Nebius Group

Phase A — Understand the business

Lens 1 · Company Overview (re-grounded — now primary-sourced)

Nebius Group N.V. (Amsterdam-HQ'd, Nasdaq: NBIS) is "a technology company building full-stack infrastructure to service the high-growth global AI industry, including large-scale GPU clusters, cloud platforms, and tools and services for developers". It is the renamed remnant of Yandex N.V., which in 2024 divested >95% of its revenue (the Russian businesses) for ~$2.6B cash + 162.5M Class A shares returned, and rebranded to Nebius in August 2024.

Four pieces, one that matters. Reportable segments:

• Nebius AI cloud — the business. Q1'26 revenue $389.7M of $399.0M group (~98%), adj. EBITDA $174.0M. Full-stack AI cloud (compute, storage, networking, managed services, in-house hardware + software; proprietary Finland DC + leased + greenfield builds).
• Avride (autonomous driving) — Q1'26 revenue $0.9M, adj. EBITDA −$34.1M. A cash-burning option, not a business.
• TripleTen (edtech reskilling) — Q1'26 revenue $11.6M (+10%, ~5,000 new enrollments), adj. EBITDA −$10.4M. The unit carrying one of the two open material weaknesses.
• Equity stakes: ClickHouse (carried $1,517.7M, measurement-alternative, no significant influence) and Toloka (deconsolidated May-2025; Nebius retains 81% economic / 49% voting, carried as $93.9M preferred + $0 equity-method).

Three 2026 acquisitions sharpen the inference story: Tavily (search infra for LLMs, $189.7M, Feb-2026, $163.3M goodwill), Eigen AI Labs (model inference/compression, up to ~$98M cash + ~3.8M shares, announced May-2026 — "#1 speed inference provider by NVIDIA"), and Clarifai (patent portfolio + core team, $75M cash, May-2026). The pattern: buy inference IP + engineers to move up the stack from raw GPU rental toward inference/agentic services. Contract structure is a mix of on-demand "pay-as-you-go" and fixed "reserved capacity," with the reserved/hyperscaler tier increasingly prepaid (deferred revenue $4,778.1M; prepayments drove $3,198.0M of operating cash inflow in Q1).

Lens 2 · Supply Chain (carried — now primary-confirmed)

The chain is two scarce inputs bracketing an in-house integration layer:

NVIDIA (silicon + capital + demand-shaper) → ODM/server assembly → Nebius in-house DC + rack design (owned Finland DC + greenfield US/EU builds + co-location leases) → end buyers: Microsoft, Meta, AI-natives, enterprises.

• Upstream chokepoint #1 — GPUs. NVIDIA is simultaneously Nebius's key supplier AND a $2.0B equity holder (pre-funded warrants, 21,065,936 Class A shares at $0.0001 strike, March-2026). The filing's own risk language: "Supply chain constraints affecting the procurement of high-performance GPUs … remain a key operational challenge … demand … continues to outpace available capacity".
• Upstream chokepoint #2 — power. The second scarce input. >3.5 GW contracted, ≥4 GW targeted by YE2026, >75% owned/contracted. New: Bloom Energy fuel cells (250 MW guaranteed / 328 MW installed, 10-yr, up to $2.6B) — a behind-the-meter power source that routes around grid-interconnection queues, a genuine structural advantage in a power-constrained build-out.
• Downstream concentration — extreme. As of YE2025, a single "Customer D" accounted for $597.0M (83%) of accounts receivable. This is the supply chain's single greatest fragility — quantified now from the primary source, not estimated.

Single-source dependencies: NVIDIA on the chip side; a near-monopsony pair (MSFT + Meta) on the demand side. Names, not generics — the lens passes.

Lens 3 · Competitive Advantages / Moats (carried — primary-confirmed)

What the moat IS: capital access + NVIDIA-preferred allocation + speed-to-deploy full-stack. The 45% AI-cloud adj. EBITDA margin is the proof that the unit economics work at scale — co-location/lease cost fell from 49% to 26% of revenue YoY on operating leverage. Owning the full stack (in-house hardware + software, owned power) is the differentiator management leans on: "Our platform is most efficient when we own the full stack".

What the moat is NOT: switching costs, IP, or network effects. This is fungible compute; the customer can move to CoreWeave, Crusoe, or a hyperscaler's own capacity. The defensibility is being early, fast, and NVIDIA-blessed with the capital to build ahead of demand — a position that erodes the moment capital gets expensive or NVIDIA spreads its allocation. The 2026 inference acquisitions (Eigen, Clarifai) are an attempt to build a stickier, higher-margin software layer on top of the commodity compute — too early to credit as a moat.

Bargaining power: weak over NVIDIA (price-taker on the scarce input), weak over MSFT/Meta (who can in-source), but improving as a scarce-capacity seller in a sold-out market — "we sold out our capacity as demand continued to exceed available supply … everything we build is sold". That is real pricing power while the shortage lasts.

Lens 4 · Segments (re-grounded — now )

Every number here is now SEC-disclosed, not paraphrased. Revenue and adj. EBITDA by segment:

Full-year FY2025: group revenue $529.8M (+479%), Nebius AI cloud $480.3M (+603%) off five new locations deployed in 2025, FY2024 $91.5M, FY2023 $9.8M.

Geography (long-lived assets, where the capital sits): US $4,558.8M (now the largest, up from $2,994.0M at YE2025), Netherlands $2,910.8M, Finland $302.0M, Israel $285.1M. The US is now the gravitational center of the asset base — consistent with the two owned US gigawatt-scale campuses (New Jersey/Vineland + the new Pennsylvania 1.2 GW site).

Trend & cause: AI-cloud is accelerating and inflecting to segment-level profitability (−$27.4M → +$174.0M adj. EBITDA YoY) on operating leverage as capacity scales into pre-sold demand. Avride and TripleTen are immaterial to revenue and net cash drains — both remain divestiture candidates (the Toloka playbook).

---

Phase B — Measure performance

Lens 5 · Earnings Result + Financing (re-run, fully re-grounded — Q1 2026 is the latest print)

Q1 2026 (reported 2026-05-13, statements filed 2026-05-20). All unless noted.

• Revenue $399.0M (+684% YoY, +75% QoQ). Nebius AI cloud $389.7M.
• Loss from operations −$128.0M (vs −$120.3M Q1'25). Operating loss is roughly flat in dollars while revenue 8×'d — that is the leverage story.
• Net income +$621.2M, but $780.6M is the non-cash ClickHouse revaluation gain. Ex-mark, pre-tax loss ~−$160M. Diluted EPS $2.11 (basic $2.40) — almost entirely the mark; do not read it as run-rate earnings power.
• Adj. EBITDA +$129.5M group / +$174.0M AI-cloud (45% margin), vs −$53.7M / −$27.4M a year ago.
• D&A $212.0M (+332%), now 40% of operating expenses — the capital intensity is showing up. The 4→5yr useful-life change cut $43.1M of it (NI +$41.6M).
• Interest expense $63.7M (vs nil) — six convert series; $15.7M of interest was capitalized [Note 7/12].
• Cash $9,298.2M (from $3,678.1M at YE2025). Operating cash flow +$2,258.0M — but $3,198.0M of that is customer prepayments (deferred revenue), i.e. financing dressed as operations; ex-prepayments, operating cash was deeply negative.
• CapEx $2,472.9M (P&E + intangibles) in the quarter alone; assets not yet in use $2,427.9M on the balance sheet [Note 7].
• Balance sheet: total assets $22,303.3M; P&E net $7,131.7M; convert debt carrying $8,432.0M (principal $10,041.8M); deferred revenue $4,778.1M; equity $7,241.9M [Note 4/7/12].

Guidance (reiterated on the Q1 call): FY2026 group revenue $3.0-3.4B; exit ARR run-rate $7-9B; group adj. EBITDA margin ~40%; CapEx $20-25B (raised from $16-20B). Margin cadence: Q3'26 returns to ~Q1 levels, higher in Q4.

The live variable is financing, and the primary sources make the structure concrete. The Q1 liquidity stack: $4.34B of new March-2026 converts (gross $4,337.5M; $2.59B @1.25% due 2031 + $1.75B @2.625% due 2033; effective rates 4.98%/5.20%) + $2.0B NVIDIA pre-funded warrants + $3.2B customer prepayments = the $9.3B reserve. The forward funding plan for the $20-25B capex, in management's own words: "with our Microsoft contract and our 2 Meta contracts, we expect to unlock the ability to raise significant capital through asset-backed financing … at attractive terms based on Microsoft and Meta credit ratings"; plus corporate debt, the ATM (25M Class A shares, never used), and prepayments. Crucially: "the capacity we projected in February is already secured by cash and contractual commitments. The incremental capacity reflected in our raised $20-25B guidance will be funded through additional financing" — i.e. the funded plan is the base build; the raised guidance is the part that depends on a facility not yet closed.

The Meta backstop (Meta must buy any unsold cluster capacity through the 5-year term, up to $15B) is the credit enhancement that makes the asset-backed route financeable — a genuinely important structural feature, now confirmed verbatim in the filing.

Net read: the operating result is strong and improving; the GAAP "profit" is a mark; the funding gap is real but the announced plan is more concrete and more credit-enhanced than a generic "they'll need to raise." The single most important catalyst — does the asset-backed facility close at a disclosed rate? — is still pending. Market reaction (drift from ~$291 → ~$239 over twelve days) confirms the market is pricing the financing overhang, not the operating beat.

Lens 6 · Earnings Calls — sentiment trend (re-grounded — 7 transcripts now on the shelf)

With the full transcript set (2024-Q3 → 2026-Q1) the tone arc is now readable from primary sources, not inferred:

• 2024-Q3/Q4 (post-divestment reset): rebuilding-from-scratch narrative; small numbers; "independent company" framing. Cautious-but-ambitious.
• 2025-Q1→Q3 (the inflection): sequential acceleration; first hyperscaler wins; tone shifts to "demand exceeds supply." Power becomes the recurring constraint.
• 2025-Q4 → 2026-Q1 (sold-out confidence): Volozh's "AI-native hyperscaler across 4 dimensions — capacity, product, customers, capital" framework. Recurring phrases: "we sold out our capacity," "everything we build is sold," "building big." The new vocabulary is capital discipline — "consistent guardrails on cost of capital and shareholder dilution, while maintaining a disciplined capital structure" — a tell that management knows the funding question is the market's whole focus.

What they stopped saying: the early "rebuilding / standalone viability" hedge is gone. What's new and load-bearing: the explicit, repeated walk-through of the asset-backed-financing plan against MSFT/Meta credit — management is pre-selling the funding story as hard as the demand story. CFO is Dado Alonso; IR Gili Naftalovich. Sentiment: aggressively confident on demand, conspicuously careful on capital.

Lens 7 · Comps (re-grounded valuation frame; closest peer re-marked)

Multiples are `` with date, or n/a. The closest comparable is CoreWeave (CRWV) — same NVIDIA-neocloud model, public, reporting.

The closest-peer comparison, re-marked: CoreWeave carries ~3x Nebius's RPO ($99.4B vs $33.6B) on a TTM revenue base ~5× larger ($5.1B+ vs ~$1.5-2B annualized), at 13.7x EV/sales TTM. On forward sales both names compress toward ~6-7x CY27E — Goldman explicitly uses 7x CY27E EV/Sales → ~$227 fair value for NBIS. So the prior dossier's "NBIS at 19-24x vs CRWV at 8-11x, a 2x premium" is overstated on a forward basis: on the multiple the Street actually uses to value these (forward EV/sales), NBIS and CRWV are roughly in line at ~6-7x, with NBIS's premium concentrated in the trailing multiple (a function of its smaller current revenue base, not a structurally richer valuation). The honest read: NBIS is not dramatically more expensive than its closest peer on the forward multiple the market uses — but it has less revenue, less backlog, and a worse controls opinion to get there.

Other reference points: hyperscalers (MSFT/AMZN/GOOGL) trade single-digit EV/sales but are not comparable (diversified, profitable, self-funding). No dividend, no positive ROE for NBIS — n/a on those columns by construction (loss-making, reinvesting).

Lens 8 · Stock-Price Catalysts (re-run — pattern confirmed)

5-year / 18-month catalyst tape:

• Aug 2024 — Yandex→Nebius rebrand; trading resumes after the Russia divestment. The reset that created the current entity.
• 2025 — sequential earnings beats + first hyperscaler contracts → the re-rate from low-teens to >$100.
• Sep 2025 — public equity offering + convert issuance (June/Sept 2025 notes) — funding events.
• 2026-03 — NVIDIA $2.0B equity + Microsoft ($17.4B) and second Meta (up to $27B) contracts → the move toward ~$290.
• 2026-05-13 — Q1 print (revenue +684%, capex raised to $20-25B) + Bloom Energy $2.6B power deal (Bloom +12% on the news).
• 2026-06-22 — added to Nasdaq-100; pop faded within days against the funding overhang.
• 2026-06-22 → 6-28 — sell-side "funding gap" teardowns + general AI-capex risk-off → drift ~$260 → ~$239.

Pattern (unchanged and now well-evidenced): NBIS trades UP on contract / capacity / NVIDIA-association headlines and DOWN on financing / dilution fear. Index inclusion (a flow event) was overwhelmed within days by funding-gap research (the structural governor). Earnings matter less than the next mega-contract or the next capital-raise headline. This is a catalyst-driven, narrative-sensitive stock where the single dominant variable is the cost and cleanliness of the next dollar of capital.

---

Phase C — Judge people & books

Lens 9 · Management (carried — now primary-confirmed)

• Arkady Volozh — founder & CEO. Built Yandex into a >$30B Russian tech champion; navigated the forced Russia exit and rebuilt as a Western-domiciled AI-infra company in ~18 months. Beneficially holds ~52% of voting power via Class B (10-vote) shares — a Controlled Company. Founder-operator archetype with absolute control; the bet is on his execution and capital-markets access.
• Bench: ex-Yandex technical core + Western hires — CFO Dado Alonso, CPO Roman Chernin, IR Gili Naftalovich. Audit-committee financial expert: Mr. Ryan.
• Capital allocation: aggressive reinvestment — $20-25B/yr capex, three bolt-on inference acquisitions, willingness to monetize equity stakes (Toloka growth round led by Bezos Expeditions; explicit openness to tap ClickHouse) to fund the core "while minimizing dilution". The Toloka deconsolidation (retaining 81% economics, 49% voting) shows a deliberate "spin-but-keep-the-upside" playbook. No buybacks (correctly — every dollar goes into the build).
• Skin in the game: very high (founder, ~52% vote). SBC $35.3M/qtr, unamortized $270.1M over ~4.2 yrs — meaningful but not egregious at this revenue ramp.
• Red flags: the NVIDIA supplier-AND-investor circularity; Controlled-Company exemptions; two open material weaknesses on his watch (though two prior ones were remediated). Founder-control is the double-edged core: it enabled the fast rebuild and concentrates all governance risk in one person.

Lens 10 · Forensic Red Flags (re-run — fully re-grounded; this is the crux)

Acting as a forensic analyst, every figure now from the filings.

1. Earnings quality — the "profit" is a mark. Net income $621.2M is $780.6M of non-cash ClickHouse revaluation sitting on a $128.0M operating loss. This is the single most important forensic fact: do not capitalize the GAAP EPS. Note there were two separate ClickHouse marks — $597.4M on the Series C (recognized FY2025) and $780.6M on the Series D (Q1'26); the prior dossiers conflated them into one ~$598M number. ClickHouse is a Critical Audit Matter (Level 3 fair value, 22.5% DLOM, back-solve from a $159.59/share private round) — i.e. the auditor itself flags the valuation judgment.
2. Cash flow vs earnings — operating cash is prepayment-fuelled. OCF +$2,258.0M, of which +$3,198.0M is deferred-revenue prepayments. Ex-prepayments, the business consumes cash. Receivables also ballooned ($720.3M → $1,479.2M; the $597M single-customer concentration). This is financeable but flatters the headline "self-funding" narrative.
3. Capitalization aggressiveness — flagged by the auditor. P&E is a Critical Audit Matter specifically on placed-in-use timing and capitalized costs. $2,427.9M of "assets not yet in use" [Note 7] is the swing factor: when those go into service, depreciation steps up hard. The 4→5yr life extension (+$41.6M NI) is a legitimate-but-margin-flattering estimate change made "right when margins are scrutinized."
4. ICFR — adverse opinion stands; two material weaknesses remain (fixed assets + TripleTen rev rec). But the financial-statement opinion is unqualified, two prior weaknesses were remediated, and going-concern doubt is removed. The fixed-assets weakness is the uncomfortable one — it's a control gap on the very asset class ($7.1B P&E) being bought with $20B+/yr of capex.
5. Auditor — Reanda Audit & Assurance B.V. (Amsterdam), since 2024; audit fees $6.3M (2025) vs $2.2M (2024). A small/mid-tier Dutch firm auditing a ~$70B-cap company is a real governance question — though it is a PCAOB-registered firm and the FY2025 opinion is unqualified. FY2023 was audited by "Technologies of Trust – Audit" (Moscow, the legacy-Yandex auditor, 2021-2024) — the "Moscow audit" the prior dossier flagged is the legacy entity, now superseded.
6. Debt structure — converts everywhere, dilution baked in. $10,041.8M principal across six series, several accreting to 120% of principal at maturity, conversion prices $180-$183 (March notes). 6,000,126 anti-dilutive shares excluded from Q1 diluted EPS; diluted share count already 308.97M vs 258.30M basic [Note 2]. The dilution path is structural, not hypothetical.

Regulatory findings (required sub-section):

• SEC Litigation Releases & AAERs: ZERO. Verified via SEC EDGAR EFTS (LR + AAER).
• Item 8 Legal Proceedings (20-F): ordinary course only — $3.3M probable-and-accrued litigation, $6.2M total claims at YE2025 ($5.9M at Q1'26); management believes no material adverse effect.
• Non-SEC enforcement (web): no material FTC/DOJ/EU enforcement actions surfaced. The relevant regulatory exposure is forward-looking — EU AI Act phased implementation, EU Energy Efficiency Directive (DC energy reporting), the proposed EU Cloud and AI Development Act, and AI-chip export-control frameworks — all named as risks in the filing, none an enforcement action.
• Verdict: No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER), web search, and 20-F Item 8 as of 2026-06-29. The forensic risk is accounting-quality and controls (the adverse ICFR, the mark-driven "profit," the capitalization judgment), not litigation or enforcement.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (re-run — guidance-anchored, now to primary actuals)

Built bottom-up from FY2025 actuals ($529.8M) + Q1'26 ($399.0M) + reiterated guidance. Revenue scenarios:

Sourcing discipline: FY2026 consensus ~$3.4B (top of guide) and a Street path of ~$11B (2027) / ~$21B (2028) are ``; note the wide analyst dispersion — one major shop models FY2027 at only $7.7B. Use the band, not a point.

EPS / profitability projection, with the arithmetic:

• GAAP EPS is not meaningful through ≥FY2027. Operating loss ~$128M/qtr and rising as "assets not yet in use" ($2.43B) go into service and lift D&A; interest expense $63.7M/qtr and climbing on a $10B+ convert stack. The base case is GAAP-negative through FY2026-27 before D&A + interest.
• Non-GAAP / adj. EBITDA is the right metric. At ~40% group adj. EBITDA margin on ~$3.2B FY2026 revenue → ~$1.3B FY2026 adj. EBITDA, scaling to ~$3-4B FY2027 at the $8-11B revenue band. But adj. EBITDA excludes the D&A that is this business's true economic cost — so it overstates owner earnings for a capex utility.
• Valuation anchor: at ~$68B EV on ~$8-11B FY2027 revenue → ~6-8x CY27E EV/sales, roughly in line with CoreWeave's forward multiple and with Goldman's 7x → ~$227 fair value. The bull needs FY2028 revenue at the $15-21B top of the band and a margin profile that earns a premium multiple to grow into the ~$66-72B cap. The ~18% de-rate from $291 buys modest margin of safety; it does not change the shape of the bet.

Next hard datapoint: Q2'26 print (~Aug 2026) — watch exit-ARR trajectory vs the $7-9B path, the FCF/financing update, and any disclosed asset-backed facility.

(Refresh run — no fiscal year closed this period, so no Brier forecast resolved. Not logging a new Brier forecast: this is an unattended re-grounding, not a fresh committed base case — per SKILL, only log when genuinely committing. The base FY2026 ≥$3.0B revenue call is high-confidence-but-trivial given reiterated guidance; the informative binary is the financing print, which is not a clean EPS/revenue Brier line.)

Lens 12 · Bull vs Bear (re-run — against the primary facts)

Bull case. A validated tier-one AI-native hyperscaler with $529.8M FY2025 → ~$3.2B FY2026E revenue, 45% AI-cloud adj. EBITDA margins proving the unit economics, $33.6B of contracted RPO, >3.5 GW power toward ≥4 GW (>75% owned), NVIDIA as a $2.0B equity partner with preferred allocation, $9.3B cash, three inference acquisitions building a higher-margin software layer, and a non-dilutive funding lever in the ClickHouse stake ($1.52B carrying / ~$4B+ notional at the $15B round) plus the Meta backstop-through-the-term that de-risks both demand AND financeability. If FY2028 revenue reaches the $15-21B band at a defensible margin, a ~$66-72B cap is early. Capacity is sold out — the constraint is supply, not demand.

Bear case (the 2-3 permanent-impairment risks).

1. A financing window slams shut mid-build. $20-25B/yr capex on ~$3.2B revenue; the raised guidance is explicitly funded by financing not yet closed. If AI-capex sentiment turns risk-off before the asset-backed facility prints, the choices are distressed terms or dilutive equity — and the convert stack ($10B principal, accreting to 120%) already bakes in dilution. Permanent-impairment-grade.
2. The "profit" and the controls don't inspire confidence in a crisis. A GAAP profit that is entirely a Level-3 investee mark, on a business with an adverse ICFR opinion and a material weakness on the very fixed-asset class being bought with the capex — if a capitalization or valuation judgment proves aggressive, the de-rate compounds the funding stress. Quality-of-earnings risk that becomes solvency risk under stress.
3. Customer concentration (83% of receivables in one customer) + GPU residual risk. If MSFT or Meta in-sources or renegotiates, the RPO and the asset-backed collateral both weaken at once. The 5-yr GPU life assumption may be optimistic against the silicon-refresh cadence. Revenue-deceleration / asset-impairment grade.

Pre-mortem (18 months out, thesis broke): A 2026-2027 AI-capex digestion phase collided with a risk-off financing window mid-build on $20B+ of commitments → Nebius raised at distressed asset-backed terms (or diluted via the ATM into a falling stock) → the forward EV/sales multiple compressed from ~7x toward CoreWeave-like-or-worse → and a capitalization/valuation question surfaced under scrutiny → the stock halved. The legible version of this exact scenario is why the multiple is fragile even though the business is real.

Are multiples too high? On forward EV/sales — roughly fair vs the closest peer and vs the Street's own 7x→$227 anchor (i.e., not the "2x overpriced vs CoreWeave" the prior dossier claimed). On trailing and on quality-of-earnings — yes, demanding: you are paying a hyperscaler-growth multiple for a loss-making, prepayment-funded, adverse-ICFR capex utility.

Contrarian view the market is refusing to see (refined against the filings): The re-grounding cuts the bull's headline too — the backlog is $33.6B not ~$47B, the ClickHouse stake carries at $1.52B not $4.2B. But the symmetric contrarian point is that the prior bear case was also overstated: there is no "$600M operating loss" (it's $128M), no going-concern doubt (removed), no 2x-CoreWeave valuation premium on the forward multiple (roughly in line), and the financing plan is more concrete and more credit-enhanced (Meta backstop, MSFT/Meta investment-grade-anchored asset-backed financing) than a generic "they'll need to raise $29B." What the market is genuinely mispricing is not direction but the width of the distribution: this is a binary-ish funding-execution bet wearing a continuous-growth valuation. The clean-financing-print outcome and the distressed-raise outcome are both plausible, and the gap between them is enormous.

Lens 13 · Devil's Advocate / short-seller (re-run — at ~$239)

Dismantling the bull at ~$239 / ~$66-72B:

• It's a levered capex utility wearing a software multiple, and the "profit" is fake. $621M net income is a $780.6M non-cash ClickHouse mark on a $128M operating loss. 45% "adj. EBITDA" excludes the D&A ($212M/qtr, 40% of opex) that is this business's true cost. Capitalize the EPS and you're wrong.
• The funding math is unforgiving. $20-25B FY2026 capex on ~$3.2B revenue; management concedes the raised guidance is funded by financing not yet closed. $10B of converts already outstanding (accreting to 120%); $9.9B of additional executed-but-not-commenced leases. Every quarter the asset-backed facility doesn't print at a disclosed rate is a quarter of refinancing risk.
• Controls + concentration are the knife. Adverse ICFR opinion; a material weakness on fixed-asset management — the asset being bought with $20B+/yr; 83% of receivables in one customer; a $70B company audited by a small Dutch firm ($6.3M fee). Any single capitalization or single-customer surprise compounds the funding stress.
• Governance: founder ~52% vote, Controlled-Company exemptions, NVIDIA as supplier-and-investor shaping demand.
• What must hold for the price: flawless GW-scale physical delivery (two owned US campuses mid-build — Vineland/NJ + Pennsylvania), open capital markets for serial refinancing, MSFT+Meta never wavering, 5-yr GPU residuals holding, and a forward multiple staying ~7x. If FY2027 revenue disappoints 20-30% (and recall one major shop already models $7.7B, well below the ~$11B Street point), it's a multiple and estimate de-rate — easily −40-50%.
• The counters I must respect (why I'm not pressing the short): (1) the forward valuation is NOT the egregious 2x-CoreWeave premium the bear consensus claims — it's ~in line at ~6-7x CY27E, so the "valuation crash from a stretched multiple" thesis is weaker than advertised; (2) $9.3B cash + the Meta backstop + the ClickHouse stake are real bridges across the gap; (3) heavily-shorted, NVIDIA-haloed, Nasdaq-100-included — squeeze risk on any clean financing print is severe.
• Single permanent-impairment scenario: AI-capex pause + closed financing window mid-build → distressed financing / forced dilution on $20B+ commitments. Plausibility low-to-moderate, rising with the capex number — which rose again this quarter (4 GW target).

Lens 14 · Management Questions (carried + sharpened by the filings — ordered by information value)

1. The asset-backed facility against the Microsoft and Meta contracts — what size, what rate, what advance ratio, and when does it close? (The whole call hinges here.)
2. Of the $20-25B FY2026 capex, exactly how much is funded today (cash + committed) vs dependent on an unclosed facility?
3. ClickHouse and Toloka monetization — under what conditions, and how much of the capex program could the stakes realistically fund without triggering tax/lock-up frictions?
4. The fixed-asset material weakness is on the asset class you're spending $20B+/yr to build — what specifically failed, and what is the dated remediation milestone (you've now missed the "end-2025" target)?
5. Customer D is 83% of receivables — name the concentration by revenue, the contract's take-or-pay floor, and your exposure if they in-source.
6. Microsoft $17.4B and Meta up-to-$27B — split firm vs optional/backstop at the contract level, and how does the backstop flow into the asset-backed collateral?
7. Vineland (NJ) and the new Pennsylvania 1.2 GW campus — give the power-on schedule, milestone-by-milestone; what slips first if anything slips?
8. 5-year GPU useful life — defend it against the silicon-refresh cadence; what's the residual-value assumption and the downside if it's wrong?
9. $2.43B "assets not yet in use" — when do these enter service, and what is the resulting step-up in quarterly D&A?
10. Dilution guardrails — you cite "cost of capital and shareholder dilution" discipline; quantify the maximum ATM usage and convert dilution you'd tolerate before pivoting funding mix.
11. Inference stack (Eigen, Clarifai, Tavily) — what revenue/margin do you expect this higher-margin layer to contribute by FY2027, and is it a moat or a feature?
12. Auditor — will you move to a Big-Four / larger firm as you scale toward a $100B-cap profile, and on what timeline?
13. Bloom Energy fuel cells — what % of the ≥4 GW will be behind-the-meter, and how does that change your interconnection-queue and cost-per-MW math vs grid power?
14. NVIDIA relationship — beyond the $2B equity, what does "preferred access" contractually guarantee on allocation and timing, and how exclusive is it?
15. Path to GAAP profitability — at what revenue scale do you expect D&A + interest to be covered, and what does steady-state owner-earnings margin look like for an owned-full-stack neocloud?

---