Nebius AI GPU Infrastructure Deep Dive: European Cloud Provider, H100 Clusters, and Sovereign AI Strategy

Nebius AI emerged from the corporate restructuring of Yandex NV in 2024, when Yandex's international operations were spun off into a separate entity headquartered in Amsterdam. The Nebius group includes the former Yandex Cloud GPU infrastructure, Toloka AI (data labeling platform), and several AI research initiatives. The restructuring created a clean break: Nebius is a Netherlands-registered company with no Russian ownership or legal ties, positioning it as a GDPR-compliant, sovereign European cloud provider. This legal structure has been critical for winning European government and enterprise AI contracts that require data residency within the EU.

As of Q2 2026, Nebius operates approximately 8,000 GPUs, primarily H100 SXM (6,000 units) supplemented by A100-80G (1,500) and H200 (500). They have data centers in five European locations: Helsinki (Finland), Amsterdam (Netherlands), Frankfurt (Germany), Stockholm (Sweden), and a new facility in Tel Aviv (Israel). Total infrastructure investment since the spin-off is estimated at $750 million, funded by a combination of the Yandex separation proceeds and a $200 million debt facility from European banks.

Nebius offers three GPU service models. Managed Kubernetes clusters (similar to GKE but with GPU-optimized node pools and pre-installed NVIDIA GPU Operator) are the flagship product, with H100 nodes priced at $2.80-3.20 per GPU-hour on-demand. Bare metal servers (SSH root access to dedicated 8x H100 nodes) are priced at $2.20-2.80 per GPU-hour on 1-year commitments. And a serverless inference product (rebranded from Toloka AI's inference layer) offers per-token pricing for popular open-source models hosted on Nebius infrastructure.

All three models run on Nebius' custom cloud platform, which is built on OpenStack with Kubernetes for container orchestration. The platform provides native integrations with European compliance frameworks-GDPR data processing agreements signed by default, data residency guarantees enforced at the infrastructure level, and SOC 2 Type II certification obtained in 2025. Nebius also offers a Sovereign Cloud tier where customer data never leaves the country of origin, which has been particularly attractive to German financial institutions and French government AI projects.

Nebius has built its market strategy around the concept of sovereign AI-the idea that European countries need AI infrastructure that is legally and operationally independent of US hyperscalers. This message resonates strongly in the wake of the EU AI Act (effective August 2024) and growing regulatory scrutiny of US cloud providers' data handling. Nebius has signed GPU-as-a-Service contracts with two EU government agencies and three large European financial institutions, all of which cited data sovereignty as the primary purchasing criterion.

The sovereign AI positioning is reinforced by Nebius' technical architecture. Their cloud platform uses exclusively open-source components (OpenStack, Kubernetes, Ceph for storage), allowing customers to audit the full stack. The data centers are powered by 65 percent renewable energy (primarily hydroelectric from Finland and Sweden), with a target of 100 percent by 2027. Nebius also participates in European AI research initiatives, including the European Language Grid and the GAIA-X federation, further strengthening their position as the European cloud for AI.

In the European GPU cloud market, Nebius faces competition from two directions. The hyperscalers (AWS Frankfurt, GCP Frankfurt, Azure Netherlands) offer more comprehensive managed services but at higher GPU prices ($3.20-3.80/hr for H100 on-demand) and with US legal jurisdiction. CoreWeave's European expansion (London, Frankfurt, Amsterdam) provides the same Kubernetes-native experience at similar pricing ($2.75-3.25/hr) but with a US corporate entity. Nebius' differentiator is European legal structure and compliance: they are the only major GPU cloud that is both price-competitive with US alternatives and fully EU-governed.

The main constraint on Nebius' growth is GPU supply. As a newer entrant without the decades-long relationships that CoreWeave and hyperscalers have with NVIDIA, Nebius has faced longer lead times for H100 and H200 allocations. Their 8,000-GPU fleet size is a fraction of CoreWeave's 45,000 or AWS's estimated 150,000+ H100s. Until Nebius can secure larger allocation volumes, they remain a niche player suitable for compliance-sensitive workloads but not for companies needing large-scale GPU clusters (256+ GPUs per job). The planned expansion to 15,000 GPUs by end of 2026 and a potential B200 procurement deal in negotiation aim to close this gap.

Nebius' pricing varies by data center location due to differences in electricity costs and cooling requirements. The Helsinki data center, powered primarily by Nordic hydroelectricity at approximately $0.04/kWh, has the lowest operating costs and consequently the lowest prices: H100 on-demand in Helsinki is $2.70/hr versus $3.10/hr in Frankfurt (German power at $0.12/kWh, higher grid fees). Stockholm and Amsterdam fall in the middle at $2.85-2.90/hr. Tel Aviv matches Frankfurt pricing at $3.10/hr due to higher cooling costs in the Middle Eastern climate.

This regional pricing model creates interesting arbitrage opportunities for price-sensitive workloads. A training job running in Helsinki costs approximately 13 percent less than the same job in Frankfurt, with the same hardware and network connectivity. Nebius encourages this by providing a unified API across all regions-users can target the cheapest region for batch workloads without code changes. For latency-sensitive inference serving where proximity matters, users pay a premium for the closest regional node.