Nebius Group

Nebius Group N.V. is the corporate successor to Yandex N.V., re-founded in substance in 2024 as a pure-play AI cloud infrastructure company following the divestment of its Russian operating businesses. It is today one of a small group of "neocloud" providers building purpose-built, multi-tenant GPU cloud infrastructure at gigawatt scale, and one of only two such providers listed on a U.S. exchange with a substantial European footprint.

Shares outstanding and headcount trend

The FY2023 headcount of 26,361 reflects the pre-divestment Yandex N.V. group and is not comparable; the FY2024 figure of 1,371 is the first post-divestment year-end count. Third-party workforce analytics (Revelio Labs) estimated approximately 1,815 employees as of March 2026, with roughly 28% of the workforce in the United States and 20% in the Netherlands — this is an estimate, not a company disclosure.

Capital markets snapshot (13 August 2026)

150-word positioning statement

Nebius Group is an Amsterdam-headquartered, Nasdaq-listed AI cloud company that designs, builds and operates full-stack GPU infrastructure — from data centre and rack architecture through networking and storage to a proprietary cloud control plane, a managed inference platform (Token Factory) and agentic tooling (Tavily). It is the principal European-domiciled challenger to the U.S. hyperscalers and to CoreWeave in the AI compute market, differentiated by in-house hardware and software engineering inherited from two decades of building large-scale platforms at Yandex, by a deliberately diversified customer base spanning frontier labs, AI-natives and enterprises, and by anchor contracts with Microsoft (up to $19.4 billion) and Meta (up to $30 billion across two agreements). Following the July 2024 divestment of its Russian businesses, the group also retains an autonomous-vehicle unit (Avride), an edtech platform (TripleTen), and minority stakes in ClickHouse and Toloka. Its central strategic wager is that securing power and capacity ahead of contracted demand is the binding constraint — and the durable moat — in AI infrastructure.

The company's own description

In the FY2025 Annual Report the company describes itself as follows: Nebius is "a global AI cloud platform" that "delivers a unified full-stack AI cloud that spans the complete AI journey — from compute capacity to software and services — that enable fast and efficient training and inference at scale." It states that it was "founded around deep in-house technological expertise" and offers "a comprehensive and integrated suite of AI and ML cloud solutions, including both hardware and software built in-house," enabling delivery of "high-performance GPU compute clusters, storage, managed services, and advanced tools for AI model training and inference at enterprise-scale." Management asserts that the group "offers one of the few global, at scale, multi-tenant clouds purpose built for AI, with a significant presence in Europe, the U.S., and other geographies around the world."

Independent characterisation

Nebius is best understood as a vertically integrated compute landlord that is progressively adding software rent. Three distinct economic layers sit inside what the company reports as a single AI cloud segment:

Layer one — capacity origination. The company secures land, power and shell capacity, either by acquiring and building its own sites ("AI factories," now the majority of contracted power) or by leasing colocation from third parties (DataOne in Vineland NJ, Patmos in Kansas City, Ark in London, Equinix in Paris, Verne in Keflavik, Mega Or in Israel, Vantage in South Wales). Since July 2026 a third mode exists: an asset-light partnership model in which a partner finances, builds and owns the facility and hardware while Nebius contributes its software stack, reference architecture and go-to-market demand in exchange for high-margin revenue with minimal balance-sheet capital.

Layer two — infrastructure monetisation. GPU capacity is sold either on-demand ("pay-as-you-go") or as fixed reserved-capacity contracts. As of Q2 2026 management characterises three contract archetypes: short-duration deals of three to six months priced at a substantial premium ($40–50 million annual contract value per megawatt); mid-term contracts of one to three years, described as the core business ($20–25 million per MW for deals closed in Q2 2026); and long-term contracts with investment-grade counterparties (Microsoft, Meta) which serve principally as collateral and financing enablers rather than as margin maximisers. The 2026 installed base is priced at roughly $12 million per MW, so the repricing dynamic is steep.

Layer three — software and value-added services. Token Factory (managed inference on open-weight models), Tavily (agentic search API), Nebius Echo (natural-language infrastructure agent), Serverless AI, storage, orchestration and the Aether cloud platform. Inference workloads on Token Factory more than tripled sequentially in Q2 2026. Strategically this layer is the answer to the bear case that neocloud economics are those of a leveraged real-estate business: software attached to partner-financed capacity produces revenue without corresponding capex.

Revenue model and mix

Revenue is overwhelmingly service revenue recognised over time as capacity is consumed or as reserved capacity is made available. There is no meaningful licensing revenue. TripleTen contributes consumer education revenue (bootcamps, project-based learning, and a growing B2B enterprise offering). Avride contributes de minimis revenue from robodelivery and robotaxi operations conducted largely through the Uber and Uber Eats platforms.

(Percentages are of total segment revenue before eliminations and therefore sum above 100.)

Customers and end-markets

Named or characterised customers disclosed across 2025–2026 filings and shareholder letters include Microsoft, Meta, Reflection, Cohere, an unnamed U.S. frontier AI "neolab," a large U.S. quantitative trading firm, AMI (co-founded by Yann LeCun), Basecamp Research, Prima Mente, Higgsfield, Sword Health, 1X Technologies, Rhoda, Core Automation, Logical Intelligence, Revolut, monday.com, Black Forest Labs, Cloudflare, Cognition, Recraft and Stanford University. Vertical go-to-market motions are organised around healthcare and life sciences, physical AI and robotics, media and entertainment, retail and commerce, financial services, and agents/copilots.

Value chain position

Nebius sits between the semiconductor supplier (NVIDIA, on which it is explicitly dependent per its own risk factors) and the model developer or AI application company. It does not build frontier models. It is increasingly moving downstream into the model-serving layer through Token Factory, Eigen AI (inference optimisation and post-training) and Clarifai (system-level inference and compute orchestration), and sideways into the agentic tooling layer through Tavily. Upstream it is moving into power procurement and generation (Bloom Energy behind-the-meter fuel cells; a privately financed 800MW generation facility supporting the Independence, Missouri campus).

Stated strategy — verbatim themes from the FY2025 Annual Report

The FY2025 Annual Report frames the business around ten "most significant factors," which function as the company's strategic pillars:

1. Competing effectively in a rapidly evolving market — "investing in the continuing development of a flexible, full-stack AI cloud solution that extends beyond bare-metal compute, and includes comprehensive software and value-added services."
2. Expanding our customer base — "AI is rapidly moving from research into large-scale production systems"; the aim is to "serve hundreds of customers, reduce concentration risk and become the platform of choice for next-gen businesses."
3. Scaling with our customers — "natively embedding this enterprise-grade security and functionality into our platform."
4. Building scalable and resilient data center capacity — "expanding our global data center footprint, building highly optimized and scalable infrastructure."
5. Leveraging our global footprint to capture new opportunities.
6. Securing power, components and talent — "our ability to navigate the evolving regulatory framework, dynamic supply chain, and constrained talent pool."
7. Access to capital to support our growth ambitions — noting "in 2025, we raised over $5 billion... In the first quarter of 2026, we raised more than $6 billion."
8. Expanding our talent base.
9. Integration of newly acquired businesses and assets — "we plan to continue to strategically acquire or invest in businesses that can expand or enhance our AI cloud platform through talent and/or technology."
10. Our other businesses also operate in high-growth sectors.

Volozh's Q2 2026 letter adds the sharpest strategic articulation to date: "We are not simply responding to where the industry stands today; we have the knowledge and experience to build the infrastructure, tools, and capabilities for where it will be tomorrow," and, on the earnings call, "We choose when to sell, to whom we sell, and on what terms, and how we finance everything."

Strategic initiatives announced in the last 24 months

Medium-term financial targets and guidance

All FY2026 guidance was reaffirmed on 12 August 2026. Management has also indicated a longer-run ambition of 20–30% EBIT margins for the AI cloud business, plans to deploy more than 1GW of new capacity per year starting in 2027, and expects nearly all 5GW of contracted power to be brought online within two to three and a half years. Formal 2027 guidance is scheduled to be provided later in 2026.

ESG and sustainability commitments

Covered in detail in Section 20. In summary: 100% renewable electricity sourcing already achieved in France, Iceland and the UK; Finland transitioning from 95% low-carbon to 100% renewable from mid-2026; 328MW of behind-the-meter fuel-cell capacity contracted; closed-loop liquid and air cooling designed to minimise water consumption; community investment programmes in STEM education, skilled-trades apprenticeships and first responders around U.S. sites.

5.1 Nebius AI cloud segment

Nebius AI Cloud (platform releases "Aether"). The proprietary, AI-native cloud control plane, rebuilt from the ground up and launched in Q4 2024. Versioning has been rapid: Aether 3.0 and 3.1 in 2025 (enterprise-grade security and functionality); Aether 3.5 in March 2026; Aether 3.6 in June 2026. Target customers span individual developers to hyperscale enterprises. Pricing is published at nebius.com/prices on both on-demand and reserved-capacity bases.

• Aether 3.5 (March 2026) introduced Serverless AI (on-demand inference without capacity planning), a Data Transfer Service enabling petabyte-scale cross-cloud data movement, improved Kubernetes Secrets integration, and expanded billing and audit-log export.
• Aether 3.6 (June 2026) introduced Nebius Echo (see below), a Key Management Service with customer-managed encryption keys, Workload Identity Federation for credential-free authentication, Budgets for FinOps spend limits, and Bring Your Own Image for hardened base images. Object storage read speed improved 30%; Shared Filesystem became 3x faster for small-file operations.

Compute. High-performance GPU clusters. Deployed generations include NVIDIA H100/H200, B200 (Israel launch, Q4 2025), B300 (UK launch, Q4 2025), GB300 NVL72 (on which Nebius achieved NVIDIA Exemplar Cloud status for training in Q1 2026, one of a small group of providers to hold the designation across multiple GPU generations), RTX PRO 6000 Blackwell Server Edition (added to Aether 3.5 for applied AI and simulation workloads), and Vera Rubin NVL72 — first systems received by Q2 2026 and under validation, with production availability planned as shipments scale. NVIDIA Vera CPUs are being adopted early to complement the GPU fleet for agentic orchestration, tool-calling and memory workloads.

Nebius Token Factory. Managed inference platform for open-weight and custom models, launched in Q4 2025 as the successor to AI Studio. Day-one support is offered for advanced open models including Kimi K3, GLM-5.2, MiniMax 3 and NVIDIA Nemotron 3 Ultra. In the Artificial Analysis Endpoint Accuracy Index, Nebius achieved full accuracy parity on GLM-5.2 while delivering among the fastest output speeds of any provider. Production inference workloads more than tripled sequentially in Q2 2026. Named customers include Revolut (removed human intervention from 80% of support chats; 1.2 million chat tickets per month) and monday.com. Separate pricing is published at tokenfactory.nebius.com.

Nebius Echo. An AI agent for natural-language infrastructure control, built into Nebius AI Cloud and running on open-source models served by Token Factory. Launched with Aether 3.6 (June 2026). Allows plain-language environment management, documentation-grounded Q&A, live resource status checks and simple operations such as instance creation. Roadmap includes automated infrastructure diagnostics and end-to-end multi-step deployments.

Tavily (Agentic Search). Acquired February 2026. Search infrastructure purpose-built for LLM and agentic applications, including LLM-optimised search APIs. Developer community grew from 1 million (February 2026) to more than 2.5 million (Q2 2026). Q2 2026 launches included keyless pay-per-search designed for autonomous agent consumption; ISO 27001 certification achieved for enterprise deployments.

Tendem (Human Validation). Originally a Toloka product, integrated into the Nebius ecosystem from February 2026. A hybrid human-AI agent that embeds vetted human experts directly into agentic workflows, callable via the Model Context Protocol.

Serverless AI, AI Orchestration, DataOps, ModelOps, Networking, AI Storage. Discrete product lines listed in the company's product taxonomy. Storage comprises object storage and Shared Filesystem. Orchestration capabilities were materially deepened by the Clarifai IP acquisition (May 2026), which covers AI inference and compute orchestration patents plus a perpetual non-exclusive licence to Clarifai's inference and orchestration technology stack.

Physical AI solution. Introduced Q1 2026. A managed platform offering combining synthetic data generation, orchestration, world models and real-world inference for robotics and physical AI customers. Announced in conjunction with an NVIDIA collaboration covering the full robotics lifecycle from simulation and training to deployment. In June 2026 Nebius launched a Physical AI Living Lab for UK and European robotics startups built with NVIDIA technologies.

Vertical solutions. Packaged go-to-market solutions for Scientific AI and Healthcare, Media & Entertainment, Retail & Commerce, Physical AI & Robotics, Agents & Copilots, Vision AI, Pre-Training, Post-Training, Inference and Simulation.

Programmes. Nebius Partner Program; Startup Program; Research Grants; Nebius Fellows; the Nebius Builder Program (early preview from Q2 2026, providing credits across AI Cloud, Token Factory, Tavily and Nebius Academy plus engineering office hours); and Nebius Certifications, offering verifiable credentials for AI cloud professionals. A Customer Advisory Board was launched at the 9 June 2026 Inflection Event with AMI, Black Forest Labs, Cloudflare, Cognition, Cohere, Core Automation, Higgsfield, Recraft, Revolut and Rhoda.

Nebius Academy. Open online and in-person courses plus corporate programmes in machine learning and generative AI, with cloud grants for partners and collaborations with academic institutions.

5.2 Avride

Autonomous driving technology for self-driving passenger vehicles and sidewalk delivery robots. Operating model is partnership-led, principally with Uber and Uber Eats.

• Robotaxi. AV-capable fleet exceeded 200 vehicles in May 2026, having nearly tripled since year-end 2025, with more than one million autonomous miles completed in 2026. More than 60,000 commercial rides completed on the Uber platform in Dallas; the Dallas operating map has doubled since launch. An R&D fleet collects data to support the transition to No-Vehicle-Operator operations.
• Robodelivery. Deliveries more than tripled year on year in Q2 2026 and have exceeded 600,000 since inception. Q1 2026 deliveries were over 174,000 (+178% year on year). Cities launched include Philadelphia (Q1 2026), Arlington VA and Miami (Q2 2026, via Uber Eats). Campus operations are live at Indiana University Bloomington and Salisbury College; a Master Services Agreement with Chartwells Higher Education was announced in July 2026, with Bowling Green State University among the first campuses.
• Enabling engineering. Pre-assembly of the rooftop sensor suite and other key electronic components has increased retrofitting throughput.

5.3 TripleTen

Edtech platform for reskilling and upskilling into technology careers, operating in the U.S. and Latin America. Blended bootcamp and MOOC formats with in-house content and career services supported by more than 40 hiring partners. Named Best Software Bootcamp in the U.S. by Fortune. Programmes span web development, software development, quality assurance, data science and — added Q2 2026 — AI Systems Engineering. A B2B enterprise offering with AI upskilling programmes has been expanded, with emphasis on synergies across the Nebius ecosystem. Approximately 5,000 new student enrolments were added in Q1 2026.

5.4 Equity-stake businesses (not consolidated)

• ClickHouse. Open-source real-time analytical database management system. Nebius holds a significant minority stake, reduced to approximately 25% following successive financings in which it did not fully participate. Carried at $1,517.7 million as of 31 March 2026.
• Toloka. AI data solutions business backed by Bezos Expeditions. Nebius holds approximately 81% economic interest with 49% voting power. Product expansions in 2026 include Toloka Arena (an independent agentic-intelligence benchmark suite with over 40 models evaluated), a self-service data-labelling platform upgrade with synthetic data generation, an expanded expert-validated dataset catalogue across coding, STEM, reasoning and RL Gyms, and accelerated investment in physical AI and robotics data including UMI-style data and teleoperation.

Comparability warning. The FY2025 Annual Report presents only FY2023, FY2024 and FY2025 on a continuing-operations basis. FY2021 and FY2022 relate to Yandex N.V., a fundamentally different entity reporting principally in Russian roubles with the Russian search, advertising, ride-hailing and e-commerce businesses consolidated. Those years are therefore not presented on any basis comparable to the current perimeter and are shown blank below. This is a genuine disclosure gap, not an estimate omission.

6.1 Income statement (USD millions, continuing operations)

Gross profit is not separately presented by the company. On a derived basis (revenue less cost of revenues, which excludes depreciation):

6.2 Earnings per share (USD)

Full-year FY2024 and FY2025 per-share amounts were not separately disclosed in the results releases reviewed. No dividend has been paid and the company states in its risk factors that it does not intend to pay dividends in the foreseeable future.

6.3 Balance sheet (USD millions)

Note that convertible notes are carried at amortised cost net of unamortised debt discount and issuance costs. The aggregate original principal amount outstanding as of 31 March 2026 was $10,041.8 million against a carrying amount of $8,432.0 million; the fair value of the convertible debt at that date was $9,955.7 million. Net debt computed on carrying value therefore materially understates the eventual accreted obligation, since five of the six series accrete toward 120% of original principal at maturity.

6.4 Cash flow (USD millions)

6.5 Ratios

Return on invested capital is not meaningful for FY2023–FY2025 given negative operating profit and a capital base that has grown roughly eightfold in eighteen months; it is not presented rather than estimated. The cash conversion cycle is likewise not a meaningful construct for this business — there is no inventory, and the working-capital dynamic is dominated by customer prepayments (a large negative cash conversion cycle in economic substance). As of 30 June 2026 deferred revenue of $5,975.2 million exceeded accounts receivable of $288.6 million by more than twenty times; management expects over $9 billion of customer prepayments in 2026.

6.6 Trend commentary and inflections

The revenue inflection. Group revenue compounded from $9.8 million in FY2023 to $529.8 million in FY2025 — a two-year CAGR of 635%. The first half of 2026 alone produced $981.3 million, already 85% above the whole of FY2025. Nebius AI cloud revenue grew 614% year on year in H1 2026 to $964.6 million. The proximate driver is capacity, not demand: management has stated repeatedly that "everything we build, we sell," and that Q4 2025 and Q3 2025 revenue "misses" against consensus reflected the timing of capacity commissioning rather than any softness in bookings.

The margin inflection. The single most important development of the past three quarters is the emergence of genuine operating leverage. Cost of revenues fell from 35% of revenue in H1 2025 to 24% in H1 2026; SG&A from 83% to 32%; product development from 51% to 26%; and D&A from 80% to 48%. Group adjusted EBITDA moved from –$74.7 million in H1 2025 to +$365.7 million in H1 2026, a $440.4 million swing.

The GAAP-versus-adjusted divergence. GAAP results remain loss-making and increasingly so at the operating line: the Q2 2026 operating loss widened to $175.9 million from $111.2 million. The wedge is threefold — D&A of $259.7 million, share-based compensation of $102.5 million (of which $74.9 million was a non-recurring charge tied to the Eigen AI acquisition), and interest expense of $119.1 million. GAAP net income in H1 2026 of $430.8 million is flattered entirely by the $780.6 million non-cash ClickHouse revaluation; strip that out and the underlying loss is substantial.

The depreciation policy change. Effective 1 January 2026 the company extended the useful life of servers and network equipment from four to five years, applied prospectively. The disclosed Q1 2026 effect was a $43.1 million reduction in depreciation expense and a $41.6 million increase in net income. This is a material, non-operational contributor to reported margin improvement and one that investors should normalise. It also stands in some tension with the risk factor acknowledging that "our results of operations may be adversely affected if we are not able to accurately estimate the value and useful lives of our long-term infrastructure assets."

The capital intensity. Capex rose from $807.5 million in FY2024 to $4,066.0 million in FY2025 to $8,130.3 million in H1 2026 alone, against 2026 full-year guidance of $20–25 billion. Free cash flow has been negative in every period. The distinguishing feature versus peers is the funding mix: H1 2026 operating cash inflow of $4,504.1 million was driven almost entirely by $4,395.0 million of increased deferred revenue — customers, not lenders, are funding roughly half the build.

Interest expense trajectory. Interest expense went from nil in FY2024 to $61.5 million in FY2025 to $182.8 million in H1 2026 — a run-rate approaching $500 million annually before the July 2026 secured facility. Contractual interest in Q1 2026 was $23.5 million against $35.1 million of debt discount accretion and issuance-cost amortisation, of which $15.7 million was capitalised into construction cost. As capitalisation ceases on completed assets, reported interest expense will step up mechanically.

6.7 Discrepancy register

These differences arose between the February preliminary release and the audited April filing and are consistent in magnitude with the fixed-asset control remediation described in the material weakness disclosure. Users should rely on the 20-F figures. Note also that third-party data aggregators, including Wikipedia's infobox, continue to carry the February figures (for example "operating income US$29 million (2025)," which conflates net income from continuing operations with operating income and uses the superseded number).

Management guidance

FY2026 guidance, reaffirmed 12 August 2026: group revenue of $3.0–3.4 billion; exit annualised run-rate revenue of $7–9 billion; group adjusted EBITDA margin of approximately 40%; capital expenditure of $20–25 billion; connected power of 800MW–1GW at year end; contracted power of 5GW at year end (raised from above 4GW); customer prepayments exceeding $9 billion. Beyond 2026: deployment of more than 1GW of new capacity per year starting in 2027; nearly all 5GW of contracted power online within two to three and a half years; a long-run AI cloud EBIT margin ambition of 20–30%. Formal 2027 guidance is scheduled for later in 2026.

Consensus expectations

Analysts polled by S&P Global carry an average twelve-month target of $255.44 against 18 contributors, with a $120–410 range. Zacks-tracked consensus expects a Q3 2026 loss per share of approximately $1.41 on the next report, scheduled for 10 November 2026 — implying that consensus models a meaningful widening of GAAP losses as depreciation and interest scale ahead of the H2 revenue ramp. The H1 2026 revenue of $981.3 million against full-year guidance of $3.0–3.4 billion requires H2 revenue of $2.0–2.4 billion, or roughly 2.1x the first half. That is the central execution question of the next two quarters.

Three bull-case arguments, grounded in the data

1. The pricing curve is inflecting faster than the cost curve. The 2026 installed base is priced at approximately $12 million of annual contract value per megawatt. Q2 2026 mid-term deals cleared above $20 million per MW. Q3 short-term deals are pricing at $40–50 million per MW, and the first capacity auction cleared 15% above the highest previous Blackwell price. Older-generation GPUs are repricing more than 30% higher than in Q1. Because the underlying capex per megawatt is broadly fixed at the time of purchase, every dollar of price increase falls almost entirely to gross margin. This is precisely what has compressed the payback period on Q2 deals to one year and ten months from a historical two-to-three years. If the base repricing persists as older contracts roll, the AI cloud segment's 49.7% adjusted EBITDA margin is a floor rather than a peak.

2. Customers, not creditors, are funding the build. H1 2026 operating cash flow of $4,504.1 million was driven by a $4,395.0 million increase in deferred revenue. Roughly 70% of Q2 deals included prepayments covering 50–60% of associated capex, and management expects more than $9 billion of prepayments in 2026 against $20–25 billion of capex. Total deferred revenue stood at $5,975.2 million at 30 June 2026. Nebius's Q2 interest expense of $119.1 million is 4.6% of revenue; CoreWeave's $640 million is 24.9%. In a sector where the bear case is a leveraged-real-estate blow-up, Nebius has structurally the least leveraged funding model of any at-scale neocloud, now supplemented by a repeatable secured template at SOFR + 250bp against more than $40 billion of customer commitments.

3. The asset-light model changes the terminal multiple, not just the growth rate. Launched in July 2026, the model lets partners finance, build and own facilities designed to Nebius standards while Nebius contributes software, architecture and demand in exchange for high-margin revenue with minimal capital outlay. Management reported dozens of interested partners on the Q2 call. If even a modest share of 2027–2028 capacity arrives this way, the business generates software-like revenue without carrying the infrastructure — which would justify a multiple structurally above that of an infrastructure landlord. Token Factory production workloads tripling sequentially and Tavily's developer base growing from 1 million to 2.5 million in four months are the leading indicators that the software layer is real.

Three bear-case arguments, grounded in the data

1. Concentration is extreme and delivery-conditional. Approximately $20.3 billion of the $21.33 billion RPO at 31 December 2025 was Microsoft and Meta; the second Meta agreement takes the two counterparties to roughly $47 billion of the $33.6 billion March RPO plus subsequent additions. The Microsoft contract explicitly grants tranche-level termination rights if delivery dates are missed after a grace period and no alternative capacity is available. D.A. Davidson cut its target on 10 August 2026 citing Vineland delays, and the Financial Times has reported commissioning questions at that site. A single missed tranche would not merely defer revenue — it would impair the collateral value underpinning the secured financing strategy.

2. Reported profitability is materially assisted by non-operating and accounting items. H1 2026 GAAP net income of $430.8 million is entirely attributable to a $780.6 million non-cash ClickHouse revaluation; excluding it, the group lost money. The useful-life extension from four to five years added $41.6 million to Q1 2026 net income alone and lowers ongoing depreciation across a $13.0 billion property and equipment base — precisely the estimate the company's own risk factors flag as potentially unreliable, and precisely the account where a material weakness was identified. Adjusted EBITDA excludes $137.8 million of H1 share-based compensation and $471.7 million of depreciation on assets with a genuine four-to-six-year economic life. The operating loss widened to $175.9 million in Q2.

3. The capital requirement compounds faster than the equity base. FY2026 capex guidance of $20–25 billion compares to a $68 billion market capitalisation and $10.3 billion of book equity. H1 2026 financing inflows were $9,160.4 million; shares outstanding rose 7.4% in a single quarter through the ATM alone. At the current share price all six convertible series sit above their conversion thresholds, adding roughly 51 million shares of prospective dilution and pushing Volozh's voting control below 50% on a fully diluted basis. The company itself warns that a more complex capital structure "could increase our financial and operational risks and restrict our strategic and financial flexibility." If AI capex growth decelerates — the scenario the 20-F explicitly contemplates, in which "excess industry capacity could develop, leading to volatility in the public markets and significant corrections" — Nebius would be mid-build on multiple gigawatt-scale sites with contracted obligations and a partly unfunded programme.

Catalysts and monitorables for the next twelve months

Analyst verdict

Nebius has, in the space of roughly eighteen months, converted itself from a post-divestment holding company with a Finnish data centre and $2 billion in the bank into the second-largest listed Western specialist AI cloud provider, with 5GW of contracted power, more than $40 billion of customer commitments, and a segment-level EBITDA margin approaching 50%. The Q2 2026 result is the first quarter in which the financial model and the strategic narrative fully converged: revenue up 454%, adjusted EBITDA of $236.2 million against a $21.0 million loss a year earlier, four contracts averaging more than $1 billion each, and pricing per megawatt inflecting sharply upward across every contract duration. The market's 34% single-day response was, on the evidence, proportionate.

The differentiating feature of the investment case is not growth — CoreWeave is growing too, from a larger base — but funding architecture. Nebius has engineered a model in which customers pre-fund half of the capital expenditure, payback has compressed to under two years, and the first secured facility priced at SOFR + 250bp against investment-grade contracted cash flows. Interest expense is 4.6% of revenue against 24.9% at its closest comparator. In a sector where the terminal risk is a credit event, that is the most valuable thing about this company. The asset-light partnership model, if it scales, extends the same logic to capacity itself.

The case against is equally concrete and should not be softened. Two material weaknesses in internal control remain unremediated, and the audited FY2025 figures came in $19.2 million below the preliminary release in exactly the account where the weakness sits. Reported profitability leans on a non-cash ClickHouse markup and a depreciation life extension. Customer concentration is among the most extreme of any company of this size. Dilution has been continuous and, with all six convertible series now above their conversion thresholds, is set to continue. And the entire edifice rests on an assumption — that AI compute demand compounds through 2028 and beyond — which the company's own risk factors describe with unusual candour as potentially reflecting "accelerated industry expansion and heightened market expectations rather than sustainable long-term adoption."

On balance the risk-reward has improved materially with the Q2 result, because the unit economics are now demonstrated rather than promised, and because the funding model has been de-risked by prepayments and the secured template. The appropriate posture is constructive with defined monitorables: H2 revenue conversion, Vineland delivery, and remediation of the control weaknesses. Failure on any one of the three would not merely trim estimates — it would call the financing architecture itself into question.

--