Nebius Group

Executive Summary

Nebius Group N.V. (Nasdaq: NBIS) is an AI infrastructure company that emerged from the $5.4 billion divestiture of Yandex's international assets in July 2024.¹ Led by Yandex co-founder Arkady Volozh, the company has rapidly pivoted from its search-engine heritage into a full-stack "neocloud" — a GPU cloud platform purpose-built for training and running large-scale AI models. In less than 18 months of independent operation, Nebius has grown revenue from $91.5 million (FY 2024) to $529.8 million (FY 2025), a 479% year-over-year increase, and exited 2025 with an annualized run-rate of $1.25 billion.²

The growth trajectory is underpinned by two landmark infrastructure deals: a $17.4–19.4 billion, five-year contract with Microsoft³ and a $3 billion, five-year agreement with Meta⁴ — together representing roughly $22 billion in committed revenue. These contracts effectively pre-fund the company's massive capital expenditure program: Nebius plans to spend $16–20 billion in 2026 alone to scale from 170 MW of active GPU compute to 800 MW–1 GW of connected capacity, with 3 GW of total contracted power across 16+ global sites.²

Source: Nebius Group Q4 2025 Letter to Shareholders

The Business

Nebius operates a vertically integrated AI infrastructure stack. Unlike hyperscalers (AWS, Azure, GCP) that offer AI compute as one service among thousands, Nebius's entire platform is optimized for a single purpose: high-performance AI training and inference. The company designs its own server racks, deploys InfiniBand networking at up to 3.2 Tb/s, and provides managed software layers (Kubernetes, Slurm, MLflow) tuned specifically for multi-thousand-GPU workloads. The core compute offering runs on NVIDIA H100 and H200 GPUs, with next-generation GB200 NVL72 systems on the roadmap.⁵

The customer value proposition is straightforward: AI labs and enterprises need massive, reliable GPU capacity at competitive prices, and the hyperscalers either can't deliver enough of it fast enough, or bundle it with services the customer doesn't need. Nebius fills that gap — purpose-built infrastructure, competitive pricing (H100s at $2.00–3.06/hour vs. industry standard $2.49+), and the ability to scale to thousands of interconnected GPUs for frontier model training.⁶

Financial Inflection

Q4 2025 marked a critical milestone: Nebius turned adjusted EBITDA positive for the first time, posting $15.0 million on $227.7 million in revenue.² The core AI infrastructure business achieved a 24% adjusted EBITDA margin in the quarter. Management's 2026 guidance projects $3.0–3.4 billion in revenue with approximately 40% adjusted EBITDA margins — implying $1.2–1.4 billion in adjusted EBITDA, a transformation from the $64.9 million full-year loss in 2025.²

Source: Nebius Group Q4 2025 Letter to Shareholders

From Yandex to Nebius

The Yandex Era

The story begins in 1997, when Arkady Volozh — a Kazakhstan-born mathematician — co-founded Yandex with geophysicist Ilya Segalovich. Over the next two decades, Yandex became Russia's dominant technology platform: search engine, ride-hailing, food delivery, e-commerce, cloud computing, and an autonomous driving program that rivaled Waymo. At its peak, Yandex commanded a $30 billion market capitalization and employed over 20,000 people across Russia and internationally.⁷

Volozh built Yandex into something unusual for a Russian company: a genuinely world-class engineering organization. The company's search algorithms competed head-to-head with Google in the Russian-language market and won. Its self-driving program, which eventually became Avride, accumulated millions of autonomous miles. Its cloud infrastructure team developed expertise in building and operating large-scale GPU clusters — skills that would prove pivotal for what came next.⁸

Arkady Volozh, Founder and CEO of Nebius Group. Source: Wikipedia

The Split

Russia's invasion of Ukraine in February 2022 triggered a cascade of consequences for Yandex. Western sanctions targeted Russian technology companies. Volozh, personally sanctioned by the EU in June 2022, stepped down as CEO. He had already left Russia in 2014 and was living in Israel, but the sanctions made his continued leadership of a Nasdaq-listed company untenable.⁷

In August 2023, Volozh publicly condemned the invasion as "barbaric" — an extraordinary statement for a Russian tech billionaire. The EU lifted his sanctions in March 2024, clearing the path for what would become one of the most unusual corporate restructurings in recent tech history.⁷

In July 2024, Yandex completed a $5.4 billion transaction that split the company in two. The Russian domestic operations — search, ride-hailing, e-commerce, everything that touched Russian consumers — were sold to a consortium of local investors. The international assets were retained by the Dutch-domiciled holding company, renamed Nebius Group N.V., with Volozh returning as CEO.¹

What Nebius Inherited

The split left Nebius with three critical assets:

1. A world-class engineering team. Approximately 1,300 engineers, many of whom had built Yandex's cloud infrastructure and AI research division. These weren't new hires — they were people who had been operating GPU clusters at scale for years, building search engines, training machine learning models, and developing autonomous driving systems.⁸
2. A data center in Finland. Yandex's Mäntsälä facility, originally built as a disaster recovery and European operations hub, became Nebius's first production data center — and the foundation for its AI cloud platform.⁸
3. $2.5 billion in cash. The proceeds from the Russian asset sale provided runway to transform from a search-engine holding company into a pure-play AI infrastructure provider.⁸

Nasdaq trading resumed in October 2024 after a suspension of more than two years. By December, Nebius had closed a $700 million strategic investment from NVIDIA, Accel, and Orbis Investments⁹ — a powerful signal that the market's most important GPU supplier saw Nebius as a serious infrastructure partner, not just another customer.

Product & Platform

The Full-Stack Neocloud

Nebius positions itself as a "full-stack" AI infrastructure provider — meaning it controls the entire vertical from hardware design to the software layer that customers interact with. This is a deliberate architectural choice. Unlike hyperscalers that adapt general-purpose cloud infrastructure for AI workloads, or colocation providers that simply rent out space with power, Nebius designs every layer specifically for large-scale AI compute.⁵

The stack breaks down into three layers:

Hardware layer. Nebius designs proprietary server racks optimized for GPU density and thermal management. The company deploys NVIDIA H100 and H200 Tensor Core GPUs interconnected via InfiniBand networking at up to 3.2 Tb/s — critical for the all-to-all communication patterns that large model training requires. Next-generation NVIDIA GB200 NVL72 systems (Blackwell architecture) are on the near-term roadmap. The InfiniBand fabric is a key differentiator: it enables clusters of thousands of GPUs to function as a single coherent compute resource, a requirement for training frontier AI models.⁵

Platform layer. On top of the hardware, Nebius provides AI Cloud (branded Aether 3.0/3.1) — a managed cloud environment with pre-configured virtual machines loaded with standard AI development libraries, managed Apache Spark for data processing, MLflow for experiment tracking, and high-performance distributed storage delivering up to 100 GB/s throughput. The platform supports both Kubernetes and Slurm workload managers, accommodating teams that want containerized microservice architectures and those running traditional HPC-style training jobs.⁵

Application layer. Token Factory, launched in Q4 2025, is Nebius's production-scale inference and post-training platform.² It targets the rapidly growing segment of AI companies that have finished initial model training and need optimized infrastructure for serving models at scale — fine-tuning, reinforcement learning from human feedback (RLHF), and low-latency inference serving. This is where Nebius aims to capture customers beyond the initial training phase and into ongoing production workloads, increasing lifetime value.

Source: Nebius

Why Not Just Use AWS?

The natural question is why companies would choose Nebius over established hyperscalers. Three factors drive the neocloud value proposition:

1. Availability. GPU capacity at the major clouds has been chronically constrained since late 2022. Waitlists for H100 clusters on AWS, Azure, and GCP have stretched months. Neoclouds like Nebius secured GPU supply through direct NVIDIA relationships and dedicated their entire capacity to AI workloads, making clusters available faster.
2. Price. Nebius offers H100 compute at $2.00–3.06 per GPU-hour (reserved to on-demand), compared to $2.49+ at CoreWeave and significantly more at hyperscalers when accounting for networking and storage costs. For a customer running 1,000 H100s continuously, even a $0.50/hour difference translates to $4.4 million annually.⁶
3. Simplicity. Hyperscalers bundle AI compute with hundreds of other services — databases, analytics, serverless functions, identity management — and their pricing reflects that complexity. Nebius strips away everything except what AI workloads need: raw GPU compute, fast interconnects, high-throughput storage, and the minimal orchestration layer to manage jobs.

Serverless AI Inference

Beyond dedicated GPU clusters, Nebius offers serverless AI endpoints — API-accessible inference serving where customers pay per token or per request rather than reserving GPU capacity. This targets the long tail of AI applications: companies deploying models that don't need dedicated infrastructure but need reliable, low-latency inference at variable scale.⁵

Infrastructure Footprint

From One Data Center to a Global Network

When Nebius emerged from the Yandex split in mid-2024, it had a single operational data center: the Mäntsälä facility in Finland. Eighteen months later, the company has secured 16 sites globally and is executing one of the most aggressive data center buildout programs in the industry.²

Source: Nebius Q4 2025 Shareholder Letter, company press releases², ¹⁰

The Independence, MO Mega-Campus

The Independence, Missouri facility at the Eastgate Commerce Center is Nebius's flagship U.S. site: approved for up to 1.2 GW across ten buildings on 400 acres — Nebius's first gigawatt-scale AI factory.¹⁰ To put that in context, 1.2 GW is roughly the power output of a nuclear plant — enough to run well over 100,000 GPUs simultaneously at full utilization. The campus is being developed in phases, with Phase 1 already operational and Phase 2 targeting mid-2025 completion.

The Power Constraint

The single biggest bottleneck in AI infrastructure is not GPU supply — it's power. NVIDIA can ship chips faster than data centers can secure grid connections, environmental permits, and utility contracts. Nebius's strategy of acquiring land and power contracts aggressively, in locations where grid capacity is available (rural Missouri, New Jersey industrial zones, geothermal-powered Iceland), is a direct response to this constraint.²

Contracted power guidance for 2026 was raised from 2 GW to 3 GW, with 800 MW–1 GW targeted to be connected and operational by year-end.² The delta between contracted power (3 GW) and connected capacity (up to 1 GW) reflects the reality that securing power contracts runs 12–24 months ahead of actually bringing facilities online.

Hybrid Deployment Strategy

Nebius uses three models for infrastructure deployment:

1. Greenfield owned sites (Independence MO, Oklahoma, Alabama, Minnesota) — highest control and margins, longest time to deploy
2. Build-to-suit (New Jersey) — a developer builds to Nebius's specifications, reducing upfront capital but locking in long-term leases
3. Colocation (Iceland, some early deployments) — fastest to market, lowest capital outlay, but lower margins and less control over the physical environment

This hybrid approach lets Nebius serve customer demand quickly via colocation while building owned capacity that delivers better unit economics over time.

Market Opportunity

The AI Infrastructure Spending Boom

AI infrastructure spending hit a record $86 billion in a single quarter (Q3 2025) according to IDC, with full-year 2025 tracking toward $334 billion. IDC projects this will reach $902 billion annually by 2029, with growth sustained above 30% through 2027.¹¹

Goldman Sachs estimates that total cloud revenues will reach $2 trillion by 2030 at a 22% CAGR, with IaaS alone reaching $580 billion. Hyperscaler AI capex is projected at $527 billion in 2026 alone — with Microsoft, Amazon, Google, Meta, and Oracle collectively committing more than half a trillion dollars in a single year to data center buildout.¹²

Gartner forecasts AI-optimized IaaS spending specifically will hit $37.5 billion in 2026, more than doubling from $18.3 billion in 2025 — with inference workloads ($20.6B) overtaking training as the primary demand driver.¹³

The Neocloud Segment

Within this broader AI infrastructure market, a distinct category has emerged: "neoclouds" — specialized cloud providers built exclusively for AI workloads. McKinsey has documented how these players have grown rapidly by offering purpose-built GPU infrastructure at prices up to 85% lower than traditional hyperscalers, exploiting a structural gap between AI compute demand and hyperscaler supply capacity.¹¹

The five largest neoclouds — CoreWeave, Nebius, IREN, Lambda Labs, and Crusoe — collectively represent the bulk of this market. Their growth is being driven by a structural mismatch: demand for AI compute is expanding faster than hyperscalers can add dedicated GPU capacity, creating a window of opportunity for purpose-built providers.

The Opportunity for Neoclouds

The dynamic is analogous to the early cloud market in 2008–2012: AWS had a massive head start, but the market was growing fast enough that Azure, GCP, and others could build multi-billion-dollar businesses without needing to take share from Amazon. Neoclouds are in a similar position today — there's simply more demand for dedicated AI compute than any single provider can serve.

Market Sizing Context

Competitive Landscape

The Neocloud Peers

Nebius competes primarily with four other neoclouds, each with distinct positioning:

CoreWeave is the largest and most established neocloud, having IPO'd in early 2025. It is an NVIDIA Elite Partner with Kubernetes-native infrastructure, strong enterprise compliance (SOC2, HIPAA), and a focus on large-scale (100+ GPU) training workloads. CoreWeave has secured its own landmark contracts, including with Microsoft, and recently went public at a valuation north of $20 billion. Its advantage is maturity and enterprise-readiness; its challenge is capital intensity similar to Nebius's.⁶

IREN (formerly Iris Energy) is a former Bitcoin miner that pivoted aggressively into AI cloud infrastructure, anchored by a $9.7 billion five-year contract with Microsoft. IREN has secured 4.5 GW of grid-connected power across Texas, British Columbia, and Oklahoma, with 150,000 NVIDIA GPUs on order (including 50,000 B300s). Its infrastructure runs on 100% renewable energy at ~$0.033/kWh average power cost and features 3.2 Tb/s InfiniBand networking built to NVIDIA reference architecture. IREN's advantage is its massive, low-cost renewable power pipeline; its challenge is executing the transition from mining revenue to AI cloud revenue while managing significant dilution.

Lambda Labs targets a different segment: startups, researchers, and mid-scale AI teams. Lambda offers a broader variety of GPU types (including B200, A100, and consumer-grade options), pre-configured ML software stacks (Lambda Stack), and startup credits of $5K–$25K. The platform is optimized for 1–8 GPU experiments and quick prototyping rather than thousand-GPU training runs.⁶

Crusoe Energy differentiates on sustainability, using stranded natural gas and renewable energy sources to power GPU clusters. Crusoe targets environmentally conscious customers and benefits from lower energy costs, but operates at smaller scale than CoreWeave or Nebius.

Neocloud Comparison

The Hyperscaler Threat

AWS, Azure, and GCP are not standing still. Each is investing tens of billions in GPU capacity, developing custom AI accelerators (AWS Trainium, Google TPUs), and offering increasingly competitive pricing. Microsoft's $17.4 billion deal with Nebius is notable precisely because it comes from a company that operates one of the world's largest cloud platforms — the implication being that even Azure's internal capacity is insufficient for its AI demand.³

The hyperscaler risk is that as GPU supply normalizes (expected by late 2026–2027), the availability advantage that neoclouds enjoy today will erode. The pricing advantage could follow as hyperscalers achieve greater GPU purchasing scale and amortize infrastructure costs across a larger revenue base. Neoclouds will need to compete on performance optimization, customer service, and switching costs rather than simply having GPUs available.

Nebius's Competitive Moat

Nebius's defensibility rests on four pillars:

1. Vertical integration. Custom server rack design, proprietary firmware, and InfiniBand networking optimizations deliver measurably better GPU utilization than commodity configurations. Higher utilization means lower effective cost per training run.
2. NVIDIA relationship. NVIDIA has now invested a cumulative $2.7 billion in Nebius — $700 million in December 2024 and an additional $2 billion announced March 11, 2026 as part of a deepened strategic partnership targeting 5+ GW of NVIDIA systems deployed by 2030. The deal includes early access to the NVIDIA Rubin platform, Vera CPUs, and BlueField storage systems, plus joint development on AI factory architecture and inference optimization. This is the deepest NVIDIA partnership outside of a hyperscaler.⁹¹⁸
3. Power contracts. 3 GW of contracted power capacity represents a 3–5 year head start over any new entrant that would need to acquire land, secure utility agreements, and obtain permits.
4. Engineering pedigree. The team that built Yandex's search infrastructure and cloud platform brings operational expertise that is difficult to hire for on the open market.

Business Model & Financials

Revenue Model

Nebius generates revenue through three streams, all centered on providing GPU compute infrastructure to AI companies and enterprises:

GPU-as-a-Service (GPUaaS) is the core business. Customers reserve GPU clusters on-demand or via long-term contracts (1–5 years). Pricing is per GPU-hour, with discounts of up to 45% for reserved capacity commitments. This is the bulk of the $214 million in core cloud revenue generated in Q4 2025 (+830% YoY).²

AI Lifecycle Management includes managed Kubernetes clusters, Slurm workload management, MLflow experiment tracking, and serverless inference endpoints. These services generate recurring revenue on top of base compute and increase switching costs as customers integrate Nebius tooling into their ML workflows.

Custom Infrastructure Solutions cover build-to-suit deployments for large customers who need dedicated, isolated GPU clusters. The Microsoft and Meta contracts likely fall into this category — bespoke infrastructure arrangements that command premium pricing in exchange for guaranteed capacity and performance SLAs.

Quarterly Revenue Progression

Source: Nebius Q4 2025 Earnings Release, Q1 2025 Earnings Release², ¹⁴

Profitability Trajectory

Nebius is in the classic infrastructure scale-up phase: revenue is growing faster than costs, but the GAAP P&L is dominated by depreciation of GPU assets, stock-based compensation, and one-time costs associated with the Yandex separation.

The Q4 2025 adjusted EBITDA turn to positive territory is significant. It demonstrates that the core AI infrastructure business has unit economics that work at scale — the question is whether those economics hold as the company deploys capital 10x faster in 2026.²

2026 Guidance

The 2026 guidance implies a revenue trajectory that, if achieved, would make Nebius one of the fastest-scaling infrastructure companies in history. For context, CoreWeave — widely considered the neocloud leader — generated roughly similar full-year 2025 revenue. Nebius is guiding for $3+ billion in 2026, a step function that depends on bringing new data center capacity online on schedule and filling it with paying customers.²

Traction & Key Contracts

The Microsoft Deal

In September 2025, Nebius signed a five-year infrastructure contract with Microsoft valued at $17.4 billion, with potential expansion to $19.4 billion.³ Microsoft will use Nebius's GPU infrastructure from the Vineland, New Jersey data center for large language model training and AI development — effectively outsourcing a portion of its Azure AI compute needs to a third-party neocloud.

This contract is remarkable for several reasons. Microsoft operates one of the world's three largest cloud platforms. It has direct relationships with NVIDIA, its own custom AI accelerator program (Maia), and data centers spanning dozens of countries. That it chose to contract with Nebius for this volume of compute suggests that internal capacity is genuinely insufficient — and that Nebius's full-stack offering delivers something Microsoft couldn't replicate internally on the required timeline.

The Meta Deal

Two months later, in November 2025, Meta signed a $3 billion, five-year agreement for AI infrastructure.⁴ Nebius disclosed that demand from Meta was strong enough that the contract had to be capped based on available capacity — suggesting Meta would have committed to a larger deal if infrastructure were available. Nebius planned to deploy the necessary GPU clusters within three months of signing.

Contract Economics

Combined, the Microsoft and Meta contracts represent approximately $22 billion in committed revenue over five years — roughly $4.4 billion annually if recognized evenly. This alone would exceed Nebius's 2026 revenue guidance of $3.0–3.4 billion, though revenue recognition timing depends on when capacity comes online and passes acceptance testing.

Critically, these contracts include customer prepayments that help fund Nebius's CapEx program. Management has stated that approximately 60% of 2026 planned CapEx ($16–20 billion) is already funded through internal cash, prepayments, and existing debt facilities.² This pre-funded model significantly reduces the execution risk of the infrastructure buildout compared to building on spec.

Beyond the Anchor Tenants

While Microsoft and Meta dominate the backlog, Nebius has been steadily building a broader customer base. The company grew from approximately 10 clients in 2023 to over 40 by mid-2025, and the Q4 2025 run-rate of $1.25 billion (versus contracted revenue of ~$4.4B annually from just two customers) suggests meaningful demand from the broader market as well.

Other Strategic Partnerships

NVIDIA. On March 11, 2026, NVIDIA and Nebius announced a major strategic partnership expansion. NVIDIA will invest an additional $2 billion in Nebius — on top of the $700 million from December 2024 — bringing total NVIDIA investment to $2.7 billion. The partnership targets deploying more than 5 gigawatts of NVIDIA systems across Nebius's platform by end of 2030, with collaboration spanning AI factory design, inference and agentic AI stack development, early adoption of the NVIDIA Rubin platform and Vera CPUs, and fleet health management. Jensen Huang called Nebius "an AI cloud designed for the agentic era, fully integrated from silicon to software." This is arguably the strongest NVIDIA endorsement of any neocloud.¹⁸

Accel. The participation of Accel — one of Silicon Valley's most established venture firms (early backer of Facebook, Spotify, Slack) — in both the $700 million December 2024 round and subsequent governance (Matt Weigand receiving board observer rights) provides institutional credibility.⁹

Capital Structure & Valuation

Funding History

Nebius has raised capital aggressively since the Yandex separation, reflecting the capital intensity of its infrastructure buildout:

Source: Company filings, press releases⁹, ¹⁵, ¹⁸

Share Structure

Nebius has approximately 205.8 million shares outstanding as of March 2026, down from roughly 235 million at the time of the Yandex split — a 12.7% reduction driven by share buybacks and cancellations prior to the December 2024 offering.¹⁵ However, the ATM program and convertible notes represent significant potential dilution: the 25 million ATM shares would increase the count by ~12%, and conversion of the $2 billion in convertible notes (depending on conversion price) could add further dilution.

Valuation

At approximately $23 billion market cap, Nebius trades at roughly 43x trailing twelve-month revenue (FY 2025: $529.8M). On 2026 guided revenue of $3.0–3.4 billion, the forward multiple compresses to approximately 7–8x — which, if the guidance is achieved, would be modest for a company growing at 500%+ with 40% EBITDA margins.²

The valuation story is binary. If Nebius executes on 2026 guidance, the stock is trading at a fraction of what fast-growing infrastructure companies typically command. If execution falters — data centers come online late, customers renegotiate terms, or demand softens — the $16–20 billion CapEx program becomes a liability rather than an asset, and the current market cap prices in far too much optimism.

Key Opportunities

Key Risks

Key Takeaways

Several dynamics will determine which narrative prevails in the next 12–18 months:

1. Capacity delivery velocity. Can Nebius bring 800 MW–1 GW of connected capacity online by year-end 2026? Every month of delay directly reduces revenue recognition.
2. Margin sustainability. The 24% adjusted EBITDA margin achieved in Q4 2025 (core AI) needs to scale to 40% on a much larger revenue base. Power costs, depreciation schedules, and pricing pressure from competitors will test this.
3. Customer diversification. Moving beyond Microsoft and Meta to a broader base of AI labs, enterprises, and government customers will reduce concentration risk and validate the platform's competitiveness in an open market.
4. Capital markets access. With $16–20 billion in CapEx planned and only ~60% funded, Nebius will likely need additional capital. Market conditions and stock price at the time of future raises will determine the dilution impact.
5. GPU supply normalization. When NVIDIA's production catches up with demand (likely mid-2027), the structural advantage that neoclouds enjoy today — simply having GPUs available — will diminish. Nebius needs to have built durable competitive advantages by then.

Nebius is, in essence, a bet on the thesis that AI infrastructure demand is so large and growing so fast that there is room for multiple major providers alongside the hyperscalers. The company's Yandex engineering heritage, NVIDIA backing, and landmark contracts provide a credible foundation. Execution will determine whether the foundation becomes a franchise.