Nebius Group

Phase A — Understand the business

Lens 1 · Company Overview

Nebius is a vertically-integrated "AI cloud" — a neocloud GPU infrastructure provider that designs its own servers, racks and data centers and rents NVIDIA GPU compute (IaaS) plus a software stack (MLOps, managed inference) to AI builders. It is the rump of Yandex: after the EU sanctioned founder Arkady Volozh in June 2022, Yandex N.V. sold its entire Russian business for ~$5.2B (the largest corporate exit from Russia), kept four non-Russian assets, renamed to Nebius Group N.V., and resumed NASDAQ trading in October 2024.

Four reportable pieces, but one is ~98% of the story:

• Nebius (core AI cloud): FY2025 revenue $480.3M, +603% YoY (from $68.3M in FY2024). This is the company.
• TripleTen (edtech / reskilling): FY2025 $54.1M, +88% — and the source of an ongoing material weakness.
• Avride (autonomous driving / delivery robots): FY2025 $1.3M; partnered with Uber (a combined Uber+Nebius commitment of up to $375M) and Hyundai.
• Material equity stakes: significant minority in ClickHouse (open-source columnar DB, spun out 2021) and Toloka (LLM data labelling, deconsolidated Q2 2025). The ClickHouse stake drove a $597.4M non-cash revaluation gain in FY2025.

Group FY2025 revenue (after eliminations) was $529.8M, +479% YoY. Contract structure on the core business is shifting from on-demand/self-service GPU rental toward multi-year, multi-billion-dollar committed capacity contracts with hyperscalers — the Microsoft and Meta deals — which carry large upfront customer prepayments (a $982.5M prepayment inflow in FY2025).

Lens 2 · Supply Chain

Map: NVIDIA (GPUs, the chokepoint) → OEM/ODM server assembly → Nebius (in-house rack/server/DC design, owned + co-located capacity) → AI builders / hyperscalers (Microsoft, Meta, AI-natives, enterprises).

• Upstream — the single-source dependency. "We currently rely on Nvidia for the GPU chips we use and on a limited number of other suppliers for other key components". The stack named in the filing: NVIDIA GB300 NVL72, GB200 NVL72, HGX B200, HGX B300, HGX H100, RTX PRO 6000, with InfiniBand NDR/XDR interconnect; Nebius will be among the first NVIDIA Cloud Partners to ship the Vera Rubin NVL72 (announced Jan 2026). This is not arm's-length: NVIDIA is also a shareholder — a $700M Dec-2024 placement (~0.5%) plus an additional $2B via pre-funded warrants targeting 5GW of AI capacity by 2030. Supplier = investor = demand-shaper. That is both a moat and a governance red flag (Lens 13).
• Mid-chain — Nebius's own value-add. In-house server/rack design, proprietary firmware, toolless racks (one engineer per thousands of servers), servers rated to 40°C enabling air-cooling. Owned greenfield (Finland, Missouri, Alabama), build-to-suit (Vineland NJ), brownfield, and co-location (France, Iceland, UK, Israel, Kansas City).
• Power — the second chokepoint. ~170 MW active power at YE2025; >2 GW contracted by Feb 2026, raised to >3 GW by Q1 2026. The whole growth case is a race to convert contracted → connected → active power. Site selection (power, zoning, permits) is repeatedly flagged as the binding constraint.
• Downstream — extreme concentration. Microsoft and Meta are now the demand base (Lens 1/3). The AR line jumping from $11.2M to $720.3M in one year is the hyperscaler-contract dynamic showing up on the balance sheet.

This lens is the thesis: Nebius sits between the two scarcest inputs in AI (NVIDIA silicon and gigawatts of power) and two of the deepest-pocketed buyers. Names, not generics — that's the point.

Lens 3 · Competitive Advantages (moats)

Nebius's pitch vs. rivals rests on full-stack vertical integration + NVIDIA-preferred access + speed-to-deploy:

• Scale + capital access as a moat: $9.3B cash (Q1 2026) and an NVIDIA/Microsoft/Meta-validated balance sheet let it pre-commit to GW-scale capacity that smaller neoclouds can't finance.
• NVIDIA relationship: "consistent track record of being one of the first-to-deploy the latest generation of NVIDIA GPU chips" + Cloud-Partner status + NVIDIA as equity holder. First-to-Blackwell/Vera-Rubin is a real, if temporary, edge.
• In-house DC/hardware design → unit-economics edge (PUE, density, maintenance). Finland reuses ~20 GWh of server heat for district heating. Adjusted EBITDA margin on the AI business hit 45% in Q1 2026 (up from 24% in Q4 2025) — evidence the unit economics are real at scale.
• Bargaining power: weak upstream, mixed downstream. Over NVIDIA: minimal (single-source, allocation-dependent). Over hyperscalers: the multi-year take-or-pay-style contracts give revenue power, but Microsoft/Meta are sophisticated counterparties who can also build their own — the filing concedes hyperscalers "may offer competing solutions."

Durability verdict: The moat is capital + NVIDIA access + execution speed, not switching costs or IP. It is real today and rented tomorrow — defensible only as long as Nebius stays first-to-silicon and best-financed. Competitors named in the filing: CoreWeave, Crusoe, Lambda (compute layer); AWS, Azure, GCP, Oracle (full-stack); Fireworks AI, Together AI (inference / Token Factory).

Lens 4 · Segments

Segment revenue, FY2023 → FY2024 → FY2025, all ``:

Geography: HQ Netherlands; capacity across Europe/Middle East (Finland, France, Iceland, UK, Israel) and the US (Kansas City MO, Vineland NJ, Missouri, Alabama, Minnesota, Oklahoma). The trend is unambiguous acceleration driven entirely by the AI-cloud segment converting GPU capacity into revenue; TripleTen and Avride are now rounding errors and increasingly look like candidates for divestiture / deconsolidation (the Toloka and ClickHouse playbook). Operating loss is not segment-split in the MD&A — the group runs one P&L (Lens 5).

---

Phase B — Measure performance

Lens 5 · Earnings Result

FY2025 consolidated income statement (continuing ops), all ``:

Read it correctly: the headline "net income of $9.8M / EPS ~$0.12" is an accounting artifact of the $598.9M ClickHouse mark, not operations. The operating business lost $611.7M and is getting more loss-making in absolute dollars as it scales. Gross margin improved sharply (cost of revenue fell from 48% → 31% of revenue) — genuine operating leverage. But D&A exploded to $417.9M (+442%) as GPUs hit the P&L, and a one-time $43.6M equipment-loss-in-transit charge hit SG&A. The four-year GPU depreciation life is the swing factor (Lens 10).

Latest print — Q1 2026 (reported 2026-05-13):

• Revenue $399M, +684% YoY, +75% QoQ; Nebius AI $390M (98% of group).
• Core AI cloud ARR $1.92B, +54% QoQ — the metric the market trades on.
• Nebius AI adj. EBITDA margin 45% (from 24% in Q4 2025).
• Operating cash flow $2.3B (vs. −$198M PY) — driven by customer prepayments, not earnings.
• Cash $9.3B at quarter-end.
• Guidance: FY2026 revenue $3.0–3.4B; ARR $7–9B by YE2026; >3 GW contracted; capex raised to $20–25B; adj. EBITDA margin ~40% target by YE.
• "Sold out" — demand exceeds supply.

Market reaction: the stock is up ~210% YTD and ~492% in a year (52-wk range $43.89 → $297.93). What's priced in is flawless backlog conversion. One source notes shares slid 11% on a "$5B financing test" — i.e., the market does react to funding risk, not just demand.

Lens 6 · Earnings Calls (sentiment trend)

No transcripts/ on disk; this lens is ``. Across the FY2025 → Q1 2026 calls, management tone has shifted from "we're building credibility / proving the model" to "we are supply-constrained and pre-selling 2027 capacity." Recurring 2026 phrases: "sold out," "build for 2027 demand not just this year" (CFO Dado Alonso defending the $20–25B capex), "contracted power," "ARR," "tier-one." What they've stopped emphasizing: the legacy non-core assets (TripleTen/Avride get less airtime each quarter), and the going-concern/Russia overhang (now resolved). The CRO hire (Marc Boroditsky, ex-Cloudflare/Twilio, May 2025) signals an explicit pivot to enterprise/hyperscaler sales motion. Net sentiment: aggressively confident, with the tell being that confidence is now expressed as capex they're willing to commit ahead of revenue — the most expensive possible way to be wrong.

Lens 7 · Comps

Peer set: listed neoclouds + hyperscale-adjacent. Multiples are `` with date or n/a. Do not over-trust cross-source market caps — they conflict by date.

The comp that matters: NBIS vs CRWV. CoreWeave does ~4-6x Nebius's revenue ($12-13B FY26 guide vs $3.0-3.4B) yet trades at a similar-to-smaller market cap and a dramatically lower EV/sales (~3.6-4.5x vs ~21-24x). On a backlog basis Nebius's ~$46B vs CoreWeave's much larger (>$50B+) book also favors CoreWeave on price-to-backlog. Nebius is the most expensive name in its own peer group on every revenue-based multiple — the premium is paid for (a) faster growth (+684% vs +112%), (b) higher disclosed margin (45% adj. EBITDA), (c) NVIDIA equity halo, and (d) a cleaner customer mix story than CoreWeave's Microsoft-heavy book. Whether that premium is deserved is the entire bull/bear (Lens 12).

Lens 8 · Stock-Price Catalysts

Moves >5%, last ~18 months (NBIS only relisted Oct 2024):

• Oct 2024 — NASDAQ relisting after Russia severance; trading resumes.
• Dec 2024 — $700M placement incl. NVIDIA + Accel; first AI-cloud validation.
• Sep 2025 — Microsoft contract (up to $19.4B); convertible notes + equity raise; stock re-rates hard.
• Dec 2025 — initial Meta $3B deal + NVIDIA $2B strategic investment, same week.
• Mar 2026 — Meta expanded to ~$27B (Vera Rubin); all-time highs.
• Q1 2026 (May 2026) — earnings beat, ARR $1.92B, capex raised to $20-25B; also an ~11% slide on a "$5B financing test" narrative.
• Jun 2026 — added to Nasdaq-100 (+8.9% pre-market); closed $643M Eigen AI acquisition.

Pattern: NBIS trades on backlog/contract announcements and NVIDIA association, and sells off on financing/dilution fear. It is a demand-narrative stock with a funding governor. Earnings matter less than the next mega-contract or the next capital raise.

---

Phase C — Judge people & books

Lens 9 · Management

• Arkady Volozh — founder, CEO (Exec Director since Aug 2024). Built Yandex from 2000 into Russia's Google; computer scientist, serial entrepreneur (InfiniNet, CompTek; early investor in Getir, Face.com). EU-sanctioned June 2022 over the Ukraine invasion; resigned, then delisted from EU sanctions March 2024; Forbes net worth ~$1.5B. He returned to run the rump. Skin in the game is enormous and entrenched: he beneficially holds ~52% of voting control (Class B = 10 votes/share); Nebius is a NASDAQ "Controlled Company" exempt from majority-independent-board rules. A family trust holds additional Class B.
• Track record: genuinely world-class as a builder (Yandex) — but the Yandex chapter also ended in sanctions and a forced fire-sale, and the entire senior bench is ex-Yandex (Chernin/CBO, Korolenko/CPIO, Shtan/CTO all Yandex infra veterans). Deep operating competence in building data centers and search-scale infra; that is exactly the relevant skill.
• Newer hires signal professionalization: CFO Maria del Dado Alonso (ex-Berlin Brands Group, Azerion, Booking, Amazon; since Jun 2025) and CRO Marc Boroditsky (ex-Cloudflare President of Revenue, ex-Twilio CRO; since May 2025) — credible Western enterprise-go-to-market and finance talent layered onto the Yandex engineering core.
• Capital allocation: aggressive reinvestment ($4.07B FY2025 capex; $20-25B guided FY2026), funded by $4.16B converts + $1.15B equity + $2B NVIDIA warrants + customer prepayments. Also opportunistic value-creation via spin-outs (ClickHouse $597M mark, Toloka). No buybacks/dividends (correct at this stage). ROIC is deeply negative now by design.
• Red flags: the adverse ICFR opinion (Lens 10), the Controlled-Company governance, the related-party-adjacent NVIDIA supplier/investor entanglement, and a tiny auditor (Reanda) for a $70B company. Archetype: founder-operator — high conviction, high control, building fast; the stage fits the archetype, but minority holders are along for the ride with little say.

Lens 10 · Forensic Red Flags

Act-as-forensic-analyst. This is the part of the Nebius story bulls underweight.

1. ADVERSE opinion on internal control over financial reporting (auditor Reanda Audit & Assurance B.V., Amsterdam, 2026-04-30). Two material weaknesses as of YE2025: (a) property & equipment controls, and (b) TripleTen revenue-process controls (TripleTen ≈ 10% of revenue). P&E is the balance-sheet line for a capex story ($846.7M → $5,553.3M in one year) — a material weakness there is exactly where you'd want clean controls. Management says remediation "ongoing… expected complete by end of 2026."
2. Going-concern doubt was raised and only recently removed. The auditors previously concluded substantial doubt about the ability to continue as a going concern; it was lifted only after the divestment cash landed. Fresh history for a company about to spend $20-25B/yr.
3. Earnings quality — the $598.9M gap. FY2025 "profit" is entirely a non-cash ClickHouse revaluation under ASC 321 from an observable-price-change (Series C). A further mark is coming in Q1 2026 (ClickHouse Series D at ~$15B valuation, $400M raised, Jan 2026). Strip the marks and the company loses ~$600M from operations. Non-GAAP "adj. EBITDA" (45%) also excludes $83.2M of SBC.
4. Operating cash flow is prepayment-fueled, not earnings-fueled. FY2025 operating CF of +$401.9M includes $982.5M of customer advances; Q1 2026's $2.3B operating CF is the same mechanism scaled up. This is financing dressed as operations — real cash, but it reverses as capacity is delivered, and it makes OCF a poor proxy for profitability.
5. Depreciation-estimate change flatters FY2026. In Jan 2026 Nebius extended server/network useful life from 4 to 5 years, which it expects to reduce FY2026 D&A by ~$167.6M. Legitimate, but it lowers reported losses via an estimate change precisely when scrutiny is highest — watch for it being used to "hit" margin targets. (For GPUs specifically, a longer assumed life is aggressive given obsolescence risk.)
6. Capital structure complexity / dilution. $4.16B converts (tranches 2029-2032) + NVIDIA pre-funded warrants + ongoing equity issuance = real dilution and refinancing risk; interest expense went 0 → $61.5M and climbs.
7. Auditor + jurisdiction: small Dutch firm (Reanda, auditor only since 2024); FY2023 numbers were audited by a Moscow firm ("Technologies of Trust – Audit"). Yandex legacy in the audit trail.

Regulatory findings. Per regulatory/regulatory-findings.md (fetched 2026-06-18, sources SEC EDGAR EFTS LR + AAER): 0 SEC Litigation Releases and 0 AAERs naming Nebius Group in 2021-2026. Non-SEC web check ("Nebius Group" (FTC/DOJ/FDA/CFPB/consent decree/settlement/fine/penalty) enforcement): no material agency enforcement action surfaced; the only "regulatory" matters are (i) the legacy EU sanctions on founder Volozh, since removed (Mar 2024), and (ii) a Dutch $180.9M dividend-withholding-tax accrual on ~117M repurchased Class A shares from the divestment, settled Feb 2025. The 20-F discloses the adverse ICFR opinion and material weaknesses as its principal control matter rather than litigation. Conclusion: No material regulatory or legal enforcement findings — verified via SEC EDGAR EFTS (LR, AAER), web search, and the 20-F as of 2026-06-18. The governance flag is the adverse ICFR opinion, not an enforcement action.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection

Nebius does not have meaningful EPS yet — it is a pre-profit hyperscale-capex story, so the right frame is revenue/ARR trajectory and the path to profitability, with EPS still negative on a GAAP basis near-term. Every line labelled; bottom-up from FY2025 actuals + management guidance.

Base inputs (sourced): FY2025 revenue $529.8M; Q1 2026 revenue $399M and ARR $1.92B; FY2026 guide $3.0-3.4B revenue + $7-9B ARR exit + ~40% adj. EBITDA margin + $20-25B capex; backlog ~$46B over 2027-2031.

Profitability: Nebius AI adj. EBITDA margin is already ~45%, but GAAP earnings stay negative through at least FY2026-27 under the weight of D&A on $20-25B of capex and interest. Next-quarter (Q2 2026) consensus is EPS −$0.68 on ~$601M revenue. The swing to GAAP profit depends on (a) GPU depreciation life assumptions, (b) utilization on delivered capacity, and (c) financing cost — all of which are management-controllable estimates, which is itself a yellow flag.

EV/sales sanity check: at ~$72B mkt cap and a ~$8B FY2027 base-case revenue, NBIS trades ~9x FY2027 sales `` — still ~2x CoreWeave's forward multiple. The bull needs FY2028 revenue near the $14-20B range and a margin profile that justifies a hyperscaler-like multiple to grow into the price.

(Per --watchlist rules, no Brier forecast logged.)

Lens 12 · Bull vs Bear

Bull case. Nebius is a validated tier-one neocloud sitting on the two scarcest inputs in AI (NVIDIA silicon, gigawatts of power) with ~$46B of legally-binding backlog from the two best counterparties on earth (Microsoft, Meta), NVIDIA itself as a $2B+ equity holder, $9.3B of cash, 45% adj. EBITDA margins proving the unit economics, and a founder who has built hyperscale infra before. ARR went $0 → $1.92B in a few quarters and is guided to $7-9B by YE2026. If it converts the backlog and keeps winning Vera-Rubin-generation capacity, the FY2028 revenue base could be $14-20B and today's $72B cap looks early. Secular tailwind: inference demand is exploding and hyperscalers are capacity-short enough to rent from a third party.

Bear case (permanent-impairment risks).

1. Customer concentration is now existential. Microsoft + Meta dominate the backlog; "a substantial portion of future revenue growth" depends on two customers who can also self-build. If either reprices, delays, or in-sources, the backlog and the multiple both crater.
2. The capex/financing bet can break the company. FY2025 FCF was −$3.66B; FY2026 capex is $20-25B against $3.0-3.4B revenue. This is funded by debt + warrants + customer prepayments. Any AI-capex air-pocket, rate move, or equity-market shut window turns the "$5B financing test" into a solvency question. Customer prepayments that flatter OCF reverse as capacity delivers.
3. GPU obsolescence + depreciation. Four-to-five-year GPU lives may be optimistic; a faster refresh cycle (Blackwell → Rubin → next) impairs the asset base and the margin story.

Pre-mortem (18 months out, thesis broke): Most likely failure — a 2026-2027 AI-capex digestion phase: hyperscalers pause incremental third-party capacity, Nebius is mid-build on $20B+ of committed spend, a convert refinancing hits a risk-off window, dilution spikes, and the stock halves from a 20x+ multiple to a CoreWeave-like ~4x as the market re-rates neoclouds from "growth" to "capital-intensive utility." Secondary failure: a TripleTen/P&E restatement out of the unremediated material weaknesses dents credibility at exactly the wrong moment.

Are multiples too high? On every revenue-based metric, yes vs its own peer (CoreWeave at ~4x fwd-sales vs NBIS ~9-24x). The premium is a bet on growth + margin + NVIDIA optionality. Contrarian view the market is refusing to see: the $46B backlog is both the bull case and the risk — it converts demand visibility into financing obligation and concentration, and the market is pricing the visibility while underweighting that Nebius must spend $20-25B before most of that revenue arrives, with an adverse ICFR opinion on the very balance-sheet line (P&E) doing the spending.

Lens 13 · Devil's Advocate (short-seller)

You are short NBIS at ~$291 / ~$72B.

• Revenue is two phone numbers. Microsoft and Meta are the company's future. Both run rival clouds and both have every incentive to insource once their own capacity catches up. A single renegotiation headline is a 30%+ down day.
• It's a levered capex machine, not a software business, wearing a software multiple. −$3.66B FCF, $4.16B converts, $20-25B FY2026 capex on $3B of revenue. Strip the ClickHouse mark and there's a ~$600M operating loss. The 45% "adj. EBITDA" excludes the D&A that is the real cost of this business and $83M of SBC. This is an infrastructure utility that should trade at CoreWeave's ~4x, not 20x+.
• The accounting tells you to be careful. ADVERSE ICFR opinion. Two material weaknesses (one on P&E — the asset they're spending billions on). Going-concern doubt removed only recently. A $15B/year-old company audited by a small Dutch firm (Reanda) since 2024, FY2023 audited in Moscow. Useful-life extended 4→5 years to cut $167.6M of FY2026 depreciation right when margins are scrutinized. Profit manufactured from a related-investee mark. Every one of these is individually explainable; together they are a forensic checklist.
• Governance: founder controls ~52% of the vote; "Controlled Company" exemptions; the key supplier (NVIDIA) is also an equity holder shaping demand — a circularity that inflates the AI-infra ecosystem broadly.
• What must hold for the price: flawless GW-scale execution, open capital markets for serial refinancing, Microsoft+Meta never wavering, GPU values holding, and a multiple staying 2-5x its closest peer. If FY2027 revenue disappoints 20-30% (backlog converts slower), the stock is a multiple and estimate de-rate — easily −50%.
• Single permanent-impairment scenario: an AI-capex pause coincides with a closed financing window mid-build → forced dilution / distressed financing on $20B+ of commitments. Plausibility: low-to-moderate, but non-trivial and rising with the capex number.

Lens 14 · Management Questions (ordered by information value)

1. Of the ~$46B Microsoft+Meta backlog, what is the revenue-recognition phasing by year (2026-2031), and how much is firm take-or-pay vs. "available capacity" the customer may decline?
2. What are the termination, repricing, and service-credit terms of the Microsoft and Meta contracts, and under what conditions can either reduce committed volume?
3. Walk us through the full FY2026-2027 funding plan for $20-25B/yr capex: how much debt, equity, warrants, customer prepayment, and asset-backed financing — and what's the contingency if equity markets close for two quarters?
4. Customer prepayments were $982.5M in FY2025 and drove Q1 2026 OCF of $2.3B — what is the prepayment balance, and how does it unwind against delivery over the next eight quarters?
5. On the adverse ICFR opinion: precisely what remains unremediated on P&E and TripleTen controls, what is the dated remediation plan, and who is accountable?
6. Why is a company of this scale audited by Reanda, and is a Big Four transition planned?
7. What is the real economic life of your GPUs, and what utilization/refresh assumptions underpin the 4→5-year change — what's the impairment exposure if Rubin compresses Hopper/Blackwell residual values?
8. What adj.-EBITDA-to-FCF bridge gets Nebius to self-funding, and in which year do you expect positive FCF?
9. How concentrated is NVIDIA allocation risk — what contractual GPU supply commitments do you have for the Rubin generation, and how does NVIDIA's equity stake affect allocation priority?
10. What is the power-delivery schedule converting >3 GW contracted → connected → active, and where are the binding permit/interconnection risks?
11. How should we think about TripleTen and Avride — core, or divestiture/deconsolidation candidates like Toloka/ClickHouse?
12. What return on invested capital do you underwrite per data center, and at what utilization does a greenfield site clear its cost of capital?
13. Given the ~52% voting control and Controlled-Company status, what minority-shareholder protections is the board committed to?
14. How do you defend the moat in 18-24 months once first-to-silicon advantage normalizes and hyperscalers expand owned capacity?
15. What is the plan if the AI-infrastructure capex cycle pauses in 2027 — which commitments can be deferred, and what is the downside operating model?

---