Nebius Group N.V. [NBIS] — Company Research

Task 1 of the Equity Underwriting Memo · As of 2026-07-27

Ticker: NBIS (Nasdaq) · Domicile: Amsterdam, Netherlands · SEC CIK: 0001513845 Filer status: Foreign private issuer — files 20-F (annual) and 6-K (interim), not 10-K/10-Q Price (Alpaca SIP daily bar, 2026-07-27 intraday): $184.92 · Last completed close (2026-07-24): $187.77 Shares outstanding: 253,898,194 as of 2026-03-31 (6-K filed 2026-05-13); 256.0m per stockanalysis.com 2026-07-27 Market cap: ~$47.3bn · Enterprise value: ~$47.5bn (near-zero net debt — see §5)

1. What this company is

Nebius Group is an AI-native cloud infrastructure provider ("neocloud") that designs, builds, owns and operates GPU data centres and sells compute, storage and a software stack (training, inference, agentic tooling) on top of them. It is vertically integrated in an unusual way for its size: it designs its own server racks and data-centre topology, runs its own facilities rather than only leasing colocation, and layers a proprietary software platform (Nebius AI Cloud "Aether", Nebius Token Factory) above the raw silicon.

The corporate history matters more than usual and is frequently mis-stated. Nebius Group N.V. is the same legal entity as the former Yandex N.V. (same CIK, 0001513845 — the EDGAR filing history runs continuously back to 2011). Following the 2022 invasion of Ukraine and the sanctions regime that followed, Yandex N.V. divested its entire Russian business in a transaction that closed in July 2024, retaining four international assets and a large cash balance, and renamed itself Nebius Group. Practically, this means:

• Financial history before FY2024 is not comparable. FY2020 revenue of $2,955.6m and FY2021 of $4,794.2m in the SEC XBRL record are Yandex Russia, a search/advertising/ride-hailing conglomerate that no longer forms any part of this company. Continuing-operations revenue for the retained businesses was $9.8m in FY2023 and $91.5m in FY2024 (SEC EDGAR XBRL, 20-F filed 2026-04-30). Any screen, comp or chart that runs a continuous revenue series through 2022 is measuring two different companies.
• The divestiture left the company with a large net cash position and no Russian operating exposure, which is the structural reason NBIS entered the AI buildout with a cleaner balance sheet than any of its neocloud peers — a point that recurs throughout this memo.

1.1 The four businesses

Two points on the non-core assets:

1. ClickHouse is a live, marked asset, not a footnote. A January 2026 Series D reportedly valued ClickHouse at ~$15bn; the re-valuation of Nebius's stake produced a $780.6m non-cash gain in Q1'26 non-operating income (6-K, 2026-05-13). "Investments in non-marketable equity securities" on the balance sheet rose from $836.6m at 2025-12-31 to $1,614.1m at 2026-03-31. This single line explains why NBIS printed +$621.2m of GAAP net income in a quarter in which it lost $128.0m at the operating line. Treat the GAAP net-income line as uninformative about operations. Management's own "adjusted net loss" for Q1'26 was −$100.3m.
2. Avride and TripleTen are immaterial to the equity value at current prices (together ~2% of revenue) but are optionality/distraction depending on your view. Avride took a strategic investment from Uber in Q1'26.

1.2 How the business actually makes money

Two revenue motions, with very different economics:

• Long-term reserved capacity contracts (Microsoft, Meta): multi-year, take-or-pay-like commitments for dedicated GPU capacity, frequently with customer prepayments. Lower price per GPU-hour, but they de-risk utilisation, and — critically — they are the collateral that makes asset-backed debt financing possible. This is the financing engine, not just a revenue line.
• On-demand / self-serve AI cloud: higher price per GPU-hour, no volume commitment, sold to AI-native startups, model builders and enterprises. Management reported record pipeline generation in Q1'26, up ~3.5x QoQ, and rising pricing on both new-generation and older-generation GPUs.

Management explicitly describes managing the mix between the two ("we remain agile on capacity allocation and contract terms"). The $27bn Meta agreement is structured to preserve that optionality — see §3.

2. Financial trajectory (all figures sourced; PR-sourced items flagged)

2.1 Annual — continuing operations

Source: SEC EDGAR XBRL company facts, CIK 0001513845, 20-F filed 2026-04-30. Facts grouped by the fact's own end date, full-year durations only (350–380 days), latest filed on duplicate period ends.

The 2023→2024 collapse in total assets ($8,755m → $3,549m) is the Russian divestiture, not deterioration. The 2024→2025 increase ($3,549m → $12,431m, +250.3%) is the AI buildout — and is the single reading that put NBIS 45th of 47 on the 2026-07-27 screen. It is adjudicated in the Valuation Analysis (Task 3, §4).

2.2 Q1 2026 — the most recent reported quarter

Source: Form 6-K filed 2026-05-13, Exhibits 99.1 (results) and 99.2 (shareholder letter). Quarterly XBRL is sparse for this foreign private issuer; the figures below are read directly from the filed exhibits, not XBRL.

Three things in that table are load-bearing for the entire memo:

1. Gross margin of 74% (cost of revenue 26% of revenue), before depreciation, up from 51% a year earlier. Genuine operating leverage as capacity scales, not an accounting artefact.
2. $4,778.1m of deferred revenue — cash already collected from customers for capacity not yet delivered. The Q1 movement alone was +$3,198.0m. This is what turned operating cash flow positive to the tune of $2.26bn on $399m of revenue. Customers are pre-funding the buildout.
3. RPO of $33,585.3m against total assets of $22,303.3m. The contractual claim on future revenue is 1.5x the entire asset base. This is the central fact in the asset-growth adjudication.

2.3 An earnings-quality flag that must not be buried

"Starting Q1'26, we revised the useful life for our server and network equipment from four years to five years to reflect usage patterns and current utilization commitments. The change in accounting estimate has been applied prospectively from 2026." — Q1'26 shareholder letter, p.9

This is legitimate under GAAP, is consistent with what several peers have done, and is arguably justified by the observed longer economic life of prior-generation GPUs. But it mechanically reduces D&A on exactly the asset base whose growth is under scrutiny, and it flatters reported EBIT, net income and "EBITDA-to-EBIT conversion" in the first year it applies. On a $7.1bn PP&E base, moving from a 4-year to a 5-year life on the equipment portion reduces annual depreciation by roughly 20% of the equipment charge — order of $250–350m/yr at current scale, growing with the fleet. Every forward margin comparison in this memo is therefore made on adjusted EBITDA (pre-D&A) or on explicitly re-modelled depreciation, never on reported EBIT trends alone. See Task 3 §4 (Factor & Anomaly Scorecard) for how this interacts with the accruals signal.

3. Customers, contracts and the backlog

3.1 The two anchor contracts

The Meta structure is genuinely clever and is under-appreciated in most commentary: the $15bn tranche is effectively a floor with upside retained — Nebius gets financing-grade revenue visibility for lenders without giving up the ability to sell that capacity into a tighter spot market. It is also the reason the headline "$27bn Meta deal" should not be read as $27bn of committed revenue; $12bn is committed, $15bn is optional-to-Nebius.

3.2 Backlog

• RPO: $21,333m (2025-12-31) → $33,585m (2026-03-31), +57% in one quarter.
• Company statement (2026-07-17 6-K): "more than $40 billion of additional contracted revenue from investment-grade customers such as Microsoft and Meta already in place." (Note: this $40bn+ figure is a press-release characterisation and is larger than the $33.6bn RPO in the filed balance-sheet disclosure — the difference is most likely the treatment of the optional Meta tranche and contracts signed after quarter-end. Flagged as PR-sourced, not XBRL-sourced.)
• Deferred revenue (cash in hand): $4,778.1m.

3.3 Customer concentration — the un-quantified risk

Nebius does not disclose a customer-concentration table in the filed exhibits reviewed. Inference from disclosed contract sizes: Microsoft ($17.4bn) and Meta ($12bn committed) together represent the large majority of RPO. This is a material, un-quantified concentration risk and is treated as such throughout — it is not something the memo can size precisely, and it is not something an investor should assume is diversified. Management's counter-narrative (record pipeline +3.5x QoQ, wins across healthcare/life sciences, physical AI/robotics — 1X Technologies, Rhoda — and Token Factory customers Revolut and monday.com) is directionally encouraging but is not quantified in dollars.

4. Capacity: the physical asset base

Nebius uses three distinct power measures; conflating them is the most common analytical error on this name.

The gap between 4 GW contracted and ≤1 GW connected is the entire investment question. Contracted power is cheap to secure and generates no revenue; connected-and-active power is what earns. The 2026 capex programme exists to close that gap.

Owned vs leased: >75% of contracted power is owned capacity, across five owned facilities totalling ~3 GW:

• Independence, Missouri — 1.2 GW owned AI factory; ground broken May 2026; operational 2027
• Pennsylvania — up to 1.2 GW; announced May 2026; phased delivery from 2027
• Finland — 310 MW when fully deployed; one of Europe's largest dedicated AI factories
• Alabama — in active construction; operational 2027
• New Jersey — servicing Microsoft; remaining capacity activating through 2026, majority in H2

Colocation footprint additionally spans Iceland, UK, France, Israel, Spain, Kansas City, Oklahoma, Minnesota. Sites with >100 MW went from 1 at YE2025 to 7 currently.

Technology position. NVIDIA Exemplar Cloud status on GB300 NVL72 for training (among the first globally, and one of few holding it across multiple GPU generations); among the first worldwide to deploy NVIDIA Vera Rubin NVL72 from H2 2026. Eigen AI's inference stack ranked #1 by NVIDIA at GTC 2026.

5. Balance sheet and financing — the crux of the story

5.1 Position as of 2026-03-31, plus July events

Post-quarter: $775m senior secured facility, 2026-07-17, maturing 2030-10-31, priced at SOFR + 250bp, backed by deployed GPU infrastructure and contracted cash flows from an investment-grade customer. (This adds both cash and debt — EV-neutral at the moment of the raise.)

5.2 The financing model, and why the July 17 deal mattered so much

Q1'26 capital raised: $6.3bn — $4.3bn convertible senior notes + $2.0bn equity investment from NVIDIA (NVIDIA disclosed a ~9.3% stake, reported 2026-07-21; the stock rose ~16–19% on the day).

The July 17 secured facility is small in dollars but large in significance, and its own press release makes the claim explicitly: "Together with cash flows under the customer agreement, the facility covers more than 100% of the capital expenditure required to deploy the underlying GPU infrastructure."

If that is genuinely repeatable across the contracted book, the equity dilution most bears assume is required to fund $20–25bn of 2026 capex does not happen. That is precisely why the stock jumped ~8% on the announcement, and it is the single most important thing to verify at the Q2 print (Task 6, catalyst C-2).

The stated funding stack, in management's own order of preference: (1) operating cash flow and upfront customer payments; (2) asset-backed and corporate debt — "raising mid-single-digit billions in the near term"; (3) an at-the-market equity programme, not yet used; (4) other vehicles.

5.3 The convertible overhang

Diluted share count in Q1'26 was 309.0m vs basic 258.3m — roughly 50m shares of if-converted dilution from the convertible notes, or ~20% of the current base. Those converts were struck when the shares were materially higher than today's $184.92, so at present they behave as debt, not equity. Practically this means: the convertibles cap the upside — a large rally re-triggers ~50m shares of dilution — while providing no downside protection to the equity. This asymmetry is explicitly modelled in Task 3.

6. Management and governance

Dual-class structure. At 2026-03-31: 220,406,311 Class A and 33,491,883 Class B shares outstanding (plus 68,142,750 Class A held in treasury). Class B carries superior voting rights. Public shareholders therefore have limited ability to influence a capital-allocation programme of this magnitude — a real governance consideration when the company is committing $20–25bn of capex in a single year.

M&A cadence in 2026 (all in-year): Tavily (agentic search, Feb), Eigen AI (inference optimisation, announced May 1, closed June 10), Clarifai (system-level inference orchestration + IP licence + team, May). Goodwill went from $0 at YE2025 to $163.3m at Q1'26. These are small, talent-and-technology acquisitions, not capacity roll-ups — worth noting because the Cooper/Gulen/Schill asset-growth anomaly is documented largely on acquisitive asset growth, which this explicitly is not.

6.1 Insider activity — sourced, with the empirical asymmetry applied

Source: SEC Form 4 filings, CIK 0001513845, parsed from EDGAR XML, 2025-07-01 through 2026-07-27 (23 filings).

Open-market purchases (transaction code P) in the window: zero.

Applying Cohen, Malloy & Pomorski (2012): insider sales are predominantly noise — diversification, taxes, 10b5-1 plans — and roughly half of the filings above are explicitly 10b5-1 plan sales. Several ex-Yandex executives hold effectively all their net worth in one stock that has risen ~254% in twelve months; diversification is a fully adequate explanation. I therefore do not present this as bearish evidence.

Two observations do survive that discount, and are stated for what they are: - Korolenko's $217m is extreme in absolute size — roughly 84% of all insider selling — and includes filings in April and July 2026 that are not marked as 10b5-1 plan sales. - The informative signal (clustered, non-routine insider buying) is simply absent. Zero purchases across 23 filings and a 12-month window. That is not bearish; it is the removal of a bullish leg. Whatever conviction supports this stock, it is not being expressed by insiders adding.

6.2 Institutional ownership

Source: fintel.io / secform4.com 13F aggregation, retrieved 2026-07-27; stockanalysis.com ownership page.

• Institutional ownership: 52.9%; insider ownership 16.3%
• 752 institutions filing 13D/G or 13F, holding 119,280,543 shares
• Largest holders: BlackRock, Fred Alger Management, Susquehanna International Group, Orbis Allan Gray, Citadel Advisors
• Most recent quarter: institutions bought 774k shares, sold 4.4m shares — net selling
• NVIDIA: ~9.3% stake (disclosed 2026-07-21; $2bn investment)

Composition read, per the consensus-bridge discipline: this holder list is not the profile of a high-conviction active long. BlackRock is largely index/mandate capital. Susquehanna and Citadel are market-makers and multi-strategy firms whose NBIS "long" is very likely the long leg of convertible-bond arbitrage or options market-making inventory, not a directional bet — see the short-interest analysis in Task 3 §3, which independently supports that reading. Orbis Allan Gray and Fred Alger are the only genuinely active, valuation-driven names near the top. The one unambiguously strategic, conviction holder is NVIDIA, which is also NBIS's principal supplier — a relationship that cuts both ways (see Risk R-7).

Limitation flagged per skill discipline: EDGAR has no single endpoint returning all 13F holders of a ticker; the above is aggregator-sourced and is a partial, lagged view, not a complete ownership table.

7. Industry structure, TAM and competition

7.1 The neocloud category

"Neoclouds" are specialist GPU-cloud operators that emerged because hyperscaler capacity could not absorb the 2023–2026 AI training and inference demand, and because AI workloads have fundamentally different requirements (dense power, high-bandwidth interconnect, bare-metal access, short procurement cycles) from general-purpose cloud. The category's competitive position rests on three scarce inputs, in this order:

1. Power — increasingly the binding constraint. Contracted, permitted, grid-connected gigawatts.
2. GPU allocation — supplier relationships with NVIDIA; allocation priority for new generations.
3. Capital — ability to fund $10bn+/year of capex at a cost of capital below the asset's return.

Nebius has an unusually strong position on all three: >4 GW contracted (guided), NVIDIA Exemplar Cloud status plus NVIDIA as a 9.3% shareholder, and — uniquely in the peer group — a net-cash balance sheet.

7.2 TAM

The honest statement is that neocloud TAM figures in circulation are analyst constructions, not disclosed company data, and I decline to cite a specific number as fact. What can be sourced:

• Nebius's own contracted power guidance of >4 GW by YE2026 with ≤1 GW connected — i.e. the company's own revenue-generating capacity in 2026 is roughly a quarter of what it has secured.
• Hyperscaler capex is the demand backdrop; the group's 2026 capital plans are of the order of several hundred billion dollars in aggregate. Nebius's $20–25bn 2026 capex is meaningful but not systemically large.
• The addressable question is not "how big is AI" but "what share of AI compute is rented from specialists rather than built by hyperscalers or the labs themselves" — and that share is precisely what the July 1, 2026 Meta-builds-its-own-cloud story called into question (see Risk R-1).

7.3 Competitive set

(Market/EV/short data: stockanalysis.com, 2026-07-27. Price returns: Alpaca SIP daily bars, computed 2026-07-27.)

Three structural observations, each material to the investment case:

1. NBIS is the only member of this group with a clean balance sheet. CoreWeave carries $35.2bn of gross debt against $2.3bn of cash and an Altman Z-score of 0.36. NBIS's Z-score is 1.49 — also in the "distress zone" by the textbook cut-off, but that reading is distorted by a company whose asset base tripled in a year and whose retained earnings line reflects the Yandex history. The cash comparison is unambiguous: NBIS has $9.3bn of cash against $8.5bn of debt.
2. The peer group is de-rating hard while NBIS has re-rated. CRWV −41% and ORCL −50% over twelve months against NBIS +254%. This kills the usual "the whole group is expensive together, so it's a sector-regime question" defence in reverse: NBIS's premium is idiosyncratic, attributable to the NVIDIA stake, the Meta/Microsoft contracts and the balance sheet — not to sector beta. It also means the relative-value warning is real: an investor can own the same theme at 2.9x 2027 EV/revenue (CRWV) instead of 4.5x (NBIS).
3. Every name in the group carries 24–29% short interest of float. This is a crowded, heavily-hedged, convertible-arb-laden corner of the market. It is not a "no one believes" setup; it is a "everyone is hedged" setup. See Task 3 §3.

The competitor that matters most is not on this list: the customers themselves. Meta, Microsoft, Amazon and Google all build their own capacity. The July 1, 2026 report that Meta may monetise excess data-centre capacity as a cloud offering took 17% off NBIS in one session — because it implies the anchor customer could become the competitor. That is the single most important competitive risk on the name.

7A. Scoping what is NOT published (v1.4.0 — required)

Filings are what everyone has. This section records what was looked for outside them, which corpus was used, and what it returned — including where the answer is genuinely not knowable.

7A.1 Transcript mention-frequency — run FIRST, generatively, before revising any view

History window: 6 quarters, Q4-2024 through Q1-2026. State the constraint precisely rather than implying more: Nebius has held only seven earnings calls as a standalone company — the Yandex divestiture completed in July 2024 and the first Nebius-as-Nebius report was Q3-2024 (2024-11-19). Pre-Q3-2024 transcripts are Yandex N.V.: a different company, different assets, different management remit, and not comparable. Of the seven, six were retrieved; Q3-2024 could not be obtained from any accessible archive. Alpha Vantage's EARNINGS_CALL_TRANSCRIPT quota was exhausted before this task began, so sources are InsiderMonkey (Q4-2024 to Q3-2025) and stockanalysis.com (Q4-2025, Q1-2026). Cross-source comparability check: the control terms "revenue" (25/30/25/25/23/25) and "customers" (27/44/15/21/25/28) show no level shift at the source boundary, so the two corpora are treated as comparable — a check made because the same two sources were found to disagree materially on a NuScale transcript in this batch (see the SMR file, §4A.1). Cached to data/transcripts/; counts in data/mention_freq_NBIS.json; chart 36_mention_frequency.png.

Every emerging and decaying term, listed as an open question with no interpretation attached, then investigated:

Hypotheses generated by this process (recorded as generative, not carried in from the prior memo):

• G-1. The unit of account has been re-based from MW to GW. Independently corroborated: Pennsylvania 1.2 GW of power and land secured (announced 2026-05-13; first 250–300 MW live end-2027, full 1.2 GW by 2030) and an Independence, Missouri 1.2 GW campus broken ground 2026-05-12. Both are inside the ">3.5 GW contracted" figure, i.e. the language change tracks a real, disclosed asset change. Confirmed.
• G-2. The KPI is migrating away from ARR. ARR mentions collapsed 19 → 1 while "contracted" and "contract" rose. RPO and backlog — the metrics that would replace ARR — are in the press release but have never been spoken on a call. Open. This is the term to watch: a company that stops quoting the metric it made its name on is usually anticipating that metric decelerating.
• G-3. "Pricing" emerging is positive, not negative — and this is the single most decision-relevant finding of the exercise. The count alone is directionless, so the text was read. Q1-2026, CRO Marc Boroditsky, in answer to Goldman Sachs: "We continue to see strong pricing across both old and new GPU generations as demand continues to exceed our available capacity. We just raised prices again in the latest quarter, and we are still selling out across all chip types at the higher prices." And in prepared remarks: "On pricing, strong market demand is translating into pricing gains in our latest deals." This is the first direct evidence on the question the v1.3.0 memo named as its own p_bear invalidator — "a demonstrated re-lease of first-generation capacity at flat-or-better pricing." It is a management assertion in Q&A rather than a disclosed metric, and is weighted as such in Task 5.
• G-4. The residual-value question is unasked, not answered. "useful life" 0/6, "re-lease" 0/6, "churn" 0 since Q4-2024, "depreciation" 5 mentions in six calls. The accounting change that adds most to EBIT was made without a spoken word. This is a disclosure gap the market has not priced because the market has not asked.
• G-5. The geographic narrative has narrowed to the US — Israel, Kansas City and Finland all to zero, including a 310 MW Finnish project announced inside the quarter of the call that ignored it.

7A.2 Sector corpora (software/infrastructure: permits, interconnection queues, job postings, GPU supply)

7A.3 What remains genuinely unknowable — with the corpora that failed to answer it

7A.4 Bottom-up capacity TAM and the required implied-penetration statement

Built from megawatts, per references/tam-sizing.md — not from an AI-datacenter industry-report headline.

Unit build

Penetration path — what each revenue level requires in average connected MW at $8.89m/MW-year

Conservative path ($7.00m/MW-year): 2031 requires 4,143 MW = 118% of contracted. Aggressive path ($11.25m/MW-year): 2,578 MW = 74%.

Time to revenue: the 2.4 GW of announced US campuses is a 2027–2030 delivery, with Pennsylvania not reaching full 1.2 GW until 2030. On the base case the pool is not fully monetised until 2031.

Sector unit metrics (the form of screen references/tam-sizing.md says is viable — within a sector, not across):

Required implied-penetration statement. At $184.92 and a $47,538m enterprise value, Nebius is priced at $13.58m per megawatt of contracted power and $52.8m per megawatt it expects to have connected by the end of 2026. The house Base DCF — which values the shares at $174.88, i.e. 5.4% BELOW the market price — already requires 3.26 GW of average connected IT load by 2031, or 93% of the 3.5 GW contracted today and 82% of the ≥4 GW targeted for year-end 2026, at $8.9m of revenue per MW-year. There is essentially no un-priced capacity left. The market is already paying for the entire contracted book to be built, energised, sold out and priced at today's rates. This is TAM capture, not TAM expansion: the gigawatts the transcript signal points to are the same gigawatts the price already discounts.

What would falsify this section: contracted power rising materially above 4 GW during 2026 with the incremental capacity pre-sold (which would create genuinely un-modelled pool), or a disclosed revenue-per-MW above $11.25m sustained for two quarters (which would mean the same megawatts are worth more than the price assumes).

8. Risk register

9. Bull case and bear case, stated fairly

The bull case — what a buyer is actually betting on

1. The contracted book is real and investment-grade. $33.6bn RPO, $4.8bn of cash already collected, Meta and Microsoft as counterparties. This is not a story stock with a pitch deck; it is a company with signed, prepaid, multi-year obligations.
2. The financing model may be self-funding. If the July 17 template — asset-backed debt at SOFR+250 plus contracted customer cash covering >100% of deployment capex — repeats across $40bn of commitments, the equity never has to fund the buildout, and the dilution the market fears never arrives.
3. Unit economics are proving out. 74% gross margin, AI cloud adjusted EBITDA margin of 45% in Q1'26, management targeting 20–30% EBIT margins. Operating leverage is visible in the actual filed numbers, not promised.
4. Scarce assets. 4 GW of contracted power, 75%+ owned, in a world where power is the binding constraint, is an asset with strategic value independent of Nebius's own execution.
5. NVIDIA's 9.3% stake is the strongest available third-party signal on technology position and allocation priority.

The bear case — what it actually turns on

1. This is a capital-destruction machine unless the terminal outcome is very large. $22bn of 2026 capex against $3.3bn of revenue; my own Base-case DCF has cumulative free cash flow of −$30bn through 2031, and 167% of the enterprise value sits in the terminal value. The equity is a long-dated call option on the AI capex cycle, not a claim on near-term cash flows.
2. Five-year assets, five-year contracts, and value that lives in year six. Contracts roughly match the depreciation life. Whatever equity value exists depends on capacity re-letting profitably after the initial terms, into a market that every competitor is simultaneously flooding with supply.
3. The customer can become the competitor (R-1), and the market showed on July 1 exactly how it will react.
4. The peer group is already de-rating. CRWV −41%, ORCL −50% over twelve months. NBIS is the last member of the cohort still trading at a premium multiple.
5. The financing is not yet proven at scale. One $775m facility is a template, not a track record. The ATM programme exists and is unused; that is the swing factor for dilution.

10. Data-sourcing note

Alpha Vantage note: EARNINGS_ESTIMATES returned an empty estimates array for NBIS (the endpoint has no coverage for this foreign private issuer). The consensus revenue/EPS figures used in Task 3 are therefore web-aggregator-sourced, not Alpha Vantage-sourced, and the estimate-revision-count data that endpoint normally supplies is not available for this name — that scorecard row is marked "not computable" rather than silently omitted. One Alpha Vantage call was successfully used on EARNINGS, which supplied the quarterly EPS surprise history used for the PEAD/SUE row.

Numbers flagged as estimates or assumptions (never presented as fact): the ~98%/1%/1% Q1'26 revenue split across Nebius AI Cloud / Avride / TripleTen (only the AI Cloud figure is disclosed; the remainder is a residual); the $250–350m/yr depreciation impact of the useful-life change (my computation, not disclosed); the inference that Susquehanna/Citadel holdings are convert-arb/market-making inventory (an inference from short interest and convert issuance, not a disclosed fact); all forward-period figures in Task 2/3.