPC) resources optimized for artificial intelligence workloads. As of September 2025, Nebius operates a network of data centers strategically positioned across Europe and North America to minimize latency for global AI developers and enterprises. Its flagship facility is located in Mäntsälä, Finland, approximately 60 kilometers from Helsinki, which currently supports the ISEG supercomputer—ranked 19th globally on the TOP500 list—and hosts thousands of NVIDIA GPUs in custom-designed servers and racks. This data center is engineered for energy efficiency and sustainability, featuring advanced cooling systems and a current capacity that is set to expand significantly. In the United States, Nebius has established its first availability zone in Kansas City, Missouri, through a colocation partnership, starting with a 5 MW NVIDIA GPU cluster that can scale to 40 MW, accommodating up to 35,000 GPUs at full build-out. Additionally, a second U.S. site in Vineland, New Jersey, is under development as a build-to-suit facility spanning six 220,000-square-foot buildings, designed for up to 300 MW of IT load to handle dense AI training and inference tasks. In Europe, beyond Finland, Nebius maintains a presence in Paris, France, via Equinix’s PA10 campus in Saint-Denis, supporting initial GPU clusters for low-latency European access, and has recently launched operations in the United Kingdom with deployments in London-area facilities. These locations are selected for their proximity to major internet exchange points, robust power grids, and access to renewable energy sources, ensuring high availability (99.99% SLA) and compliance with GDPR and other regional data sovereignty standards.
At the hardware level, Nebius’s infrastructure is built around a reference architecture co-developed with NVIDIA, emphasizing scalability, interconnectivity, and performance for large-scale AI models. Core compute nodes feature NVIDIA’s latest GPU offerings, including the GB200 NVL72 and HGX B200 for next-generation Blackwell architecture, H200 and H100 SXM modules with up to 141 GB HBM3e memory per GPU, and L40S for inference workloads, all paired with Intel Xeon processors such as Emerald Rapids and Sapphire Rapids for CPU acceleration. Servers are in-house designed by Nebius engineers to optimize airflow, power density (up to 100 kW per rack), and thermal management, incorporating liquid cooling for high-TDP components to achieve PUE ratings below 1.2 in its Finnish facility. Networking is powered by NVIDIA InfiniBand fabrics delivering up to 3.2 Tbit/s per host via Quantum-2 or Spectrum-X Ethernet switches, enabling low-latency RDMA over Converged Ethernet (RoCE) for distributed training across clusters of thousands of GPUs. Key suppliers include NVIDIA as the primary GPU and networking partner—evidenced by Nebius’s status as a Reference Platform Cloud Partner—Intel for CPU and storage controllers, and colocation providers like Equinix for edge deployments in Paris and Patmos Computer for the Kansas City site. For the New Jersey expansion, Nebius has partnered with DataOne (a subsidiary of BSO) for construction, leveraging their expertise in hyperscale builds to integrate custom power distribution units (PDUs) and backup generators capable of supporting 300 MW with N+1 redundancy. Storage solutions draw from enterprise-grade NVMe SSDs and distributed file systems like Ceph, ensuring petabyte-scale durability for AI datasets.
Nebius’s announced expansion plans underscore its ambition to become a gigawatt-scale AI infrastructure powerhouse, with a $1 billion investment committed to Europe by mid-2025 and an additional $700 million in strategic equity financing secured in late 2024 to fund global build-outs. In Finland, the Mäntsälä data center will triple its capacity from 25 MW to 75 MW by early 2026, incorporating thousands more NVIDIA Blackwell GPUs to support exascale computing for AI research consortia. The U.S. footprint is accelerating with the New Jersey facility’s phased rollout—initial 100 MW online by summer 2025, scaling to 300 MW—alongside Kansas City’s progression to 40 MW, collectively adding over 300 MW of GPU-optimized capacity to tap into North American AI demand. Broader ambitions include signing letters of intent for two new European data centers (locations undisclosed but likely in Western Europe for redundancy) and deploying NVIDIA B300 Blackwell Ultra clusters in the UK to enhance cross-Atlantic hybrid cloud capabilities. CEO Arkady Volozh has publicly outlined a target of 1 GW total connected power by the end of 2026, achieved through greenfield developments and partnerships, aiming to house over 100,000 GPUs while prioritizing sustainable power sourcing (e.g., 100% renewable in Finland).
Compared with peers: CoreWeave has disclosed an $11.9 B, five-year OpenAI capacity deal and large Nvidia-collateralized debt programs; Lambda has raised equity and GPU-backed credit; Crusoe is building hyperscale campuses tied to Stargate/Oracle efforts. Nebius’s Microsoft contract is competitive in headline value and—importantly—explicitly conditions obligations on financing confirmation, reflecting tighter execution discipline after recent delivery-risk headlines elsewhere in the sector. If Nebius hits delivery milestones and brings Vineland online on schedule, it strengthens Microsoft’s diversified AI supply chain while giving Nebius contracted cash flows to accelerate 2026 growth.
