Nebius: the quiet European compute play building a multi-site GPU empire

Nebius, the Amsterdam-based AI infrastructure company, published a detailed breakdown of its global data centre footprint this week, revealing a multi-site strategy that spans six facilities across four countries and two continents. The company published the page as part of a broader push to position itself as a serious alternative to the hyperscalers for AI compute.

The most striking detail: Nebius claims an initial Q4 2025 deployment of 4,000 NVIDIA Blackwell Ultra GPUs at its UK colocation site in Surrey, calling it “one of the first in Europe of this type of GPU.” That is a bold claim. NVIDIA only announced the Blackwell Ultra architecture at GTC 2025 in March, with general availability expected in the second half of 2025. If Nebius lands those chips in a European data centre by the end of this year, it would put the company ahead of most cloud providers on the continent.

The UK site sits at Ark Data Centres’ Longcross Park campus in Surrey, a facility described as “designed from day one for emerging AI workloads” with liquid-ready cooling and high-density power. Nebius says the colocation will serve UK startups, research institutes, enterprises, and public-sector organisations including NHS England. That last customer is notable: NHS England has been exploring AI for medical imaging, diagnostics, and administrative automation, but has been constrained by compute availability on domestic soil.

Nebius’ strategy is not UK-only. The company operates its own data centre in Finland, 60 kilometres from Helsinki, where it built a supercomputer and a “supercluster of thousands of GPUs.” That facility runs on Finland’s grid, which is among the cleanest in Europe, with over 40% of its electricity from nuclear and most of the rest from hydro and biomass.

In the United States, Nebius partnered with DataOne on a 300 MW greenfield site in New Jersey. The first phase is slated for summer 2025. The key innovation there is behind-the-meter electricity generation, meaning Nebius will generate its own power on site rather than pulling from the grid. That approach bypasses the interconnection queues that have delayed many US data centre projects by years. The site is currently a greenfield, shown in a render on the company’s page.

Missouri offers a different model. Nebius is colocating at a Patmos-owned data centre in Kansas City, in a building that was formerly the Kansas City Star printing press. Patmos repurposed the facility for AI workloads. The first phase starts at 5 MW but can scale to 40 MW, or roughly 35,000 GPUs at full build-out. Nebius says those GPUs will be a mix of NVIDIA Blackwells and Hopper H200s.

Iceland and France round out the portfolio. In Keflavik, Nebius is deploying a 10 MW cluster at a Verne-operated facility that runs entirely on Iceland’s 100% renewable hydroelectric and geothermal energy. Physical deployment and software installation are underway, with full operational availability expected by the end of March. In Paris, Nebius has a colocation at Equinix’s PA10 campus in Saint-Denis, which it says was among the first facilities in the world to adopt NVIDIA H200 GPUs. The Paris site features an urban farm on the roof, heated by waste heat from the servers.

The breadth of this footprint is unusual for a company that is not one of the big three cloud providers. Nebius is effectively building a distributed GPU fabric across six sites, each with different power profiles, regulatory environments, and customer bases. The Finland and Iceland sites offer cheap, clean power for training runs. The UK and France sites offer low-latency access to European enterprise and public-sector customers. The New Jersey site offers US mainland access with on-site power generation. The Missouri site offers a repurposed industrial building with room to scale.

The question is whether Nebius can execute. Building data centres is hard. Getting 4,000 Blackwell Ultra GPUs in Europe by Q4 2025 is harder. NVIDIA’s supply chain is constrained, and European demand for high-end AI accelerators is surging. Nebius will need to secure allocation from NVIDIA, build out the physical infrastructure at Longcross Park, and integrate the software stack to make those GPUs usable for customers.

Nebius has some advantages. The company was spun out of Yandex’s international operations in 2022, inheriting a team with deep experience in large-scale infrastructure. The company’s cloud platform, Nebius AI, already runs GPU instances for AI training and inference. The data centre page is a signal that Nebius is serious about owning the full stack, from silicon to site.

For AI builders in Europe, the Nebius footprint matters. European researchers and startups have long complained about the difficulty of accessing high-end GPU compute on the continent. AWS, Google Cloud, and Azure offer GPU instances in European regions, but supply is tight and prices are high. Nebius offers an alternative, especially for customers who want to run workloads on European soil for data sovereignty reasons.

The NHS England relationship is a test case. If Nebius can deliver Blackwell Ultra compute to a public-sector healthcare organisation at scale, it would validate the model. If the Q4 2025 deadline slips, the narrative shifts from ambitious to overpromising.

Nebius is not trying to compete with the hyperscalers on global scale. It is trying to win on specificity: purpose-built AI facilities, liquid cooling, behind-the-meter power, and a multi-site network that lets customers choose where their data lives. Whether that is enough to attract enough customers to fill 35,000 GPUs in Missouri and 4,000 Blackwells in Surrey is the open question.