Full Report

Industry — AI Cloud Infrastructure

1. Industry in One Page

The AI cloud infrastructure industry rents GPU-accelerated compute, network, storage, and the orchestration software that make them usable for training and running AI models. Customers pay either by the hour (on-demand) or, increasingly, through multi-year reserved-capacity contracts that pre-pay billions of dollars of compute so an operator can finance the data centers and GPUs that satisfy them. Profits today exist because demand for GPUs outstrips installed supply: NVIDIA H100 rental prices rose roughly 40% in six months (SemiAnalysis, April 2026), and NBIS CEO Arkady Volozh told Reuters on May 13, 2026 that "several customers are competing for every GPU we bring online." The cycle pivots on three physical constraints that bottleneck at once — GPUs, grid-connected megawatts of power, and data-center construction lead times — and is amplified by the model-training arms race among hyperscalers and frontier labs. This is not a software business; it is a capital-intensive utility-style buildout where long-term winners need access to power and balance-sheet capacity, not just code.

2. How This Industry Makes Money

The unit of sale is compute time on a GPU, priced in either GPU-hours (on-demand IaaS), reserved-capacity dollars per year over multi-year terms, or — at the higher-margin layer — tokens processed (inference). Cost of revenue is dominated by depreciation of GPUs and data-center equipment, plus electricity, colocation rent, and bandwidth. The economic shape resembles a regulated utility crossed with a leasing business: high upfront capex, multi-year asset life (GPUs depreciated typically 4–6 years), and revenue that ideally covers depreciation plus a spread that compounds at high utilization.

Key terms in plain English:

• Neocloud: A specialized cloud whose product is GPU compute, in contrast to general-purpose hyperscalers (AWS/Azure/GCP) whose AI offering is one product among many. CoreWeave, Nebius, Crusoe, and Lambda are the canonical neoclouds.
• ARR (Annualized Run-Rate Revenue): Last-month revenue × 12. Used because GPU revenue ramps faster than reported quarterly GAAP revenue catches up.
• Contracted backlog: Dollar value of signed multi-year capacity commitments not yet delivered.
• Contracted power: Megawatts of grid-connected power secured by signed land/utility contracts. Power has overtaken GPUs as the binding constraint on growth.
• Hyperscaler vs. neocloud: Hyperscalers sell broad enterprise software and increasingly buy GPU capacity from neoclouds (e.g., Microsoft buying $17B of capacity from Nebius). Neoclouds rent specialized AI infrastructure back to hyperscalers, model labs, and AI-native startups.

Margins are highest at the silicon layer and lowest in raw GPU rental once supply equalizes. Today the neocloud layer earns scarcity rents — gross margins comparable to a software business — but those margins compress when GPUs become abundant. Whether each neocloud can layer enough software (managed services, inference, fine-tuning, orchestration) on top of raw compute to keep blended margins above commodity IaaS once the shortage breaks is the strategic question.

Capital intensity is the single most distinguishing fact of the industry. Nebius spent $2.47B of capex against $399M of revenue in Q1 2026 — a capex-to-revenue ratio above 6x. CoreWeave has guided $30–35B of 2026 capex against a $5.1B 2025 revenue base. These businesses look like utilities mid-buildout, not asset-light SaaS.

3. Demand, Supply, and the Cycle

Demand has three engines that pulse on different rhythms. Frontier model training (the largest LLM and multimodal runs) is concentrated in a handful of buyers — Microsoft/OpenAI, Meta, Google, Anthropic, xAI — and tends to come in lumpy multi-billion-dollar commitments tied to model generations. Enterprise AI adoption is broader but slower, sitting in pilots and early production through 2025–2026. Inference — running, not training, models — is the fastest-growing pool and is projected by analysts to account for roughly two-thirds of compute demand by 2026 as deployed AI apps scale.

Supply is constrained by three things in order: power (multi-year grid interconnect queues, particularly in the US), GPUs (allocation from NVIDIA, which prioritizes large committed buyers), and data-center construction (18–36 month builds with HVAC, water, cooling, and dense fiber). When even one of the three is short, the others sit idle. Today all three are tight simultaneously, which is what produces the pricing power neoclouds are currently earning.

The industry has not experienced a full downturn since AI compute became a category in 2023. The most likely template is the telecom-fiber bust (2001–2002) and the crypto-mining unwind (2022–2023): a glut of capacity, GPU rental prices that fall faster than depreciation, and operators carrying long-dated debt against assets that depreciate four times faster than data-center shells. The first signal of a turn would show up in spot GPU pricing and in shorter contract tenors at renewal.

4. Competitive Structure

The market is a barbell. At one end sit three hyperscalers — AWS (28% share of Q1 2026 enterprise cloud spend per Synergy Research), Microsoft Azure (21%), and Google Cloud (14%) — which combined hold roughly 63% of total cloud spend and increasingly buy AI capacity from neoclouds rather than only sell their own. At the other end sits a fast-growing tier of specialized AI cloud builders — CoreWeave, OpenAI's own infrastructure, Oracle (a hyperscaler-class entrant via OCI), Crusoe, Nebius, Anthropic's stack, ByteDance — that Synergy explicitly singled out as Q1 2026's fastest-growing tier-2 providers. The middle of the market is thin.

Three structural observations matter. First, the market is not winner-take-all at the IaaS layer — frontier customers actively want to diversify away from a single hyperscaler. Microsoft buying $17B of compute from Nebius and Meta buying $27B from Nebius are explicit second-sourcing decisions. Second, customer concentration is the single biggest risk for every neocloud: Nebius's two hyperscaler contracts alone are reported at roughly $46B of contracted backlog, and CoreWeave's revenue is similarly concentrated. Third, the bitcoin-miner-to-AI pivot peers (IREN, APLD, HUT) are real participants but trail neoclouds on the software stack, so they tend to sell power + halls + hosting rather than fully managed AI cloud.

5. Regulation, Technology, and Rules of the Game

The industry sits at the intersection of three regulatory frameworks (AI conduct, export control, data/energy) plus a fast-moving technology refresh cycle. Each can move unit economics by 10–30 percentage points if it cuts the wrong way.

The two regulations most likely to move investor judgement are US export controls (which both constrain NBIS's geographic expansion and concentrate global supply in the operators that can comply) and grid-interconnect cost allocation (the new bottleneck on greenfield data-center economics — the FY2025 20-F flags it directly). The technology refresh that matters most this year is NVIDIA's Vera Rubin NVL72 platform arriving in H2 2026, which will drive the next training-capex cycle.

6. The Metrics Professionals Watch

Most general-cloud KPIs (DAUs, ARPU, ARPC) are useless here. The metrics below are what AI infrastructure analysts actually model.

7. Where Nebius Group N.V. Fits

Nebius is a specialized AI cloud (neocloud) with an integrated full-stack model: it designs servers and racks in-house, owns most of its data-center capacity (>75% of contracted power is owned, not leased), holds NVIDIA Reference Platform Cloud Partner status, and is layering an inference / token-priced platform (Token Factory, accelerated by the May 2026 $643M Eigen AI acquisition) on top of raw GPU rental. Within the cohort, Nebius sits between CoreWeave (the largest pure-play neocloud) and the BTC-pivot peers (IREN/APLD/HUT) — closer to CoreWeave in product depth, but distinctively European-domiciled with most owned capacity and a multi-business holding structure (Nebius core + Avride autonomous vehicles + TripleTen edtech + equity stakes in ClickHouse and Toloka).

8. What to Watch First

Know the Business — Nebius Group N.V.

Nebius is a capital-intensive, customer-financed AI infrastructure utility-in-the-making wrapped inside a Dutch holding company that also owns an autonomous-vehicle business (Avride), an edtech (TripleTen), and minority stakes in two software companies (ClickHouse, Toloka). The economic engine is the core AI Cloud — 98% of Q1 2026 revenue, the only segment generating positive Adjusted EBITDA, and the recipient of more than $20 billion of planned 2026 capex. The market is most likely overestimating how much of the long-dated $46 billion contracted backlog will translate cleanly into equity value, and underestimating the embedded option value in two assets that do not appear on the income statement: the ClickHouse stake (remeasured to ~$1.6B in Q1 2026) and the owned-power optionality in two US gigawatt sites.

1. How This Business Actually Works

Nebius rents GPU compute to AI customers — but the real engine is the conversion of customer cash, debt, and equity into power, GPUs, and data-center concrete, then into multi-year revenue contracts that earn a depreciation-plus-spread. This is a utility-style buildout disguised as a software business. In Q1 2026 the company spent $2.5B of capex against $399M of revenue — a 6.2x capex-to-revenue ratio — and pulled in $2.3B of operating cash flow almost entirely from a $3.2B build-up in deferred revenue (customers prepaid for capacity not yet delivered). The business in one sentence: customers prepay, NBIS builds, depreciation begins, and the spread accretes for 4–5 years.

The unit economics that matter are utilization and contract tenor, not headline gross margin. The Nebius core segment hit a 45% Adjusted EBITDA margin in Q1 2026 against an industry expectation of 20-30% in steady state — but Adj. EBITDA strips out the $212M of quarterly D&A that is the cash cost of GPU depreciation already incurred. GAAP operating loss was $128M in Q1 2026 because depreciation on a 4–5 year useful life now flows through faster than ramp-stage revenue can absorb it. The math only works if utilization stays high through the depreciation window — which is why the Microsoft and Meta reserved-capacity contracts are the core financing instrument, not just revenue.

The leverage is operating leverage on a fixed base, not pricing leverage. SG&A and product development moved from 192% of revenue to 53% in twelve months because revenue scaled, not because absolute spend fell. The same dynamic reverses if utilization falls before the next capex wave is fully ramped.

2. The Playing Field

The peer set is a barbell: one hyperscaler (Oracle), one direct-comparable specialized AI cloud (CoreWeave), and three power-led infrastructure pivot stories (IREN, Applied Digital, Hut 8). Nebius sits closest to CoreWeave on product depth and customer profile, but is materially smaller and trades at a higher EV/Revenue multiple because the market is valuing forward ARR — not LTM revenue — on a base that grew 684% in Q1 2026.

Three things the peer set reveals. First, Nebius is being valued on its forward run-rate, not its trailing print: $51.8B EV against $530M of FY2025 revenue (98x) compresses to roughly 27x against $1.92B Q1-exit ARR — still richer than CoreWeave (18.5x EV/LTM revenue) but no longer in a different universe. Second, the BTC-pivot peers (IREN, APLD, HUT) are real participants on power and capex but have not yet built the software layer that lets a neocloud earn software-like gross margins; APLD's 30% gross margin is the warning sign. Third, Oracle is the only peer with a positive operating income at scale — the model for what a neocloud's economics could look like at maturity, but only if the software layer holds margins when the GPU shortage breaks.

What "good" looks like in this industry is a combination CoreWeave does not yet have and Oracle is too diversified to demonstrate cleanly: owned power, vertical hardware integration, and software margin on top of compute. Nebius is the only peer in the set that explicitly claims all three. Whether that translates into durable returns on the $20-25B of 2026 capex is the central investment question.

3. Is This Business Cyclical?

It is, but the cycle has not happened yet. AI cloud has been in a single-direction up-cycle since 2023 — H100 spot prices rose 40% in the six months to April 2026 — and Nebius has only ever operated in a shortage. The cycle will hit four places at once when supply catches up: GPU rental price, utilization, contract tenor at renewal, and ABS / convertible market terms that fund the next capex wave. The historical templates are the 2001–2002 telecom-fiber bust and the 2022–2023 crypto-mining unwind: oversupply, prices falling faster than depreciation schedules, and operators carrying long-dated debt against assets that depreciate in 4–5 years.

Capex precedes revenue by roughly 18 months, and the gap currently sits at ~7x. If demand falters before the 2026 capex base rolls into the rate base, the operating leverage that produced 32% Adj EBITDA margins in Q1 2026 reverses sharply because depreciation is locked in but revenue is not. The first leading indicator of a turn would be shorter contract tenors at renewal — the day a hyperscaler signs a 12-month deal where it used to sign 5-year, the equity multiple compresses regardless of GAAP earnings.

4. The Metrics That Actually Matter

The five metrics below determine whether the equity compounds or impairs.

NBIS is the only peer that scores strongly on all five of the metrics that determine durable returns in this industry. CoreWeave matches it on ARR and contracted power but is materially levered. The BTC-pivot peers match on owned capacity but have no software layer. Oracle has all of these in spades but at a scale where AI cloud is a sub-segment of a much larger business.

5. What Is This Business Worth?

Value here is determined by two things in tension: the run-rate cash flow that the AI cloud build will produce when 4 GW of contracted power is fully ramped, discounted by the financing cost of getting there, and the carry value of two non-operating assets (ClickHouse stake, Avride) that consolidated metrics hide. A clean DCF on the AI cloud alone is premature — the depreciation curve, utilization assumption, and terminal multiple all swing the answer by an order of magnitude. The right framing is sum-of-the-parts, because the consolidated income statement mixes a build-phase utility, two cash-burning pre-revenue businesses, and two minority software stakes that are not priced in operating cash flow.

The investment math tilts on the Nebius core. ClickHouse contributes maybe $5-7 per share at the current carry; Avride and TripleTen are option-value rather than valuation drivers; net cash is earmarked. Roughly 80-90% of the $52B equity market cap is the present value of the AI cloud business — and that value is mostly forward-looking. At $52B EV against $1.92B Q1-exit ARR, the market is paying ~27x ARR for an asset that mgmt expects to grow to $7-9B by year-end. The valuation works if ARR compounds at the guided rate without margin compression and if the next $20B of capex earns returns above its financing cost; it breaks if either condition fails. The margin of safety is in the contracted backlog, not the multiple.

6. What I'd Tell a Young Analyst

Stop modeling this as a software company. The right mental model is a regulated utility in mid-buildout, with the twist that revenue is locked in by reserved-capacity contracts before the asset is built — so customer prepay is the cheapest financing in the business and the most important number on the cash flow statement. Watch four things in order. First, contracted power additions each quarter — they are the cleanest leading indicator of forward revenue, more reliable than ARR which lags. Second, contract tenor at renewal — when a hyperscaler signs a shorter deal than the prior one, the cycle has turned. Third, the spread between depreciation expense and revenue per MW — the silent killer if utilization slips. Fourth, the ClickHouse mark — the only piece of equity value that does not depend on the AI cloud build going right.

The market is most likely wrong right now to treat consolidated revenue and consolidated EBITDA as the right denominators. Almost all the EBITDA is the Nebius segment (45% margin in Q1 2026), almost all the loss is Avride and TripleTen, and almost all the future depreciation is GPUs already on the balance sheet. If you mix them you will under-value the core and miss the optionality in the stakes. If you are in this name, you are underwriting execution on power, GPU allocation, and the financing stack — not TAM and not a software multiple. The thesis breaks the day a Microsoft or Meta contract is renegotiated, or the day spot GPU prices roll over before the next capex tranche is in the rate base.

Competition — Nebius Group N.V.

Competitive Bottom Line

Nebius has a real but contingent advantage anchored on three things its closest public peer does not stack together: owned greenfield power (>75% of contracted MW), in-house full-stack hardware and software (Aether 3.5 + Token Factory + Reference Platform NVIDIA Cloud Partner status), and a clean net-cash balance sheet ($0.9B net cash) into a $20-25B 2026 capex cycle. The competitor that matters most is CoreWeave (CRWV) — same customer base (Microsoft, Meta, AI labs), same go-to-market motion, and the only public peer in the same product orbit. CRWV is roughly 10x NBIS's trailing revenue but levered ($14.7B long-term debt) and primarily a tenant in third-party data centers. The hyperscalers (Oracle and the AWS/Azure/GCP cohort) are not "beat-able" in absolute terms, but they are the customers and second-source backstops for neoclouds rather than direct hand-to-hand fighters at the sub-$10B revenue tier. The BTC-to-AI pivot peers (IREN/APLD/HUT) compete for power and land but today sell halls and hosting — not managed AI cloud — so they erode capacity availability rather than software margin. The moat is execution-on-power + GPU allocation + financing access; it weakens fastest the day GPU spot prices roll over before the next gigawatt rolls into the rate base.

The Right Peer Set

The peer set is anchored on the four cohorts named in the NBIS FY2025 20-F competition disclosure (Item 3.D): specialized AI clouds (CoreWeave, Crusoe, Lambda — the latter two private), named hyperscalers (AWS, Azure, GCP, Oracle — Oracle is the cleanest public comp by AI-cloud focus), and HPC/AI pivot operators that compete for the same power and customer pool (IREN, APLD, HUT). Microsoft and Alphabet are retained as supplementary context because they are both NBIS's largest customer (Microsoft $17B reserved capacity contract) and a competitor (Azure) — a dual relationship that recurs across every named hyperscaler. Crusoe and Lambda are excluded because they are private; Cerebras is excluded as custom-silicon, not infrastructure substitute; pure BTC miners (Riot/MARA/CIFR/WULF/Bitfarms) are excluded because they lack a credible AI/HPC pipeline.

First, Nebius is the smallest revenue-base name in the cohort apart from APLD/HUT, but the enterprise value puts it firmly in the specialized-AI-cloud orbit — not the BTC-pivot orbit — because the market is pricing forward ARR ($7-9B guided exit 2026) rather than trailing print. Second, CRWV is the only meaningful pure-play public comp; everything else is either many times larger and diversified (Oracle), or many times smaller and not yet a managed AI cloud (IREN/APLD/HUT). Third, all five peers have a top-customer-concentration profile that makes any single hyperscaler renegotiation a binary event.

Where The Company Wins

Four advantages that show up in third-party filings or NBIS disclosure — not management marketing.

1. Owned greenfield power on a hyperscaler-grade scale. Nebius reports >75% of its contracted ~3.5 GW as owned (greenfield + build-to-suit), versus CoreWeave which leases the bulk of its 43 data centers and only began developing its own DCs in 2025 (Kenilworth, NJ JV). Owned capacity locks long-run unit economics (power cost, racks, density), avoids landlord pricing risk at renewal, and is the binding constraint on growth that GPUs no longer are. CoreWeave's own FY2025 10-K explicitly flags grid power as the top business risk and notes the November 2025 delay of "certain data centers to be provided by a third-party data center provider." That is operational exposure NBIS owned sites do not have.

2. Net-cash balance sheet into a $20B-plus capex year. NBIS exited Q1 2026 with $9.3B cash against $8.4B long-term debt — roughly $0.9B net cash — after a $4.3B convertible raise and a $2B prefunded warrant equity injection from NVIDIA. CoreWeave's FY2025 balance sheet shows $3.2B cash against $14.7B long-term debt (roughly $11.5B net debt). For a capex-cycle business, that 12-billion-dollar net-leverage gap is the difference between issuing on offensive terms and issuing in a downturn.

3. NVIDIA "Reference Platform" Cloud Partner status + Vera Rubin first-to-deploy. Nebius is one of a small number of operators elevated to NVIDIA Reference Platform NCP — first access to Blackwell and Vera Rubin NVL72 generations, and an embedded relationship reinforced by the Q1 2026 $2B NVIDIA equity investment. CoreWeave also holds NVIDIA preferred-partner status (they have led NVIDIA's GB200/GB300 NVL72 deployments per their 10-K), so this is parity at the very top tier — but no other peer in the set is in that tier. APLD/IREN/HUT are GPU buyers, not reference-platform partners.

4. In-house full-stack software (Aether + Token Factory). Nebius designs its own servers and racks, runs its own orchestration platform (Aether 3.5), and operates a token-priced inference layer (Token Factory, accelerated by the May 2026 $643M Eigen AI acquisition). CoreWeave is the only competitor with a similar depth of proprietary software (Mission Control, SUNK, Weights & Biases acquisition) and an industry-only Platinum ClusterMAX rating. The BTC-pivot peers (IREN/APLD/HUT) have effectively no software layer — they sell power, halls, and hosting. APLD's own 10-K describes itself as a "data center infrastructure solutions" provider; Highrise AI (HUT's AI cloud) operated just 1,000 H100 and 96 H200 GPUs at YE 2025.

Nebius is the only operator in the cohort that scores 5 across both owned power and in-house software-class infrastructure. CoreWeave matches it on software and NVIDIA tier but is materially levered and primarily a tenant. The BTC-pivot peers match on owned land but lose four of the five rows. Oracle wins on everything except owned-power exclusivity, but at a scale where the relevant question is what fraction of $57.4B revenue is the OCI AI-cloud segment — not whether Oracle can outcompete Nebius for a Meta-style $27B reserved-capacity contract.

Where Competitors Are Better

Four places where competitors are demonstrably ahead — the gaps an investor underwrites.

1. CoreWeave is roughly 10x larger and has more committed customer dollars under signed contract. CRWV reported $60.7B of remaining performance obligations (RPO) at YE 2025 with a ~5-year weighted-average duration — directly comparable to NBIS's reported ~$46B contracted backlog. CRWV also has more deployed power (>850 MW active versus NBIS's "operational" footprint that is smaller in MW even if total contracted is similar). At today's scale, CRWV gets first call on the next NVIDIA allocation tranche and on the next hyperscaler reserved-capacity carve-out. The Q1 2026 CRWV revenue print (extrapolating from $5.1B FY2025 plus growth) is more than 10x NBIS's $399M Q1.

2. Oracle is the only peer earning steady-state-grade economics today. Oracle FY2025 (ended May 2025) reported $57.4B revenue, $17.7B operating income (30.8% OI margin), and $12.4B net income. Every other peer in the set is GAAP-loss-making at the operating level: NBIS Q1 2026 GAAP operating loss $128M, CRWV FY2025 operating loss $46M (and $1.2B net loss), IREN positive op income $17M on $501M revenue (BTC mining cash + nascent AI cloud), APLD operating loss $17M, HUT net loss $226M. Oracle is the existence proof that an AI-cloud-adjacent business can earn ~30% operating margin at scale — but doing so requires the diversified software base Oracle already has and NBIS has not built.

3. Oracle and the hyperscalers have geographic and customer diversification NBIS does not. Nebius's two largest customers (Microsoft, Meta) reportedly account for approximately $44B of the ~$46B contracted backlog. CRWV faces the same concentration risk and discloses it as a top-3 risk factor. Oracle's customer base is in the tens of thousands across applications and infrastructure; AWS/Azure/GCP's runs to millions. Geographic exposure is similar: NBIS has US, Finland, Israel, and announced PA + Missouri + Alabama sites; CRWV operates in six countries; Oracle operates in 175 countries. If a single hyperscaler renegotiates, NBIS's contracted-backlog story is binary.

4. Power-led BTC peers (IREN, HUT, APLD) hold more permitted long-tail power than NBIS does today, and earned cash flow to fund their pivot. IREN reports 810 MW operating + 2 GW under development (1.4 GW Sweetwater 1 + 600 MW Sweetwater 2) in Texas; HUT has 1,020 MW under management plus 1.2 GW in development across multiple sites; APLD has 286 MW operating + 400 MW HPC under construction (Polaris Forge 1). IREN is the only peer in the cohort that printed positive operating income in FY2025 ($17.3M), with $86.9M net income, funded almost entirely by BTC mining cash flow. That cash is what pays for the AI/HPC build-out — Nebius has to do the same build using customer prepay, convertibles, and equity. Whichever operator gets the next gigawatt site interconnected first wins the next contract; this is not a race NBIS is winning yet at the long-tail-power layer.

Nebius is not behind on power-permitting or on software, but is behind on trailing scale, GAAP profitability, customer diversification, and organic cash generation. The bull case is that the contracted backlog converts those gaps into parity within 18 months; the bear case is that the gaps to CRWV and the cash-generative pivot peers widen if NBIS's next financing window closes before the next capex tranche hits the rate base.

Threat Map

Six threats — the ones that would actually re-rate NBIS, not generic "competition" risk. Severity reflects probability x equity-value impact over an 18-month window.

Three of the four High severity items (hyperscaler in-sourcing, CRWV scale advantage, customer dispute) share a single underlying exposure: customer concentration. The fourth (capital-markets close) is a financing-channel risk, not a competitor risk. None of the Medium threats — Oracle bundling, BTC-pivot peers, sovereign subsidies — is large enough on its own to re-rate Nebius in 18 months, but together they compress the differential margin the software layer is supposed to defend.

Moat Watchpoints

Five measurable signals — in order of leading-indicator power — that show whether Nebius's competitive position is improving or weakening.

A note on what is not on the watchlist. "Gross margin" is not a moat signal in this industry yet — it is dominated by GPU depreciation policy (NBIS extended useful life from 4 to 5 years in Q1 2026, an accounting choice). "GAAP operating income" is similarly dominated by build-phase D&A. "Revenue growth rate" is mostly a function of capacity available to deploy, which is driven by power additions, not sales effort. Read those metrics as confirmations, not leading indicators.

Current Setup & Catalysts

1. Current Setup in One Page

The stock is at $207.27, a fresh all-time high posted intraday on May 13, 2026, the morning after a blowout Q1 print that crushed expectations (revenue +684% YoY, ARR $1.92B, group adjusted EBITDA flipping to +$130M / 32% margin, Nebius AI segment EBITDA 45%). What the market has spent the last 3-6 months repricing is the move from "neocloud bet" to "second-source hyperscaler" — Microsoft $17.4B (Sep 2025), Meta $27B (Mar 2026), NVIDIA $2B equity (Mar 2026), $4.3B convertibles (Mar 2026), three software acquisitions (Tavily Feb, Eigen AI May 1, Clarifai May 12) — and that repricing is largely complete: the stock is +522% over twelve months and 100% above its 200-day SMA. The next 6-12 weeks will be quiet on confirmed dated catalysts; the single event that will define the rest of the year is the Q2 2026 ARR print on July 29, 2026. Management has telegraphed a Q2 EBITDA-margin step-down (back-end-weighted Q3 capacity step-up), so the market is set up for a mechanical margin dip and an ARR re-acceleration, not the other way around. The unresolved questions sit underneath: how much of the $46B backlog is firm vs flex (Meta's $27B is split $12B firm / $15B optional), whether Deloitte signs off cleanly on FY2026 ICFR after Reanda's adverse 2024 ICFR opinion, and whether ARR can hold its 47%-per-quarter pace as Vera Rubin NVL72 supply ships in H2 2026.

Recent setup: Bullish — Q1 2026 print on May 13 crushed expectations across revenue, ARR, EBITDA, and contracted-power additions.

---

2. What Changed in the Last 3-6 Months

The story arc since February: a Q4 2025 print that missed on revenue ($228M vs $247M consensus) on Feb 12, 2026 — followed by eight consecutive positive catalysts — has rerated the equity from a speculative AI-cloud beta into a contracted-revenue compounder. The current narrative is "the only sub-hyperscaler neocloud that prints hyperscaler-class economics on a net-cash balance sheet."

Narrative arc. Investors used to debate whether Nebius was a real second-source hyperscaler or a power reseller; that debate is settled in favor of the bull. What they care about now is execution velocity (does Q3 capacity activate on schedule, does Q2 ARR hold the 30%-QoQ pace), earnings quality (does the Q1 margin print survive Deloitte's first ICFR review, does the useful-life extension hold), and concentration (does anyone other than Microsoft or Meta sign a $1B+ deal in 2026). The two unresolved questions that still drive the next move are (1) whether the $46B headline backlog is really $44B firm or closer to $31B firm with $15B optional and (2) whether the customer-prepay-funded $2.3B Q1 operating cash flow is repeatable.

---

3. What the Market Is Watching Now

The live debate is no longer "is the demand real" but "how much of the demand is firm and repeatable." Each of the six watch items is a different way of pressure-testing that. Q2 ARR is the cleanest; the prepayment / OCF question is the cleverest; the third hyperscaler is the biggest swing factor on multiple expansion.

---

4. Ranked Catalyst Timeline

Ranked by decision value to a hedge-fund underwriting NBIS today, not by chronology. The Q2 print and the Q3 capacity step are the only catalysts likely to move the stock by more than 10% on a single day inside the next six months.

---

5. Impact Matrix

The five catalysts below are the ones that actually resolve the bull/bear debate. The rest add information without changing the thesis.

The matrix is two pure-bull resolves (Q2 ARR, Q3 capacity + Rubin), two pure-bear resolves (concentration, OCF repeatability), and one forensic resolve (Deloitte ICFR). The Q2 print is the only one of the five that is hard-dated inside the next 90 days; the others are the watchlist that runs through year-end and into Q1 2027.

---

6. Next 90 Days

The shape of the next 90 days: quiet through June (AGM and possibly an ABS deal), one major event on July 29 (Q2 print). The tape is technically extended, positioning is crowded, and the catalyst calendar is thin — which means the next big move is largely a single date away. A PM who is long here is essentially making one bet (Q2 ARR holds the trajectory) wrapped in a passive flow tailwind. A PM who is short here is fighting +522% momentum into a known-bullish event window with 21% of float already short.

---

7. What Would Change the View

Three signals would force the debate to update. First, the Q2 2026 ARR print on July 29 — anything outside $2.2-2.7B materially changes the bull/bear balance, with the asymmetric bear outcome (sub-$2.2B) more powerful than the symmetric bull (above $2.7B) because the bull case is already substantially priced in at $207. Second, any disclosed $1B+ contract from a third hyperscaler (AWS, Google, Oracle, or a sovereign program) that drops Microsoft + Meta concentration below 70% — this is the single observable signal that converts the binary "two phone numbers" bear thesis into a real diversified moat and would force a cover regardless of forensic concerns. Third, the Deloitte FY2026 ICFR opinion in spring 2027 — a clean opinion resets the forensic grade from Elevated to Medium and validates the useful-life extension, the customer-prepay treatment, and the segment-margin print; a qualified or adverse opinion does the reverse and creates a re-rating event independent of operations. Tied to bull/bear/moat/forensic concerns: a 90-day window where ARR holds, ABS prices inside converts, and concentration partially de-risks would compound into a multiple expansion toward the $250-291 high-PT range. The opposite combination — ARR slip, ABS pulled, no new hyperscaler, third material weakness — would compress the multiple toward the $105-125 Morgan Stanley / Cantor end of dispersion. Everything else is incremental.

Bull and Bear

Verdict: Lean Long, Wait For Confirmation — the bull's quantitative case (ARR trajectory, $0.9B net cash into a $20–25B capex year, NVIDIA equity backing, EV/exit-2026 ARR of ~6.6x) carries more weight than the bear's, but the bear's forensic flags are credible enough that the prudent path is to wait for the Q2 2026 ARR print (early August 2026) to confirm or break the case before sizing in.

The decisive tension is customer concentration: ~95% of the $46B contracted backlog sits with Microsoft and Meta. The bull reads that as the strongest counterparty validation an infra business can get; the bear reads it as a put written on two customers who are themselves spending $80B+ and $65B+ on owned AI capacity. Every other debate in this name — the 45% segment EBITDA margin, the $402M reported operating cash flow, the 27x ARR-to-trailing-revenue ratio — collapses or compounds depending on whether those two contracts hold their disclosed tenor and pricing.

The single piece of evidence that would change the conclusion is the Q2 2026 ARR print. Above $2.5B keeps the bull's upside scenario intact; below $2.2B materially impairs it and the verdict moves to Avoid. A disclosed third $1B+ hyperscaler anchor would convert this to a clean Lean Long regardless of accounting noise.

Bull Case

Upside scenario ~$300 / 12–18 months. Method: $8.5B exit-2026 ARR × 9x EV/ARR (50% discount to CoreWeave's current multiple on lower leverage and faster growth) ≈ $76.5B EV; plus $1B net cash, $2B ClickHouse mark, and ~$2B Avride / minority-stake option value ≈ $81.5B equity ÷ ~272M diluted shares ≈ $300. Primary catalyst is the Q2 2026 ARR print (early August 2026); a reading above $2.5B (>30% QoQ) supports the $7–9B exit guide. Disconfirming signal: Q2 2026 ARR under $2.2B (less than 15% QoQ), or any disclosed renegotiation or tenor-shortening on the Microsoft or Meta reserved-capacity contract.

Bear Case

Downside scenario ~$105 / 12–18 months. Method: peer-multiple compression to ~10–12x EV/forward revenue (CoreWeave-like, vs current ~16x on FY26 guide and ~27x on Q1 2026 ARR), assuming ARR lands at the bottom of the $7–9B guide and the market re-rates on customer-concentration concerns; EV ≈ $35–40B less ~$3–5B incremental net debt from FY26 capex funding ÷ ~290M fully diluted shares ≈ $105–115. Primary trigger is the Q2 2026 ARR print under $2.5B (less than 30% QoQ); secondaries include any disclosed shortening or repricing on either anchor contract or a third material weakness in Deloitte's FY2026 ICFR opinion. Cover signal: a new $1B+ contracted hyperscaler customer (Amazon, Google, or sovereign program) that drops top-2 concentration under 70%.

The Real Debate

Verdict

Lean Long, Wait For Confirmation. Bull carries more weight on the near-term observables that drive the next 12 months — a 6.6x EV/exit-2026 ARR multiple against 12–18x at CoreWeave, a net-cash balance sheet entering a $20–25B capex year, and NVIDIA's $2B equity sitting behind supply allocation — and a single Q2 2026 ARR data point gates the upside scenario. The single most important tension is customer concentration: ~95% of the $46B backlog sits with Microsoft and Meta, and that one fact is what makes every bull number compound and every bear number explode in the same direction. The bear could still be right because the forensic file is unambiguous — a same-quarter useful-life extension worth $167.6M of depreciation relief, organic CFO of approximately −$1.56B once prepayments and the AP swing are stripped, two ICFR material weaknesses, and a 587-day DSO are not isolated marks. The verdict moves to Avoid on a Q2 2026 ARR print under $2.2B (less than 15% QoQ) or any disclosed shortening or repricing on either anchor contract; it moves to clean Lean Long on a disclosed third $1B+ hyperscaler anchor that drops top-2 concentration under 70%. Until the Q2 print lands in early August 2026, the asymmetry favors patience over anchoring conviction to either side.

Moat — What Protects This Business

1. Moat in One Page

Conclusion: Narrow moat — and contingent on the cycle. Nebius has assembled four candidate advantages that are real today but only two of which look durable through a supply normalization: (1) owned, gigawatt-scale grid-connected power (>75% of contracted ~3.5 GW), (2) NVIDIA "Reference Platform" Cloud Partner status with first-to-deploy access to Blackwell and Vera Rubin NVL72, (3) a full-stack in-house software layer (Aether 3.5 orchestration + Token Factory inference + the May 2026 $643M Eigen AI acquisition), and (4) multi-year reserved-capacity contracts with Microsoft (~$17B) and Meta (~$27B) that pre-fund the build. The two that should survive a cycle are owned power (a 4-7 year permit + interconnect lead time competitors cannot copy on demand) and NVIDIA NCP Reference Platform tier (a designation NVIDIA grants to a small number of operators). The other two are softer: the software layer is unproven against CoreWeave's Mission Control + SUNK + Weights & Biases stack, and the contracts — while binding — are revocable in the sense that ~95% of backlog sits with two customers who are themselves building owned AI capacity ($80B+ Microsoft capex run-rate, $65B+ Meta 2026 capex). The honest read is this looks more like a scarcity-rent + access advantage during a supply shortage than a textbook moat; the business has never been tested in a glut. A beginner investor should understand that "wide moat" is what NVIDIA, AWS, and Oracle have; what NBIS has is privileged access to scarce inputs and one of the few public balance sheets clean enough to build out faster than competitors.

Moat rating: Narrow — owned power + NVIDIA NCP Reference Platform tier durable through cycle; software layer + reserved-capacity contracts softer. Weakest link: customer concentration (~95% of backlog with Microsoft + Meta).

2. Sources of Advantage

A moat must have an economic mechanism — not just a story. Below are the seven candidate sources of advantage for Nebius, each tested against three questions: how could it protect the business, is there company-specific evidence, and what would erode it?

The pattern across these seven sources: only owned power scores a High on proof quality, and even that is matched in gigawatts permitted by BTC-pivot peers. The other six are Medium or Low. There is no single hard moat; there is a stack of contingent advantages whose strength depends on the GPU + power shortage continuing. The mechanism that ties the stack together is the reserved-capacity contract: it converts scarcity rent into prepaid cash that funds the next build, which extends the lead, which lets NBIS sign the next contract. That virtuous loop is the moat — but only while the loop is fed.

3. Evidence the Moat Works

A moat is a claim about the future, but the only way to test it is to look at what the business has actually achieved. Below are eight evidence items — five that support the moat thesis and three that refute or qualify it.

The ledger is 5 supports vs 3 refutes, which is narrow-moat shape — not no-moat, not wide-moat. The two refuting items that bite hardest are (1) the customer-prepay-funded cash flow (without it the business is cash-negative) and (2) the binary concentration risk on two customers. The supporting items lean heavily on owned power, NVIDIA backing, and the contracted backlog — all real, but all interdependent.

4. Where the Moat Is Weak or Unproven

Four places where the moat claim should be interrogated harder.

1. Two-customer concentration is not a moat — it is a counterparty exposure dressed as one. A reserved-capacity contract creates a switching cost for the customer. But when 95% of contracted backlog sits with Microsoft and Meta — both of whom are simultaneously building owned AI infrastructure at $80B+ and $65B+ run-rates respectively — the contract is more like a put written on counterparty intent than a moat. If Microsoft decides to fulfill more of its compute internally, the renewal economics shift, and NBIS does not have hundreds of mid-sized customers to absorb the loss. CoreWeave faces the identical exposure, which is why both names trade as concentrated-counterparty stories, not diversified moats.

2. The software layer is not yet differentiated against CoreWeave. Token Factory is real, the Eigen AI acquisition is real, but CoreWeave's Mission Control + SUNK orchestration + the Weights & Biases acquisition is at least as deep, and SemiAnalysis rates CRWV's stack as the only Platinum-tier "ClusterMAX." Hyperscalers (Oracle, AWS, Azure, GCP) bundle equivalent services with broader enterprise software customers do not buy from NBIS. Until NBIS discloses inference-segment revenue and gross margin, this leg of the moat is a story, not evidence.

3. Owned power is matched in gigawatts by BTC-pivot peers. IREN has 810 MW operating + 2 GW in development; HUT has 1,020 MW under management + 1.2 GW pipeline; APLD has 286 MW operating + 400 MW HPC under build. NBIS's >3.5 GW contracted is larger than any single BTC-pivot peer, but the gap is "first-mover-into-AI" rather than "exclusive control of scarce power." Whoever interconnects the next greenfield gigawatt first wins the next reserved-capacity contract — that is a race, not a moat.

4. The financing-channel advantage is a cycle position, not a structure. NBIS's $0.9B net-cash vs CoreWeave's $11.5B net-debt gap is a real differentiator today because it gives NBIS the ability to issue when peers cannot. But a reopening of credit markets, or a sufficiently dilutive equity raise, closes the gap. Capital intensity is a barrier to entry (which is industry-level) more than a moat (which should be company-specific).

5. Moat vs Competitors

The right benchmark for NBIS's moat is what does the next-best operator have, and where is NBIS demonstrably better? The table below picks one moat source per peer and grades NBIS against it.

The pattern: NBIS is a 5 on owned power, NVIDIA tier, and balance-sheet headroom, ties CoreWeave on contracts and trails it on software stack depth, falls to a 1 on customer diversification and through-cycle proof, and only Oracle scores 5 on durability and diversification. The honest read: NBIS has the best stack of structural advantages among the neocloud cohort but the worst track record under stress (because there has been no stress) and the thinnest customer base (two anchor counterparties). That is precisely the shape of a narrow, contingent moat.

6. Durability Under Stress

A moat only earns the label "moat" if it survives stress. Below are seven plausible stress cases mapped to NBIS-specific evidence.

The diagnostic: the moat survives the technology shift and NVIDIA demotion stress cases reasonably well because NBIS has a backup (multi-accelerator software, deep NVIDIA cross-investment). It does not survive the single-customer renegotiation or GPU spot-price collapse well, because both of those collapse the cycle that funds the build. A "narrow moat" is exactly the right rating: real, evidenced, but conditional on cycle and counterparty intent.

7. Where Nebius Fits

The moat — to the extent it exists — sits almost entirely inside one segment, one customer cohort, and one geography mix. The rest of the consolidated entity is option value, not moat.

The reader should walk away with this distinction: the moat is the Nebius AI Cloud segment, anchored on reserved-capacity contracts with two customers and owned greenfield power in a small number of geographies. Everything else is option value or accounting consolidation. Avride is a venture-stage AV business that happens to be inside the same legal entity. TripleTen is an edtech. The ClickHouse stake is a financial mark. None of those have anything to do with whether the AI cloud earns durable returns.

8. What to Watch

The six signals below tell a reader, in order of leading-indicator power, whether the moat is broadening or fading. The first one is the canary — everything else lags it.

The first moat signal to watch is contract tenor at renewal. Everything else is reactive: ARR, EBITDA margin, and capex all describe the output of a working moat. Tenor describes the input — it is the only signal that tells you whether the customers themselves still need the lock-in. A Microsoft or Meta renewal at five years (or longer) confirms the moat. The same renewal at one or two years prints "no moat" before any income statement does.

Financial Shenanigans

Nebius is a post-divestment, hyper-growth AI cloud rebuild of the legacy Yandex N.V. legal entity. Reported numbers are not fraudulent, but the reader should not take FY2025 economics at face value: operating cash flow is flattered by $982.5M of customer prepayments, pretax income is flipped from a $585M loss to a $14M profit by a $597M non-cash ClickHouse remeasurement, and the auditor flagged two material weaknesses in internal controls covering fixed assets and TripleTen revenue. The company is also mid-transition from Reanda Netherlands to Deloitte & Touche LLP and has extended server useful lives from four to five years effective January 2026, cutting FY2026 depreciation by an estimated $167.6M. Forensic risk score: 55 — Elevated.

The Forensic Verdict

Shenanigan Coverage Scorecard

Breeding Ground

The structural conditions for accounting strain are present: a founder-controlled, dual-class issuer freshly out of a complex divestment, audited by a small Dutch firm with limited PCAOB track record, scaling capex 60x in two years to honor multi-billion-dollar hyperscaler contracts. None of these on its own is disqualifying, but combined they explain why two material weaknesses re-emerged for FY2025 even after management remediated two of the three flagged the prior year.

The auditor history is the single most striking governance signal. Reanda Audit & Assurance B.V., a small Dutch firm engaged only after the 2024 divestment, issued a clean opinion on FY2024 financial statements but an adverse internal-control opinion. The board is now seeking shareholder approval to engage Deloitte for FY2026. That sequence — small firm signs the books, flags severe control weakness, then gets replaced by a Big Four — is the textbook breeding-ground pattern. It is not, by itself, evidence of wrongdoing; it is evidence that the controls environment has not yet caught up with the business.

Earnings Quality

Reported FY2025 net income of $82.5M (group) or $9.8M (continuing operations) is almost entirely a non-cash, non-operating story. Operating loss from continuing operations was $611.7M. The path from a $611.7M operating loss to a $13.8M pretax profit runs through three line items that an industrial cash-flow buyer would mark to zero: a $598.9M equity revaluation gain on ClickHouse, $55.2M of money-market-fund gains, and $27.2M of foreign exchange gains.

Without the $598.9M ClickHouse revaluation, FY2025 pretax income would be approximately −$585M. With that one entry, it is +$13.8M. The accounting is defensible — ASC 321 requires remeasurement on observable price changes when a third-party investment provides a fresh valuation mark — but ClickHouse just announced a Series D at a $15B valuation in January 2026, so a second large remeasurement gain will land in Q1 2026 results (already visible in the $621M Q1 2026 net income). Investors should value the operating business and the ClickHouse stake separately; conflating them in a P/E ratio overstates earnings power.

Revenue quality

Revenue growth of 479% in FY2025 is real (Microsoft and Meta capacity tranches were delivered on time). But the accompanying balance-sheet movement is unusual: receivables grew from $17.4M to $851.7M, a 4,795% increase, while accounts payable simultaneously grew from $228M to $1,210M. The MD&A attributes part of the receivable build to VAT receivables, contract assets, and the significant financing component on customer prepayments. That explanation is plausible, but the magnitude warrants reading the FY2025 20-F receivables disaggregation note before sizing a position.

The TripleTen material weakness is narrower but pointed: management acknowledges that IT general controls and business-process controls over TripleTen revenue recognition were not consistently documented, and that the segment represents approximately 10% of FY2025 revenue (~$53M). Compensating procedures were performed, but the underlying control gap remained at year-end.

Expense classification

Management flagged $43.6M of "equipment loss during transportation" in SG&A and excluded it from adjusted EBITDA as a one-off restructuring item. For a company that just deployed five new global data center locations in 2025 and plans nine more in 2026, transportation/installation losses of equipment are an operational hazard that may recur. Excluding it from EBITDA is acceptable disclosure but should not be assumed away in run-rate models.

The useful-life extension on servers and networking — from four to five years, prospective from January 1, 2026 — is GAAP-compliant and disclosed transparently. It will reduce FY2026 depreciation by $167.6M based on the FY2025 asset base. NVIDIA's H100/H200 hardware obsolescence cycle is closer to three to four years in practice; the five-year assumption is at the optimistic end of industry practice and should be benchmarked against CoreWeave (which uses six years for GPUs) and the hyperscalers (which extended useful lives in the past two years).

Cash Flow Quality

The headline of "$401.9M positive operating cash flow in FY2025" and the Q4 figure of "$834M positive operating cash flow" both deserve forensic re-statement. Two mechanisms drive almost all of that swing: a $982.5M lump of customer prepayments from the new Microsoft and Meta strategic contracts, and an $982M ($228M to $1,210M) increase in accounts payable as the company scaled procurement. Strip both lifelines out and FY2025 operating cash flow from continuing operations is approximately −$1.5B against a $4.1B capex bill.

The 2022 and 2023 group CFO and FCF numbers are dominated by the legacy Yandex Russian business (discontinued operations). The continuing operations cash-flow story is much cleaner: management's own MD&A reports continuing-ops CFO of −$222M (FY23), −$269.9M (FY24), then +$401.9M (FY25), with the FY25 flip almost entirely explained by the customer-prepayment mechanism and working-capital expansion. Free cash flow after capex from continuing operations was approximately −$3.66B in FY2025, funded by $4.16B of convertible note issuance and $1.15B of equity issuance.

Cash conversion will be tested in 2026

Q1 FY2026 already shows $399M in revenue and $2.47B in capex — a quarterly capex run-rate consistent with management's $20-25B FY2026 guidance. Management has stated that "cash inflows from our larger contracts throughout 2026 further underpin our growth plans," explicitly acknowledging that operating cash flow remains structurally dependent on continued upfront customer payments. The financing model is essentially: customers pre-fund a portion of the capacity they reserve, the company borrows / issues equity for the remainder, GPUs depreciate over five years, and the operating business runs at high adjusted EBITDA margins once at scale. This works arithmetically; it does not work if customers stop prepaying or if the take-or-pay obligations on Microsoft/Meta capacity reservations are renegotiated.

Metric Hygiene

The gap between GAAP net income from continuing operations and reported segment adjusted EBITDA is large but explainable: D&A of $417.9M in FY2025 alone closes most of it, with SBC of $83.2M and the $598.9M ClickHouse gain accounting for the rest. The non-GAAP framing is not abusive on its face, but the FY2025 change in SBC treatment (full exclusion vs partial in 2023-2024) and the use of "equipment loss" as a recurring-but-excluded item should be tracked. ARR is the metric most likely to mislead in 2026: a contract that ramps from 50% utilization to 100% over six months will look like 2x revenue growth in ARR terms even though the underlying contract has not changed.

What to Underwrite Next

This forensic profile is Elevated, not High — meaning the accounting risk is a position-sizing limiter and a valuation haircut, not a thesis breaker. Reported FY2025 net income of $82.5M should be ignored. The relevant numbers for underwriting are: operating loss of $611.7M from continuing operations, capex of $4.07B, customer prepayments of $982.5M, and a contracted backlog with Microsoft and Meta worth a combined $44-46B over five years.

The five items to monitor next:

1. Deloitte's FY2026 ICFR opinion, expected in the 20-F filed in spring 2027. A clean unqualified ICFR opinion would be the single largest upgrade signal. A second adverse opinion would move the forensic score to High.

2. Quarterly receivables and contract-asset disclosure. Track ending receivables and contract assets against trailing-three-month revenue. A ratio that stays above 300 days into mid-2026 — when Microsoft and Meta should be in steady-state servicing — would suggest collection risk, not just financing-component accounting.

3. CFO ex-prepayments. Each quarterly 6-K should disclose the customer-advance component. If the company stops prepayment disclosure, treat that as a stop-light red and re-derive ex-prepayment CFO from the change in contract liabilities note.

4. Useful-life impact. FY2026 quarterly depreciation will jump because of the capex base, but the rate per unit of PP&E should fall. If depreciation rate per dollar of net PP&E falls by more than the disclosed $167.6M effect implies, the useful-life assumption has been stretched further.

5. Adjusted EBITDA bridge. The "one-off restructuring" line includes the divestment-related expenses in 2023-2024 and the $43.6M equipment loss in 2025. If 2026 introduces a new "one-off" category greater than $25M, mark it as recurring in your own model.

The thesis is not "Nebius is cooking the books." The thesis is "Nebius is a high-capex, prepayment-funded ramp where the income statement, the balance sheet, and the cash-flow statement each carry one material distortion that goes the same direction." That is by definition an elevated forensic profile, and it argues for a smaller position than the growth narrative alone would justify, plus a valuation framework that disaggregates the operating business from the ClickHouse stake. Until Deloitte signs off and prepayments stop being the swing factor in CFO, treat reported earnings and reported operating cash flow as upper-bound estimates, not facts.

The People

Governance grade: B−. Nebius is a Controlled Company in which founder Arkady Volozh wields 51.6% of votes through a single family trust while owning only 11.3% of economics — the board has real industry talent, but the CEO chairs the nominating committee, the auditor is a small Amsterdam firm (Reanda), and every recent insider transaction has been a sale.

1. The People Running This Company

The team is unusual: a founder-CEO with a 25-year track record of building a top-three global search engine, paired with a deliberately external bench — the CRO came from Cloudflare and Twilio, the CFO from European e-commerce, the GC from Israeli industrials. The three co-founders (Volozh, Chernin, Korolenko) and COO Nave each hold 625,000 "premium-priced" options struck at $100 that vest only if the stock holds above strike — a clean alignment instrument given the stock currently trades materially above that level. The gap to watch is succession: there is no announced deputy CEO, and Volozh's voting control means no external party can force one.

2. What They Get Paid

Cash compensation is unusually light for a US-listed AI infrastructure company of this scale — roughly $740K per senior executive on average, well below S&P 500 tech peers where CEOs alone routinely clear $10M cash + bonus. Equity, however, is heavy: 6.35M options and 226K RSUs against ~254M total shares outstanding equals ~2.6% annual stock-based dilution to senior management, and the plan reservation tops out near 33M shares (~13% of float). The structural choice — modest cash, dense equity, four-year quarterly vesting, premium-priced options requiring stock to hold above $100 — is shareholder-friendly in design. The caveat is disclosure: as a foreign private issuer, Nebius reports aggregate compensation only. There is no Summary Compensation Table, no CEO-pay-ratio, no clawback policy disclosed in the proxy. Investors cannot tell what Volozh personally earned in cash or grants.

3. Are They Aligned?

Ownership and Control

Volozh's family trust holds 28.66 million Class B shares, each carrying ten votes versus the one vote on each Class A share. The result: the founder controls a majority of votes with roughly one-ninth of the economic equity. Nasdaq formally classifies Nebius as a Controlled Company, which exempts it from the requirement that a majority of directors and the entire nominating committee be independent. Volozh has used both exemptions. The economic stake — about $5.8B at current prices — is large enough to align him with shareholders on outcome, but the voting structure means a minority investor has no procedural recourse to disagree with him.

Insider Activity

Every reported insider transaction in the last six months has been a sale, executed under Rule 10b5-1 plans pre-arranged with the company. That format mutes the negative signal — these are not opportunistic trades — but the absence of any open-market buying, in a stock that has roughly tripled from the relisting price, is conspicuous. Volozh's $3.5M April sale is also tiny relative to his economic stake (~0.06% of his shares), reinforcing the read that this is portfolio diversification rather than a vote of no-confidence.

Dilution and Capital Allocation

The December 2024 private placement brought in $700M with NVIDIA (0.5% stake) and Accel (3.1%) as anchors — both strategic. A reported NVIDIA $2B follow-on announced in early 2026 reinforces the commercial alignment with the largest GPU supplier in the company's supply chain. There are no buybacks (none expected at this stage), no dividends, and no related-party transactions disclosed — the 20-F note on RPTs reads simply "None", which for a controlled-company structure is a meaningful green flag.

4. Board Quality

The board's professional depth is genuine: an Accel partner on the venture-funding side, a De Brauw partner for Dutch corporate compliance, a finance veteran chairing audit, and an AI researcher with operator credentials chairing compensation. The structural problem is concentrated in two places. First, Volozh chairs the nominating and corporate governance committee while serving as CEO — he selects the people who evaluate him. This is permitted under Controlled Company rules but is the single sharpest governance weakness in the proxy. Second, the audit firm is Reanda Audit & Assurance B.V., an Amsterdam-based non-Big-4 firm retained when the company relisted. For a US-listed company carrying $12.4B in assets and operating data centers across three continents, a Big-4 (or even a tier-2 international firm such as Grant Thornton or BDO) would carry more credibility with global institutional holders.

5. The Verdict

Governance grade: B− — founder skin-in-the-game and clean related-party note offset by Controlled-Company voting structure, CEO-chaired nominating committee, and non-Big-4 auditor.

The case nets to B−. Nebius has the rarest asset in governance: a founder with a 25-year operating record who has explicit economic skin in the game at billion-dollar scale, paired with a senior bench hired from credible external companies (Cloudflare, Twilio, Booking, Azerion). The 20-F's clean related-party note is unusual for a controlled company and is the strongest single signal that economic value is not being siphoned. What pulls the grade down is structural: a single trust holds majority voting power, the CEO chairs the committee that picks his board, and the auditor is small enough that a Big-4 escalation would meaningfully change the credibility of the financials. None of these are dealbreakers — they are repairable. The one to watch most closely is the auditor; the second is whether Volozh ever steps off the nominating committee. Either move would justify an upgrade to B+.

The story, and how it changed

Few public companies have rewritten their identity as completely as this one. In four years, the same legal entity went from a $20B Russian internet conglomerate (Yandex N.V.), to a frozen, sanctioned, delisted shell, to a $2.5B cash-rich shell stub, to one of the few sub-hyperscaler AI clouds with multi-year Microsoft and Meta contracts and a $2B Nvidia equity stake. The narrative arc is unusually clean: management telegraphed the strategy at relisting in October 2024, missed its first ARR data-point three months later, then beat or raised every subsequent number through Q1 2026. The current story is simpler than at any point in the company's history — full-stack AI cloud, owned gigawatt-scale capacity, hyperscaler-anchored revenue — but it is also more capital-stretched, more accounting-sensitive, and more concentrated than the post-divestment communications acknowledge.

1. The narrative arc

Five inflection points carry the entire story. The chart below treats them as annotated milestones along a single axis; everything else is connective tissue.

Underneath those milestones, the financial trajectory tells a single story: revenue went from rounding-error to a third of a billion in five quarters, while capex went from already-aggressive to vertical.

The ARR line is the most-watched number in the bull case. Dec-2024's $90M was the only break in trend; from $249M at end-Q1 2025 it has compounded roughly 47% per quarter — a pace driven first by AI-native startups, then by hyperscalers as Microsoft and Meta capacity was lit up.

2. What management emphasized — and then stopped emphasizing

Reading filings and shareholder letters in sequence, certain words are loud, then quiet, then absent. The heatmap below scores the prominence of each theme by year (0 = absent, 1 = mentioned, 2 = central, 3 = dominant) based on the FY2021–FY2025 20-Fs and the eight quarterly shareholder communications since relisting.

Three quiet pivots are worth flagging because they reframe how the equity should be valued, not just how it is described:

• "Neocloud" framing has been retired. Through mid-2025 management described Nebius alongside CoreWeave, Crusoe, and Lambda. From Q3 2025 onward the comparison set is Azure-tier; the FY2025 20-F lists Microsoft, AWS, Google, and Oracle as competitors first, the bare-metal peers second.
• Toloka quietly exited the consolidated story. In Q3 2024 management called Toloka "the second largest contributor to the Group's total revenue." By Q2 2025 it was deconsolidated after a Bezos Expeditions-led investment, and prior-period figures were recast into discontinued operations. The minority stake is now a sidebar.
• Bare-metal language has been replaced by software-led framing. Aether 3.0/3.1/3.5, Token Factory, and the Tavily/Eigen AI/Clarifai acquisitions in Q1 2026 all reposition the equity as a platform-as-a-service story with higher-margin attach revenue rather than a GPU-rental business.

3. Risk evolution

The same legal entity has filed 20-Fs as a Russian internet conglomerate (FY2021–FY2023), a transitional stub (FY2024), and a pure-play AI cloud (FY2025). The risk-factor summary table changed wholesale in each step.

The two newly visible risks are the ones to monitor. Customer concentration is a 2025 invention in the risk language — and the FY2025 20-F now devotes a dedicated risk-factor to "limited experience in delivering, implementing and managing longer-term customer contracts," explicitly naming Microsoft and Meta. Power and capacity moved from generic operational risk in FY2024 to the top of the 2025 risk stack as the company contracted past 2 GW.

The most striking disappearance is convertible-note default, which dominated FY2022 disclosure (Nasdaq halt triggered redemption rights on the $1.25B 2025 notes) and is now gone entirely. The 2022 settlement that delivered ~5.7M Class A shares to former noteholders is a footnote in the FY2024 transition document.

4. How they handled bad news

There has really been one meaningful miss since relisting, and one accounting choice worth pressure-testing. Both are worth showing in management's own words.

The Russia-divestment communications across 2022–2024 are harder to grade because legal constraints limited what management could say while negotiations were active. The disclosure when the divestment closed in July 2024 was reasonably complete, and the transition to USD reporting (with prior periods recast as if it had happened on Jan 1, 2022) was the right call. There has been no scandal, restatement, or SEC investigation in the public record.

5. Guidance track record

Management has issued seventeen specific, verifiable guideposts since the October 2024 relisting. Sixteen have been hit or beat — and many were raised between issuance and outcome. The single miss (Dec-2024 ARR) is the one already discussed.

The pattern of "guidance set, then raised, then beaten" is most visible in the contracted-power line, which is now the only number management updates more often than ARR.

Why 8 rather than 9. Three caveats keep the score below the top tier: (i) capex run-rate has run more than double the original 2025 $2.0B guide, even allowing for hyperscaler-driven upside — meaning the funding plan keeps re-baselining; (ii) the Q1 2026 useful-life extension materially flatters the headline adj EBITDA margin and was not flagged in the CEO commentary; (iii) the medium-term 20–30% EBIT margin target was set when GPU economics, depreciation policy, and customer mix were all materially different and remains untested. None of these is dishonest, but each is the kind of seam that becomes important in a downturn.

6. What the story is now

The current Nebius story is, in three sentences: a vertically integrated AI cloud with two multi-year hyperscaler anchors (Microsoft ~$17–19B over five years; Meta ~$30B+ across two contracts), gigawatt-scale owned-capacity pipeline (>4 GW contracted by end-2026, >75% in owned facilities), and a deepening Nvidia relationship sealed by a $2B equity investment in May 2026. Adj EBITDA turned positive in Q4 2025 at the group level and in Q2 2025 at the cloud business; the company exited Q1 2026 with $9.3B of cash after raising $6.3B in a single quarter.

Financials — What the Numbers Say

Read this section together with the Industry and Business tabs. NBIS is the renamed/restructured continuation of Yandex N.V. after the July 2024 sale of its Russian assets. Continuing-operations revenue effectively starts in FY2024, so almost every financial number on this page describes a company that, in commercial terms, is two years old.

1. Financials in One Page

Nebius is an early-stage AI-infrastructure roll-up that has bolted a hyperscale-grade balance sheet onto a sub-$1B-revenue cloud business. FY2025 revenue of $530M (up ~6x YoY) and Q1 2026 revenue of $399M (up 684% YoY, 75% QoQ) reflect a step-function capacity build, not steady-state operations. Operating margin is deeply negative (-115% in FY2025), but Q1 2026 marked an inflection: group adjusted EBITDA flipped to +$130M (32% margin) and Nebius AI's segment EBITDA margin reached 45%. Reported free cash flow is sharply negative — -$3.7B in FY2025 on capex of $4.1B — and management funded the gap with $4.2B of late-2025 convertibles, an additional $4.3B of Q1 2026 convertibles, and a $2B equity investment from NVIDIA in Q1 2026, leaving an end-Q1 cash position of $9.3B alongside about $4.9B of YE2025 reported debt (LT debt rises to ~$8.4B on a Q1 2026 basis after the new converts). Valuation is the central tension: ~$52.6B market cap (May 2026) implies ~16x forward revenue at guidance midpoint and ~6–7x exit-2026 ARR, prices the business as a tier-1 AI compute platform, and leaves zero room for capacity slippage. The financial metric that matters most is annualized run-rate revenue (ARR) versus the $7–9B year-end-2026 guide — that single number sets whether the price is rational or stretched.

---

2. Revenue, Margins, and Earnings Power

Plain-English primer. Revenue is the cash value of cloud capacity sold. Gross profit is what's left after the direct cost of running GPUs, power, and bandwidth. Operating income is what's left after Nebius pays its engineers, salespeople, and administrative staff. Operating margin = operating income ÷ revenue; for a software-like business it should be positive at scale; for an infrastructure business in build-out mode it is usually deeply negative.

Annual revenue and operating income (FY2022–FY2025)

The legacy Yandex P&L was reset to near-zero in FY2022 once Russian operations were classified as discontinued. The chart below is the continuing-operations view that ties to today's company.

Revenue scaled 5.8x in FY2025 to $530M, but operating losses widened from $400M to $612M as the company hired ahead of revenue and absorbed the depreciation of new GPU fleets coming online. This is the expected shape of a capacity buildout: opex shows up before revenue does. The relevant question is whether each subsequent quarter narrows that gap.

Quarterly revenue trajectory — the actual story

Three observations:

1. The 75% sequential jump in Q1 2026 is faster than any prior quarter, which is unusual at this scale and reflects new owned-capacity sites coming online plus utilization-driven uplift on existing fleets.
2. Q1 2025 was the only sequential decline — caused by an end-of-quarter timing effect on capacity activation, not lost customers. The trend reasserted in Q2 2025.
3. Group revenue and Nebius AI revenue are now ~98% aligned ($399M vs $390M in Q1 2026), so the consolidated number is effectively a cloud number with two tiny early-stage attachments (Avride and TripleTen).

Annualized run-rate revenue (ARR) and the FY2026 setup

ARR is calculated by multiplying Nebius AI's last-month-of-quarter revenue by twelve. It is the cleanest read on the run-rate the business will exit the year at, and it is what management guides on.

Management's FY2026 guide is $3.0B–$3.4B in reported revenue and $7B–$9B exit ARR, alongside ~40% adjusted EBITDA margin. The implied math: ARR exits the year at roughly 2.4x reported revenue, which only works if a large block of new capacity activates late in 2026. That is consistent with the company's stated power-online schedule.

Margins: where the inflection is happening

Gross margin moved from negative to 69% in FY2025, which is a credible cloud-infrastructure level. Operating margin remained deeply negative because most depreciation hits the income statement before utilized revenue ramps. Q1 2026 is the first quarter to show real leverage: gross margin 74%, adjusted EBITDA margin 32% at group level and 45% at the Nebius AI segment. If that EBITDA margin holds, the FY2026 ~40% adjusted EBITDA guide is reachable.

---

3. Cash Flow and Earnings Quality

Plain-English primer. Operating cash flow (OCF) is the cash a business produces from running its operations, before paying for new equipment. Capex is the cash spent on new long-lived assets (GPUs, data center build-outs). Free cash flow (FCF) = OCF + capex. For a capacity-buildout business, FCF can be very negative even when OCF is positive — that is the case here.

Net income vs operating cash flow vs free cash flow

The FY2022 and FY2023 figures embed gain-on-sale and discontinued-operations effects from the Yandex divestiture — they should not be read as cash earnings power. The two clean years are:

• FY2024: OCF +$246M, capex -$808M, FCF -$562M. Reported net income -$641M was distorted by buyback-related items; OCF is the cleaner read.
• FY2025: Net income flipped to +$83M largely from investment gains and divestiture residuals; OCF was +$385M; capex of -$4,066M (5x prior year) drove FCF to -$3.7B.

Free cash flow conversion is structurally negative — that's the design

Capex/revenue ran at ~770% in FY2025 — that is, every dollar of sales was matched by roughly $7.70 of investment in new GPU fleets, networking, and owned data-center build-outs. By contrast, mature hyperscalers run capex/revenue at 10–25%. The ratio will compress as revenue scales, but the company will be capex-heavy for at least three more years.

Cash-flow distortions to watch

Three things matter here:

1. SBC of $83M is ~16% of FY2025 revenue. That is high in absolute terms but lower than the 25–40% typical of pre-profit software peers. If the company hits its scale guide, SBC dilution becomes a smaller share of equity.
2. Q1 2026 reported $2.3B of operating cash flow despite revenue of only $399M — this is almost certainly inflated by customer prepayments on multi-year contracts (Microsoft, Meta, Eigen-related). Treat this as financing-like, not recurring OCF.
3. D&A is set to accelerate. PP&E grew 7x in FY2025 to $6.5B. Trailing depreciation of $469M will likely double over the next 12 months, pressuring reported operating income even as cash margins expand.

---

4. Balance Sheet and Financial Resilience

Plain-English primer. Net debt = total debt minus cash. Net debt / EBITDA is the standard leverage gauge; <2x is conservative, >4x is stretched for an operating business. For a build-out company that pre-funds capex, cash runway matters more than the leverage ratio in a single year.

Cash, debt, and equity evolution

The FY2025 jump in debt is the $4.2B convertible senior notes raised in late 2025 alongside a concurrent Class A share offering. Combined with +$2B from NVIDIA's strategic equity investment in Q1 2026 and additional convertibles, end-Q1 2026 cash was $9.3B. That is roughly two-and-a-half years of capex at the FY2025 run rate without any operating contribution — substantial runway.

Liquidity, leverage, and resilience snapshot (FY2025)

Interest coverage screens poorly (EBITDA/interest of -2.3x in FY2025) — but that is a feature of the build phase, not a stress signal. The relevant questions are:

• Does the cash runway carry the company to operating-cash-flow breakeven? At FY2026 guide ($3.0–3.4B revenue, ~40% adj EBITDA), cash EBITDA will roughly cover ongoing capacity replacement capex even as growth capex continues to be funded externally.
• Are the convertibles equity-like or debt-like? They are zero-coupon convertibles with conversion premiums set above current prices at issue; if NBIS continues to compound, dilution is the more likely outcome than cash repayment.
• Is there off-balance-sheet capacity commitment? Yes — the company has contracted >4 GW of power and >100 MW of multiple new sites. Future GPU and power purchase commitments are large and disclosed in the 20-F; investors should size these alongside on-balance-sheet debt.

PP&E grew 7.3x in a single year. The risk is asset stranding if GPU prices fall faster than depreciation schedules assume; the opportunity is that owned capacity (now >75% of contracted power) gives margin leverage that lease-and-resell competitors cannot match.

---

5. Returns, Reinvestment, and Capital Allocation

Plain-English primer. Return on invested capital (ROIC) answers "for every dollar of capital deployed, how much annual profit do we earn?" Return on equity (ROE) answers the same question from a shareholder lens. Both are negative or noisy here because the denominator (invested capital) has scaled faster than the numerator (operating profit) — by design.

Returns are still informational, not investable

Returns will remain negative or noisy through at least FY2027. The investable question is the exit ROIC once capacity fills: at a 40% adjusted EBITDA margin on $8B exit-2026 ARR, theoretical EBITDA run-rate is ~$3.2B against invested capital trending toward $20–25B. That implies a steady-state EBITDA-yield-on-invested-capital of ~13–16%, which is healthy if depreciation lives are honest and if pricing power holds.

Capital allocation — how Nebius is spending its capital

There are no dividends, no buybacks, no debt paydown — essentially every dollar (and every dollar borrowed and raised) is going into capacity. That is the correct strategy if and only if demand matches the buildout schedule. Two recent signals that demand exists:

• Microsoft, September 2025: $17.4B multi-year capacity agreement
• Meta, April 2026: ~$27B five-year capacity agreement (referenced in Morningstar fair-value revision)
• NVIDIA, Q1 2026: $2B strategic equity investment (signals technical alignment and supply priority)
• Eigen AI acquisition, May 2026: $643M to add inference-optimization technology

Share count: dilution is real but partly offset by capacity-driven per-share value

The drop in FY2024 reflects the Russian-business carve-out cancellation of legacy Yandex Class A shares. Since then, period-end share count has crept up from 236M to 254M as the company has raised equity. The bigger overhang is in diluted-share count: Q1 2026 diluted basic was 258M but fully diluted share count was 309M, reflecting potential issuance from convertible notes and NVIDIA-linked instruments. That is a ~20% buffer between basic and diluted — meaningful but understood.

---

6. Segment and Unit Economics

Nebius reports three operating businesses, but one (Nebius AI cloud) carries effectively all of the consolidated revenue and economics today.

Group revenue mix — Q1 2026

Where the value is.

• Nebius AI Cloud — the GPU cloud / AI infrastructure business. Q1 2026 revenue $390M, ARR $1.9B, segment adjusted EBITDA margin 45%. This is the entirety of the investment thesis.
• Avride — autonomous driving and delivery robots. Reportedly valued at $15B in a January 2026 fundraise (strategic investment from Uber). Revenue is small but the equity stake is a real option.
• TripleTen — edtech reskilling business. Revenue grew 10% YoY in Q1, much smaller than Nebius AI.
• Equity stakes — ClickHouse and Toloka are minority positions, but ClickHouse in particular is now a multi-billion-dollar database company and the carrying value is potentially understated.

Geography concentration

Nebius does not break out revenue by geography in summary form, but disclosed data-center capacity is concentrated in:

Why this matters financially. Roughly 75% of contracted power is now owned rather than leased — important because owned capacity removes a third-party landlord margin from the P&L. That is the structural reason the segment EBITDA margin moved from 24% to 45% in one quarter.

---

7. Valuation and Market Expectations

This is the most important section in the page. The financial profile is improving, but valuation is the active risk.

What the market is paying today

The single most informative multiple is EV / exit-2026 ARR ≈ 6.6x at the midpoint guide, falling to ~5x at the high end of the ARR range. CoreWeave (the closest public peer) trades at roughly 18–22x trailing revenue with comparable hyper-growth dynamics. By that yardstick NBIS is cheaper than CoreWeave on forward ARR — but that comparison only holds if the $7–9B ARR guide is met.

Historical valuation context

Trailing EV/Sales multiples are noisy because revenue was reset post-divestiture. The honest read: Nebius has never traded on trailing earnings. It trades on capacity contracts and ARR.

Analyst expectations and price-target dispersion

The price-target dispersion of $126 to $291 — a 2.3x range — is itself the story. It tells you the bull case (full ARR delivery, hyperscaler-class economics) and the bear case (capacity slippage, multiple compression) are both rational; the difference is execution.

Simple bear / base / bull frame

Bear assumes ARR plateaus mid-range, multiple compresses to CoreWeave-like 4x on missed guidance. Base roughly maps to the current price assuming guidance is met. Bull requires both upper-end ARR delivery and a multiple re-rating toward hyperscale-class. At today's price, the market is paying for the base case. To make money from here, you need to believe in something closer to the bull case.

---

8. Peer Financial Comparison

Public peers fall into three camps: (1) the direct AI-cloud comp, CoreWeave; (2) the hyperscaler reference point, Oracle (OCI); and (3) the power-first HPC pivot stories — IREN, Applied Digital, and Hut 8.

What the table says.

• CoreWeave is the closest peer, with roughly 10x the trailing revenue ($5.1B vs $530M) and a market cap only ~20% larger ($62B vs $52.6B). CoreWeave is FCF -$7.3B, NBIS is FCF -$3.7B — both are burning cash, but CoreWeave is carrying nearly 3x the leverage ($14.7B debt vs $4.9B). On growth, NBIS is now growing faster (684% vs 168% YoY in latest reported period).
• Oracle is the only peer with real FCF discipline, but it grows at 9% — entirely different business. Useful as a "where the multiples eventually settle" reference, not as a near-term comp.
• IREN, APLD, HUT are direct power-first comps. APLD trades at ~90x trailing EV/sales because revenue is still re-pivoting; IREN and HUT are cheaper but lack NBIS's pure-AI revenue mix.

Peer positioning chart

NBIS sits in the high-growth, high-multiple quadrant — closest in shape to CoreWeave, with the highest growth rate among the comp set.

The peer-gap verdict. NBIS earns a premium for (1) faster growth, (2) lower leverage than CRWV, (3) the optionality of equity stakes and Avride, and (4) a more vertically integrated product stack with owned compute IP. The premium is fragile: it disappears if Q2 or Q3 of 2026 shows any sequential ARR deceleration. NBIS at this price is cheaper than CoreWeave on forward ARR but more expensive than IREN or APLD on trailing sales — defensible only if execution holds.

---

9. What to Watch in the Financials

Closing read

What the financials confirm is the inflection: gross margin moved from negative to 69%; segment adjusted EBITDA margin jumped from 24% to 45% in a single quarter; cash and contracted power both expanded materially; and customer commitments now read like a hyperscaler client list (Microsoft, Meta, NVIDIA partnership). The business is no longer an early-stage venture, even if its income statement looks like one.

What the financials contradict is the easy bull case. Reported earnings are negligible. Free cash flow was -$3.7B in FY2025 and will be worse in FY2026 before it gets better. Asset stranding risk on a 7x PP&E expansion is real. Diluted share count is 20% above basic. The trailing EV/sales multiple of ~99x is only defensible if you swap "trailing" for "forward exit-ARR."

The first financial metric to watch is the Q2 2026 ARR print. If end-June ARR breaks $2.5B (about 30% QoQ growth), the FY2026 $7–9B exit guide becomes a layup and the stock can compound through the year. If end-June ARR comes in below $2.2B (15% QoQ growth or less), the bull case starts cracking — capacity activation has slipped or pricing is under pressure, the $7B low-end becomes the over-under, and the multiple compresses. That single number, due in early August 2026, will define the rest of the year.

The Bottom Line from the Web

The web reveals two facts the filings don't put on a single page. First, Nebius dismissed its small-tier auditor (Reanda Audit & Assurance B.V.) and is replacing it with Deloitte for FY2026 — announced February 12, 2026, following an adverse internal-control-over-financial-reporting (ICFR) opinion on 2024 financials (TipRanks; Seeking Alpha Jan 16, 2026). Second, the $27B Meta agreement signed March 16, 2026 is structurally unusual — only ~$12B is a firm compute commitment; ~$15B is capacity Nebius can resell to Meta or to other customers at market rates, which means the headline "$46B backlog" overstates contracted revenue (CNBC, Mar 16, 2026; Q1 2026 letter via DelMorgan). Both items materially change how an investor should read the contracted-revenue narrative and the audit quality grade.

What Matters Most

1. Auditor change: Reanda dismissed, Deloitte proposed — after an adverse ICFR opinion on FY2024

2. Meta $27B deal is half firm, half optional

3. Microsoft $17.4B–$19.4B five-year deal anchored the rerating

Nebius announced its first hyperscaler anchor: a five-year, $17.4B GPU-capacity agreement with Microsoft, with options to reach $19.4B total. Shares jumped ~47% after the bell on September 8–9, 2025. Microsoft CFO Amy Hood had told investors Microsoft expected to remain capacity-constrained through 2025; Nebius will deliver dedicated GPUs from a new Vineland, NJ datacenter starting late 2025. Sources: Reuters via Fast Company; Redmond Channel Partner (Sep 12, 2025).

4. Q1 2026 print: revenue +684% YoY, but $621M net income flattered by ClickHouse mark-up

5. Capex plan $16B–$20B for 2026 vs. 2026 revenue guide $3.0B–$3.4B — funding gap is real

Management told the Q4 2025 call: 2026 capex of $16B–$20B, with ~60% funded, the remainder to come from new debt, asset-backed financing or equity. Sources: MarketBeat Q4 2025 transcript notes; Reuters via Yahoo Q1 2026 capex coverage. To bridge the gap, Nebius closed a $4.3B convertible offering on March 20, 2026 (upsized from $3.75B; 2031 notes at 1.250%, 2033 notes at 2.625%), on top of a separate ~$4.2B Sept 2025 raise. Source: BusinessWire (Mar 23, 2026).

6. NVIDIA's $2B equity stake — strategic alignment but a confounder for "independence" claims

NVIDIA invested $2B in Nebius in March 2026, deepening a relationship that began with NVIDIA's ~$33M stake in the December 2024 $700M PIPE (~0.5% then). Nebius said it will be among the first NCPs to offer the next-gen NVIDIA Vera Rubin NVL72 (H2 2026). Sources: The Next Web; Nebius newsroom (Jan 5, 2026).

7. ClickHouse stake ~28%; Series D priced at $6.35B (later $15B) — a hidden SOTP asset

8. CEO Volozh + insiders selling actively — $17.7M over trailing 90 days

9. Short interest 21.05% of float — and a +522% one-year run

NBIS short interest reached 43,138,451 shares (21.05%) as of April 30, 2026 (ChartExchange). Combined with a +522% one-year total return and 1,642% turnover, positioning is unusually crowded on both sides — squeeze potential coexists with elevated overhang from convertible strikes and forward dilution.

10. Analyst dispersion is extreme: $110 low to $291 high; consensus Overweight/Buy

WSJ research ratings show 11 Buy, 1 Overweight, 4 Hold, 0 Underweight, 1 Sell (current); 12-month price targets span $110 (low) to $291 (high), median $205, average $197.46 (current price $207.27 on May 13, 2026). Notable initiations: Cantor Fitzgerald started coverage at Overweight, $129 PT (March 2026 per The Fool); D.A. Davidson at $250. Sources: WSJ; Barron's; Motley Fool (Apr 10, 2026).

11. GPU useful-life extended to 5 years — peer comparison hidden in plain sight

NBIS extended its GPU/server useful life from 4 to 5 years in Q1 2026, lifting reported margins. CoreWeave depreciates the same equipment over 6 years (per its S-1). NBIS remains more conservative than CRWV but less aggressive than the prior 4-year baseline. Sources: Bizety (Sep 23, 2025) GPU Depreciation: CoreWeave vs. Nebius; Q1 2026 transcript via Motley Fool.

12. Eigen AI $643M acquisition (May 2026) — moves the IaaS-to-PaaS pivot from talk to spend

Nebius agreed to acquire Eigen AI for ~$643M on May 1, 2026 to deepen Token Factory inference. Stock rose ~21.7% on the news (Datagrom / SimplyWallSt). Source: Datagrom; TheAIWorld; SahmCapital. Press also reports Nebius weighing an AI21 Labs acquisition (last valued $1.4B in 2023) per The Information / SimplyWallSt.

Recent News Timeline

What the Specialists Asked

Governance and People Signals

CEO Arkady Volozh (founder). Resigned from Yandex N.V. June 2022 after EU sanctions; delisted from EU sanctions March 2024; returned as CEO of Nebius Group August 2024. Owns 11.53% (~29.5M shares); recently trimmed 33,358 shares on April 1, 2026 at ~$103.73 — modest in context of total stake, but the first publicly reported CEO sale post-relisting. CEO tenure 1.8 years (Simply Wall St). Compensation figures not publicly broken out.

Board independence. No ISS or Glass Lewis recommendations specific to the 2025 AGM, Volozh's nominating-committee role, or new directors Grimme and Weigand surfaced in the dataset.

Insider sales — past 90 days.

Net pattern: most sales are pre-planned (10b5-1) with plans adopted in late 2025, before the most recent rally — consistent with diversification rather than front-running. Total trailing-90-day insider dispositions ≈ 146,441 shares / ~$17.7M per Blockonomi compilation.

Audit committee. Dismissed Reanda Audit & Assurance B.V. and proposed Deloitte & Touche LLP for FY2026, conditional on shareholder approval. The dismissal follows Reanda's adverse ICFR opinion for FY2024.

Ownership concentration. Fiscal.ai top-holders snapshot lists Volozh at 11.53% and institutions including BlackRock, Geode, Alger, UBS, Accel, Morgan Stanley, Goldman Sachs, State Street, Janus Henderson, Capital Research, and Invesco — specific share counts beyond Volozh not exposed in public-API view.

Industry Context

Demand backdrop. Reuters and CNBC report Microsoft, Amazon, Meta, and Alphabet plan ~$630B aggregate 2026 datacenter capex, with ~70% flowing to NVIDIA chips. Microsoft CFO Amy Hood publicly noted capacity-constraint through 2025. The structural picture: hyperscalers are routing a meaningful slice of incremental capacity to third parties (NBIS, CRWV) to keep capex off their own balance sheets.

Supply-side bottleneck. Power interconnect queues are the binding constraint, per multiple sources (RMI on PJM, Nov 2025; UtilityDive on ERCOT/CAISO/PJM/ISO-NE scorecard). PJM is the worst-affected; ERCOT/CAISO score best. NBIS's announced 310 MW Finland buildout and Kansas City / Vineland NJ campuses route around US-grid friction.

Pricing data point. Independent comparison (Spheron, Mar 2026) shows NBIS public on-demand: H100 $2.95/hr, H200 $3.50/hr, B200 $5.50/hr, L40S from $1.55/hr. Reserved-contract pricing for hyperscalers is not public; the $17.4B Microsoft deal implies ~$3.5B/yr base run-rate but ramp is back-loaded.

Regulation. The BIS IaaS KYC rule (rooted in EO 13984, January 2021) remains pre-final. If enforced, it would impose KYC obligations on US-touching cloud providers and could raise onboarding cost across the neocloud cohort. Compliance-industry coverage (Voveid) suggests enforcement is "being assembled in real time."

Depreciation policy comparison. CoreWeave depreciates GPU/server equipment over 6 years (S-1). Nebius used 4 years through FY2025 and extended to 5 years in Q1 2026. The 5-year choice sits between CRWV's most aggressive bookend (6) and the historic conservative (4) — material to gross margin comparability.

Hyperscaler in-sourcing risk. No direct evidence that Microsoft is dialing back third-party neocloud reservations; the September 2025 Microsoft contract argues the opposite. Applied Digital's largest tenant for the 250 MW Polaris Forge 1 leases is CoreWeave, not Microsoft/Meta — so the "BTC-pivot peer" risk is moderate, not high.

Where We Disagree With the Market

The market is pricing Nebius against a $46B contracted backlog, a 45% Q1 segment EBITDA margin, and a "second-source hyperscaler" narrative validated by Microsoft, Meta, and NVIDIA — and the consensus 12-month PT of $205 (vs $207 spot) says fair value is roughly the base case. Our quarrel is not that the bull is wrong on direction; it is that three of the load-bearing inputs the market quotes are quietly weaker than printed: roughly $15B of the $46B backlog is a flex resale option (not firm take-or-pay), at least ~10pp of the Q1 segment-margin step embeds a same-quarter useful-life accounting change rather than pure operating leverage, and the $401.9M FY2025 / $2.3B Q1 2026 operating cash flow numbers cited as "self-funding" are essentially customer prepayments and AP build — strip both and FY2025 organic CFO from continuing ops is approximately −$1.56B against $4.1B of capex. The disagreement is testable, not philosophical: each line resolves inside a single earnings cycle, and the Q2 2026 ARR print on July 29 is the primary disconfirming event.

Variant Perception Scorecard

The score earns its 72 because each disagreement is anchored on a disclosed number (the Meta firm/flex split per the Q1 2026 letter; the $167.6M useful-life depreciation impact per the FY2025 20-F; the customer-prepayment magnitude per the cash-flow statement). Consensus is unusually clean here — 11 Buy / 1 Overweight / 4 Hold / 1 Sell with a median PT of $205 — which makes the gap to evidence easier to measure than usual. Time to resolution is short: the Q2 2026 ARR print, the first asset-backed financing terms, and a partial backlog firm/flex re-disclosure should all land inside 6 months.

Consensus Map

The market view here is unusually crisp — a Strong Buy consensus, a stock at fresh all-time highs after a Q1 blowout, and a sell-side band whose median PT sits within a dollar of spot. The table below pins down each market belief to a single observable consensus signal so disagreement can be measured, not asserted.

The Disagreement Ledger

Three disagreements pass the materiality + testability bar. The first is the one we would defend to the death; the second is the cleanest non-consensus accounting read; the third is a different cycle template than the one the multiple reflects.

Disagreement #1 — The $46B backlog is closer to $29-31B firm. Consensus would say: "Microsoft and Meta together signed ~$44B of multi-year reserved capacity. That is contracted revenue with hyperscaler counterparties; the backlog-to-revenue ratio of 87x is unmatched in cloud." The evidence disagrees because the Q1 2026 shareholder letter discloses that ~$15B of Meta's $27B is optional capacity Nebius can resell at market rates to other customers — not a take-or-pay obligation. The Microsoft contract is similarly structured with an embedded options layer ($17.4B base, up to $19.4B with options). If we are right, the market would have to concede that contracted visibility is roughly 60-65% of the printed number, that the switching-cost moat narrative depends on Meta exercising flex (not signing it), and that EV/firm-backlog is closer to 1.7x than 1.1x. The cleanest disconfirming signal is Meta exercising the full $15B flex tranche at or above market price within the next 12 months — that would reclassify the optional capacity as firm and validate the consensus print.

Disagreement #2 — At least 10pp of the Q1 45% segment EBITDA is a useful-life accounting choice. Consensus would say: "Segment Adj EBITDA stepped from 24% in Q4 2025 to 45% in Q1 2026 because owned-capacity utilization scaled — operating leverage at work." The evidence disagrees because the same quarter, Nebius prospectively extended server useful life from 4 to 5 years effective January 1, 2026, cutting FY2026 depreciation by a disclosed $167.6M on the FY2025 asset base alone. The quarterly run-rate of that change against Q1 Nebius-segment revenue is roughly 10.7pp of margin — meaning a normalized segment margin is closer to ~34%, not 45%. If we are right, the market has to concede that the "hyperscaler-class economics" framing has not yet been earned on a like-for-like basis, that CoreWeave's 6-year and the hyperscalers' 6-year lives are the relevant comp (not the prior 4-year baseline), and that the operating-leverage thesis needs at least two more quarters of confirmation. The disconfirming signal is Q2 segment EBITDA holding ≥40% with depreciation accelerating on the new PP&E base (Q1 PP&E grew 7x to $6.5B); margin that holds even as D&A ramps would prove the operating leverage was the larger driver after all.

Disagreement #3 — The market is pricing the wrong cycle template. Consensus would say: "AI demand is so far ahead of supply that the cycle is uncyclical for at least 2-3 more years; the relevant risk is execution, not oversupply." The evidence disagrees because Vera Rubin NVL72 ships H2 2026 with a disclosed ~5x supply step at next-gen launch, NBIS's own useful-life is now 5 years against contracts of 5 years (depreciation curve = contract curve, no recovery cushion), and the 2022-2023 BTC GPU-mining unwind is the closest historical template — operators carrying long-dated debt and amortizing fast-depreciating assets impaired even though end demand never fell. If we are right, the market has to concede that a normalization of GPU supply alone — without a demand recession — is sufficient to compress NBIS's multiple by 30-50%, and that this scenario is neither in the bull ($300) nor median ($205) PT bridge. The disconfirming signal is H100/B200 reserved-rate pricing holding into Q4 2026 even as new generations ramp; that would mean the shortage is structural rather than late-cycle.

Evidence That Changes the Odds

These are the six items that move the probability of each variant view — not generic facts, but pieces of evidence that an underwriter can audit in a single sitting.

How This Gets Resolved

Six signals that an institutional reader can audit on a calendar. Each is observable in a filing, an earnings call, or a price-data print — not in commentary or "execution color."

What Would Make Us Wrong

The fairest case against our backlog disagreement is that Meta's flex structure functions like firm capacity in practice. The $15B optional component is not free for Meta to walk away from — Meta has spent multi-quarter integration effort to onboard Nebius capacity, and resellable capacity in a GPU-shortage market clears at or above contracted price (sometimes higher) — so the economic difference between "firm" and "Nebius-resells-at-market" may be smaller than the legal language suggests. If H100/B200 spot rents stay elevated through 2027, "resellable at market" becomes "resellable at a premium," and the $15B portion converts to revenue at or above book value without Meta ever exercising. We would have to retract the backlog disagreement and concede the moat is firmer than printed.

The fairest case against our useful-life disagreement is that the prior 4-year baseline was genuinely too conservative. NVIDIA H100 hardware has retained value at 18-month resale checkpoints and CoreWeave's 6-year choice is the auditor-approved hyperscaler standard. If the Q2/Q3 margin holds at 40%+ even as quarterly D&A grows with the new PP&E base — meaning the operating leverage effect dominates the policy effect — we would be wrong to attribute 10pp to the accounting choice and would owe the bulls a margin upgrade.

The fairest case against our cycle disagreement is that AI compute is structurally different from telecom fiber: the 2001 fiber bust required customer adoption to lag installed capacity; in AI, every new generation of model training and inference creates net new demand for the latest hardware while older capacity drops in price. If Vera Rubin demand absorbs the supply step (which is what NVIDIA, NBIS, and CoreWeave all telegraph) and reserved-rate pricing holds into Q4 2026, the cycle template is not 2001 — it is "industrial buildout that does not bust because demand expands faster than supply." We would have to widen the multiple and re-think the implied discount rate.

The deepest risk to all three disagreements is that we are right on the facts but wrong on the timing — the market may continue to price the headline numbers for another 2-4 quarters before the firm/flex split, normalized D&A run-rate, or supply normalization shows up in any disclosed metric. A long-only is paid to wait; an arbitrage is paid to be approximately right at the resolution date.

The first thing to watch is the firm-vs-flex backlog disclosure in the Q2 2026 shareholder letter on July 29, 2026 — that single line item resolves the largest of the three disagreements and sets the tape for the second half of the year.

Liquidity & Technical

Liquidity is decisively not the constraint: NBIS trades roughly $2.7B per day and a 5% fund position is implementable within five sessions for AUMs up to $67B at 20% ADV. The tape is screaming bullish — price just closed at $207.27, a fresh all-time high, +15.7% on the Q1 2026 earnings print, sitting 99% above its 200-day SMA with momentum still expanding.

1. Portfolio implementation verdict

2. Price snapshot

3. The critical chart — price vs 50/200-day SMA

Price is 99% above its 200-day SMA — not "extended," not "elevated," explicitly bullish to the point of stretched. The 50d ($135.66) sits above the 200d ($104.15), the golden cross is intact since 18-Jun-2025, and the most recent death cross was over four years ago (10-Jan-2022). The trend is up, full stop.

4. Relative strength

5. Momentum — RSI and MACD

RSI closed at 69.7 — one tick from the 70 overbought line, the same zone it tagged through August–October 2025 before a 30% correction in November. MACD just flipped back to a positive, expanding histogram after a clean reset through early 2026; the cross above signal happened in late February and momentum is accelerating again. Near-term setup is bullish-but-stretched: a single bad print is enough to trigger profit-taking from 70-RSI levels.

6. Volume, volatility, and sponsorship

The three largest spikes on record are Yandex-era prints from a different business, and useful only as caveats — the relevant book of comparable Nebius-era events is short. The two recent prints both went up: 09-Sep-2025 (+49% on 6.6x) was the Meta cloud deal, today's +15.7% on 2.2x is the Q1 2026 earnings. Volume is confirming each leg of the new uptrend — no sign of distribution.

Realized vol is 91% annualized — above the 10-year p80 ("stressed") band of 79.5%, and roughly 3x the long-run median of 38%. Two observations: (1) the entire post-relisting Nebius regime has lived in the stressed band, so it is the new normal for this name, not a transient shock; (2) any institutional sizing exercise needs a volatility-adjusted weight — a 5% notional position carries roughly 4.5%-of-portfolio daily standard deviation, which is large relative to a typical mega-cap holding.

7. Institutional liquidity panel

ADV strip. $2.7B notional traded every session, equivalent to 5.3% of market cap — the entire float effectively recycles every 19 trading days. Annualized turnover of 1,642% places NBIS in the top decile of US-listed mid/large caps; this is a very actively traded name despite being a structurally complex post-restructuring story.

Fund-capacity table

Liquidation runway

Price-range proxy. Median daily range over the trailing 60 days is 3.4% — elevated and worth noting for execution. Marketable orders into the book should expect roughly 30–40 bps of impact on a normal session and meaningfully more on event days like the +15.7% earnings reaction today. For block prints, prefer VWAP or implementation-shortfall strategies over aggressive sweeps.

Bottom line on liquidity. At 20% ADV participation, a fund can establish a 5%-weight position within five trading days for AUMs up to ~$67B; at the more conservative 10% participation, AUMs up to ~$34B fit the same brief. A 2%-of-market-cap issuer position ($1.03B) liquidates in two sessions at 20% ADV or four at 10%. Liquidity is decisively not the bottleneck — volatility-adjusted sizing is the real constraint.

8. Technical scorecard and stance

Net score: +3 of a possible +6. Bullish on the 3-to-6 month horizon. The tape is in a powerful uptrend confirmed by both earnings-day volume and a freshly re-bullish MACD; the offset is that realized vol is in stressed territory and RSI sits right at overbought, so adding here is a high-conviction-but-late-cycle entry rather than a clean breakout. Bullish confirmation level: a daily close above $230 (cleanly above today's intraday range and ATR-1 extension) opens the next leg. Bearish invalidation level: a daily close below $172 breaks the recent base and would put price back inside the volatile chop-zone toward the 50-day at $135.66 — a sequential break through that would force a reassessment to neutral or trim.

Implementation. Liquidity is not the constraint. Funds up to $34B AUM can build a full 5% position at 10% ADV over five sessions, and most institutional book-building strategies should plan around the volatility regime, not capacity. The cross-reference for fundamentals analysts: the latest tape reaction (+15.7% on Q1 2026 print) is the market validating, not rejecting, the AI-infrastructure capex story — if the financials tab flags compression in unit economics or operating cash burn, the technical bullishness should be read as sentiment-driven and re-evaluated on the next earnings print.