Nebius is not an opportunistic start-up that arrived to rent out GPUs at the peak of the AI cycle. It's a company born from a group that already knew how to operate complex infrastructures, design cloud architectures, optimize data centers, develop autonomous technologies and build critical software. The war in Ukraine shattered the old model (cf: the income statement evolution chart below, showing a trough in 2023 and a gradual recovery in business), but it also forced a radical reallocation of assets towards a market where scarcity commands a premium - that of AI computing. The investment thesis therefore lies in determining whether this metamorphosis can turn Nebius into something more than just another neo-cloud: an integrated platform capable of capturing a durable share of the value created by the industrialization of artificial intelligence.
A "forced" birth, but an unusually solid starting point
The story of Nebius begins with Yandex, a search engine launched in Russia in 1997, which evolved over time into one of the major tech groups in its domestic market. Yandex N.V., its Dutch holding company, was registered in 2007 and went public on the Nasdaq in 2011. The 2022 invasion of Ukraine brought this structure down. Trading was suspended, the group found itself caught in a web of regulatory, political, and financial constraints, with the holding company having to organize its exit from Russia.
The restructuring concluded in 2024 with the sale of Russian assets to Russian investors. What remains is Nebius Group N.V., an Amsterdam-based company refocused on AI infrastructure, led by Arkady Volozh and holding international assets. This point remains important, as the Russian past is still one of the negative arguments regularly associated with the case. However, the new group's separation from its former Russian operations, along with the trust placed in Nebius by major American partners, has reduced the operational weight of this argument. This is not to entirely dismiss perception risk, but to acknowledge that the company is now judged on its ability to deliver AI infrastructure at scale.
This origin gives Nebius an atypical profile. Many companies positioned in the AI cloud have to start from scratch: recruiting teams, finding sites, securing power, designing architectures, negotiating GPU access, building the software layer, and convincing clients. Nebius, by contrast, inherits a pool of engineers, cloud experience, an optimized data center in Finland, stakes in tech companies, and an execution culture built at Yandex. The group relies on over a thousand engineers from the former ecosystem, with skills covering hardware, software, AI, and data centers. In a sector where talent is one of the rarest resources after electricity and GPUs, this heritage constitutes a real advantage.
The neo-cloud: Selling the new AI bottleneck
Nebius belongs to the neo-cloud category. These players do not try to replicate the full model of generalist hyperscalers like Amazon, Microsoft, Google (Alphabet), or Oracle. Their role is more targeted: providing the critical infrastructure for AI workloads, i.e. GPUs, dense clusters, storage, networking, cooling, software orchestration, training and inference.
The distinction is fundamental. Hyperscalers possess colossal scale, but their offering covers an infinite range of computing uses. Neo-clouds, meanwhile, specialize in the most constrained needs of AI. They do not win because they are larger, but because they can be faster, more specialized, and better suited to clients needing immediately available capacity.
The true bottleneck of AI is not limited to the Nvidia chip. It lies in usable capacity: land, electricity, data centers, cooling, interconnection, equipment, software, engineering, and speed to production. It takes months, sometimes longer, to transform a capacity promise into operational, revenue-generating clusters. As long as demand for compute exceeds available supply, players capable of securing and activating this capacity hold significant economic power.
Nebius is positioned precisely at this point of tension. The company indicates that its contracted capacity now exceeds 3.5 GW and has raised its target to over 4 GW by the end of 2026. It has secured new sites, including a 1.2 GW AI factory in Pennsylvania, another 1.2 GW site in Missouri, and a Finnish site of over 300 MW. The company also anticipates 800 MW to 1 GW of connected power by the end of the year. In this industry, these figures are not mere real estate announcements: they directly dictate the future revenue trajectory.
A simple business model, but complex execution
The business model of Nebius is primarily based on selling compute capacity and AI cloud services. The company provides developers, start-ups, enterprises, and labs with the resources needed to train, fine-tune, deploy and run models. Clients consume GPU capacity, reserve resources, access managed platforms or sign longer-term contracts. The company can thus generate revenue through usage, capacity reservation, contractual commitments, and associated software services.
The advantage of the model is its clarity. AI demand translates into compute needs; compute translates into capacity utilization; utilization translates into revenue. However, this apparent simplicity masks considerable operational complexity. Nebius must finance GPUs, build or lease sites, guarantee performance, optimize energy consumption, maintain clusters, manage clients, and secure supplies. Success therefore depends not just on demand volume, but on the ability to deliver at the right cost, on time, with a high utilization rate.
Q1 2026 figures show the model is beginning to produce significant leverage. Group revenue reached $399m, up 684% year-on-year. AI cloud accounts for $389.7m, or approximately 98% (almost all) of total revenue, with a staggering 841% year-on-year growth and 82% quarter-on-quarter growth. ARR reached $1.92bn at the end of March 2026, compared to $1.25bn at the end of December 2025. The company now targets $3bn to $3.4bn in revenue for 2026 and $7bn to $9bn in ARR.
Adjusted profitability is also reaching a new dimension. The group's adjusted EBITDA for reached $129.5m in Q1 2026, compared to a loss of $53.7m a year earlier. The AI cloud business generates $174m in adjusted EBITDA, representing an adjusted EBITDA margin of 45%. Cost of revenue fell from 49% of revenue to 26%, product development expenses from 72% to 17%, and SG&A from 120% to 36%. This improvement illustrates the potential scale effect: once capacity is installed and utilized, revenue can grow faster than certain costs.
Facing CoreWeave: The battle for vertical integration
The comparison with CoreWeave is natural. Both companies embody the rise of neo-clouds, driven by GPU capacity scarcity and the explosion of AI workloads. But Nebius wants to distinguish itself by avoiding becoming a mere provider of raw capacity. The ambition is to be a full-stack platform: data centers, servers, optimized hardware, cloud software, inference tools, agentic services, and enterprise solutions.
This vertical integration may be advantageous if it allows for a better cost per compute unit, better performance per GPU, higher service quality, and features better suited to clients. The Finnish data center illustrates this philosophy: using free cooling, heat recovery, energy optimization, and a technical design intended to limit operating costs. The company is also developing in-house designed servers, optimized hardware configurations, and energy efficiency superior to more standardized solutions.
This logic extends into the software layer. Aether 3.5 enhances the platform with serverless AI, large-scale data transfer, security features, governance, billing, and auditing. Token Factory aims to facilitate the deployment and optimization of open-source or custom models in production. The goal is to reduce clients' operational costs, improve cost per token, and make Nebius less exposed to the commoditization of GPU rentals.
This is a central point for the investment thesis. If the GPU cloud market becomes progressively more competitive, players selling only capacity risk seeing their prices under pressure. Those who control more layers, improve client productivity, and lower the total cost of ownership can defend better profitability. Nebius is clearly trying to position itself in this second category.
Token Factory, Aether 3.5, and the full-stack ambition
Moving up the software stack has become one of Nebius's most important strategic axes. The company does not just want to provide clients with GPU access, but to support them across the entire AI lifecycle: training, fine-tuning, inference, orchestration, agentic search, optimization, governance, and deployment.
Token Factory is particularly strategic. Open-source models are progressing rapidly and are increasingly used by companies wanting to avoid complete dependence on proprietary models. But moving from an experimental model to reliable production inference remains difficult. It requires optimizing costs, latency, quality, scalability, and the underlying infrastructure. Token Factory addresses this need by offering a managed inference platform. Revolut uses Token Factory to automate a significant portion of support conversations, while monday.com chose it to power its AI features.
Recent acquisitions reinforce this strategy. Tavily brings agentic search capabilities. Eigen AI improves inference execution and model optimization. Clarifai brings system orchestration technologies, optimization capabilities, and research talent. These building blocks are not incidental: they progressively shift Nebius from the infrastructure layer toward a more complete AI platform.
This evolution also answers the question of the five-year business model. Can Nebius become a full-stack AI training and deployment provider rather than a simple GPU lessor? The company is clearly moving in this direction, notably through Toloka for data, Token Factory for inference, Tavily for agentic workflows, and its own cloud capabilities for compute. Data to measure the future contribution of this model is still lacking, but the strategic coherence is strong.
Physical AI: When AI compute leaves the screen
Physical AI constitutes one of the most interesting growth drivers in the case. It refers to AI applications that interact with the real world: robots, autonomous vehicles, industrial systems, physical simulation, world models, synthetic data, and embedded inference. These uses require vast amounts of compute to train models, simulate environments, test behaviors, and deploy reliable systems.
Nebius is working with Nvidia on an end-to-end platform covering the robotics lifecycle, from simulation to training and then real-world deployment. The company mentions clients like 1X Technologies, which develops general-purpose robots, and Rhoda, which works on robotic intelligence using vast multimodal datasets. For Nebius, this vertical is natural: it combines GPU needs, simulation, orchestration, inference, and data.
Physical AI can also support Nebius's move upmarket. A robotics client is not just looking for raw capacity; they need a complete environment capable of managing complex data, simulations, multimodal models, testing, and deployment. If Nebius succeeds in structuring a specialized offering for these uses, the company can strengthen its differentiation relative to more generalist providers.
Avride: Direct exposure to autonomous mobility
Avride is one of the group's most original businesses. Born from autonomy work developed during the Yandex era, it operates in autonomous vehicles and delivery robots. This business does not carry the primary thesis for Nebius, but it adds an interesting option on robotics and physical AI.
In Q1 2026, Avride more than doubled the size of its autonomous-capable vehicle fleet compared to the end of 2025. The company is continuing to expand its operational perimeter in Dallas and is improving its pre-assembly processes for electronic components, notably sensors installed on vehicles. It is also developing its R&D fleet to accumulate the data necessary to improve its AI stack and transition toward driverless operations.
The delivery robot segment appears more commercially advanced. Robotic deliveries grew 178% year-on-year in Q1, with over 174,000 deliveries in the quarter and over 500,000 cumulative deliveries since launch. Avride also launched operations in Philadelphia, linked to its multi-year partnership with Uber to deploy delivery robots and autonomous vehicles to Uber and Uber Eats customers in the United States.
The value of Avride is difficult to isolate due to a lack of detailed financial figures, but the asset reinforces the image of a group with technological depth beyond the cloud. It also gives Nebius a direct testing ground for the simulation, data, training, and inference needs of autonomous systems.
TripleTen: Tech edtech as a secondary asset
TripleTen is the group's edtech business. It offers online training in fields such as web development, software engineering, data science, quality assurance, business intelligence, cybersecurity, and AI. Its model relies on training programs, enrollment fees, coaching, practical projects, and B2B offerings.
In Q1 2026, TripleTen posted revenue growth of 10% year-on-year. This pace is obviously much more moderate than that of the AI cloud, but the business retains interest. It meets the demand for tech skills, integrates more AI content into its programs, and develops upskilling offers for companies. Markets in Latin America and Brazil are growing more steadily than the US, while the group places greater emphasis on efficiency and profitability.
TripleTen does not justify a bullish thesis on Nebius on its own. It should rather be seen as a secondary asset, useful for group diversification and potentially aligned with the rising demand for AI skills.
Toloka: Data as a reliability brick
Toloka provides data to developers of LLMs, GenAI, and AI agents. Its role consists of producing, selecting, annotating, evaluating, and structuring quality data for model training and improvement. In AI, data quality remains a major determinant of performance. Models are not differentiated solely by their architecture or the compute power used, but also by the examples, corrections, evaluations, and human judgments that guide their learning.
Nebius no longer holds voting control of Toloka since Q2 2025 but retains a significant economic stake. The planned integration of Tendem into the Nebius ecosystem is notable. Tendem allows for the insertion of verified human experts into agentic workflows and makes their judgment accessible via the Model Context Protocol. In other words, Toloka can provide a layer of human reliability to increasingly automated systems.
This building block reinforces the idea of a Nebius more complete than a simple AI cloud. With Toloka for data and evaluation, Token Factory for inference, Tavily for agentic search, and cloud infrastructure for compute, Nebius can potentially cover more needs in the AI production chain. It is not yet a major quantified driver for the group, but the strategic direction is consistent with enterprise demand: less technical complexity, more integrated solutions, and greater reliability.
ClickHouse: The hidden option that could become much more visible
ClickHouse is the most interesting financial asset outside the AI cloud. Nebius holds 28% of this company, according to its 20-F available on EDGAR. ClickHouse develops an open-source analytical database designed for real-time processing, low-latency queries, and analysis of massive data volumes. In an economy where AI generates ever more logs, pipelines, observability data, agentic queries, and analytical processing, this type of infrastructure sits very close to the fundamental data consumption layer.
ClickHouse raised $400m at a valuation of $15bn, implying a theoretical value of approximately $4.2bn for Nebius's 28% stake. Relative to a market capitalization of around $57bn for Nebius, this represents nearly 8% of the group's market value. It is therefore not a detail, even if the core of the thesis remains driven by the AI cloud. It should also be noted that this value is not entirely invisible in the accounts, as Nebius has already recorded a non-cash revaluation gain linked to its stake. However, a potential ClickHouse IPO would change the market reading: the asset would become listed, comparable, liquid, and much more directly integrable into a sum-of-the-parts valuation.
This is where the option becomes interesting. If ClickHouse goes public at a valuation higher than that of the last private round, Nebius's stake could take on a whole new importance. A doubling of ClickHouse's valuation would mechanically increase the theoretical value of Nebius's stake from approximately $4.2bn to $8.4bn, before any holding company discount or liquidity considerations. For a company financing a very capital-intensive expansion in AI factories, having a monetizable asset of this size can become a strategic lever: a potential source of funding, a signal of the quality of the portfolio inherited from Yandex, and a revaluation catalyst if the market begins to better distinguish the group's different pockets of value.
The fact that Situation Awareness has made it its top position reinforces market interest around this optionality. The fund reportedly placed Nebius at the top of its portfolio, with about 15% of assets under management. One possible reading is that it is not only playing the growth of the AI cloud, but also the materialization of certain adjacent catalysts, including the potential ClickHouse IPO. This point alone does not validate the investment thesis, but it signals that specialized investors might consider Nebius as a hybrid vehicle: direct exposure to the AI compute bottleneck, augmented by a revaluation option on a high-quality software asset.
Meta, Nvidia, and commercial validation
Two partnerships structure Nebius's recent news. The first is the agreement with Meta, worth up to $27bn over five years. It includes a first part of $12bn linked to the purchase of compute capacity starting in 2027, and a second part of $15bn that allows Nebius to sell capacity to Meta under predefined conditions or to other clients at market conditions. This structure provides visibility, supports growth financing, and validates Nebius's ability to serve very large clients.
The second is the strengthening of the partnership with Nvidia. Nvidia invested $2bn in Nebius, following an earlier stake. The collaboration covers hardware and software integration, AI factory design, inference, and agentic platforms around Token Factory. Nebius also indicates it has obtained Nvidia Exemplar Cloud status on GB300 NVL72 for training workloads, which bolsters its technical credibility.
These partnerships constitute strong signals. In a still-young market, the trust of strategic clients and suppliers can accelerate commercial credibility. But they also create dependencies: on Nvidia for equipment access, on a few large clients for contractual visibility, and on external financing for continued expansion.
An ambitious but risky capital allocation
Nebius has a strengthened balance sheet following its recent raises. In Q1 2026, the company secured $6.3bn in capital, including USD 2 Bn via Nvidia's investment and $4.3bn via convertible securities. It ended the quarter with $9.3bn in cash, including $2.3bn in positive operating cash flow for the quarter. This liquidity provides the means to accelerate capacity expansion.
But the model remains very capital-intensive. Q1 capex reached approximately $2.5bn, primarily for GPUs, cluster-related hardware, and data center expansion. The company indicates it will continue to invest heavily and may resort to several funding sources: cash generated by operations, client prepayments, debt, asset-backed financing, corporate debt, at-the-market programs, and other instruments.
The capital allocation challenge is therefore major. As long as new capacity finds strong demand at high prices with good utilization, the model can create significant value. But if timelines slip, if the cost of capital rises, if GPU rental prices fall, or if demand shifts, the capital intensity becomes a risk. Nebius must finance today capacity whose profitability will depend on tomorrow's market conditions.
Valuation: The opportunity is no longer obvious; it has become demanding
Valuation requires nuance. At $259 per share, and based on approximately 254 million shares in early June 2026, the implied market cap approaches $66bn. At this level, Nebius is no longer a company ignored by the market. The stock already bakes in massive growth, solid operational execution, the success of new sites, rising ARR, and some recognition of adjacent assets. Annualized Q1 revenue remains modest against this market cap, even if the ARR momentum and 2026 guidance are rapidly changing the group's scale. The bullish thesis therefore rests on Nebius's ability to grow the denominator faster than the market anticipates: more connected capacity, higher utilization, more clients, more inference, more software, more recurring revenue, and better profitability.
ClickHouse adds an extra dimension. If the market primarily values Nebius as a neo-cloud, the 28% stake in ClickHouse already represents significant theoretical value. An IPO could make this asset much more visible and trigger a partial re-rating of the case. But caution is warranted: a private stake, however promising, is not always fully valued in a market cap, and an IPO depends on market conditions, ClickHouse's growth, and investor appetite for cloud databases.
Key Risks
Beyond the valuation, which remains a real debate (given its explosive growth, the stock could be valued at 25x 2030 earnings, or 10x depending on the speed of business development), you are buying into a project well-positioned in a bottleneck, a sovereign necessity, a strategic asset for which demand should be structurally strong for the next decade.
The first risk is execution. Nebius must transform contracted capacity into connected capacity, then into active and billable capacity. Each step carries risks: permits, construction, power connection, GPU delivery, installation, cooling, reliability, scaling, and service quality. In such a capital-intensive model, delays are costly.
The second risk is financial. The company has a strong cash position, but its needs remain high. Debt, convertibles, client prepayments, and asset-backed financing can support growth, but they also increase financial complexity. Dilution risk is not zero, particularly if the company were to further accelerate its investments or if market conditions tightened.
The third risk is competitive. Hyperscalers may further specialize their AI offerings. CoreWeave and other neo-clouds are pursuing the same opportunity. Players from colocation, mining, or energy infrastructure are also seeking to monetize sites suitable for AI workloads. Current scarcity protects sector margins, but this protection could diminish if too much capacity arrives simultaneously.
The fourth risk is technological. If models become much more efficient, if inference consumes fewer resources per query, if new architectures reduce GPU needs, or if clients heavily optimize their consumption, capacity demand could remain high in volume but less favorable in price. The risk is not the end of AI, but the economic normalization of infrastructure.
The fifth risk is concentration. Large contracts, notably with Meta, provide visibility but also increase dependence on a few structural clients. Finally, the Yandex past may still weigh on the stock's perception, even if the group's new structure and Western partnerships strongly mitigate this argument.
A rare asset
The Nebius thesis does not rely on a single driver. The first remains AI cloud expansion, with sharply rising contracted capacity, structural partnerships, and demand still exceeding available supply. The second is the move up the software stack, with Token Factory, Aether 3.5, Tavily, Eigen AI, Clarifai, and the ambition to go beyond raw GPU rental. The third is the optional value of the portfolio inherited from Yandex, particularly ClickHouse. This 28% stake, theoretically valued around $4.2bn at the last private round, could become a much more visible catalyst in the event of an IPO, especially if the public valuation significantly exceeds the $15bn of the last funding round.
It is this mix that makes Nebius atypical. The market primarily values a hyper-growth neo-cloud, but the investor also buys a company equipped with software assets, strategic stakes, and growth drivers like Avride, Toloka, or TripleTen. The risk is that the current market cap already integrates a large part of this promise. The opportunity is that certain assets, notably ClickHouse, may not yet be fully appreciated in a simplified reading of the case centered solely on GPU capacity revenue.
Nebius can therefore be analyzed as an industrial hyper-growth stock at the intersection of cloud, AI, energy, and software. The company has the attributes of a potential winner: technical heritage, secured capacity, top-tier clients, Nvidia partnership, rising inference power, Physical AI exposure, and the ClickHouse option. But its market price demands near-perfect execution. At this stage, the case remains attractive for those who believe in a durable scarcity of AI compute and in Nebius's ability to convert its contracted gigawatts into profitable revenue. It becomes much more vulnerable if scarcity dissipates, if financing tightens, or if adjusted profitability does not translate into durable value creation.
Choose MarketScreener 
