NBISNebius Group N.V.

NBIS — Nebius Group N.V. — dossier 2026-09-23

Slot names-ai-capex-5, cloud research routine, unattended. Model-drafted, not human-verified. Stages closed here: S1, S2 (incl. capital allocation), S3 (gate-minimum, incl. earnings quality), S5, S7 valuation table, S8, S9. S0/S4/S6/S7 are carried forward from the ai-capex cluster files.

Every figure below carries the document it was read from. Quotes were located by exact substring search in a text extraction of the filing at the stated URL, or in the transcript the ROIC.ai connector returned — never re-typed. Nebius renders some figures with a space after the dollar sign in the notes ($ 597.0); the spacing is the source's.

Sources

Not read (gate-minimum, §3): the FY2024 20-F, older interim reports, transcripts beyond the two most recent, the investor-day material, and the founder's quarterly shareholder letters — which is where Nebius puts its ARR targets and which live on the IR site, outside this environment's egress policy. That is a real gap and it is why the S8 table starts at the February 2026 guide rather than at the 2025 ARR promise. As a foreign private issuer Nebius files no 10-Q, no DEF 14A and no quarterly 8-K Item 2.02, so there is no proxy skim and no earnings-release exhibit series to work from; the 6-K exhibits above are the whole record.

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), coverage sec_fpi, cik 1513845, cluster ai_datacenter. Watchlist note: "20-F/A 2026-05-22 plus 6-K and Form 4." Not a Tier 0 promotion — this is the round-robin's turn for the name.

S2 — Kill test

The operating numbers

FY2024 and FY2025 from 6K-Q4-PR; 2026 columns from 6K-Q2-MDA and 6K-Q2-PR.

Revenue growth is not in doubt: +479% in 2025, +529% in H1 2026, +454% in Q2. The AI cloud segment is 98% of it ($574.9m of $582.3m in Q2; 6K-Q2-MDA). TripleTen, the edtech unit, shrank 19% YoY to $10.0m, and Avride is $1.0m.

Reported profit is not operating profit. FY2025's $29.0m of net income and H1 2026's $430.8m both come from outside the business: a $780.6 million gain from revaluation of the ClickHouse stake in H1 2026 ("attributable to the remeasurement of our investments in ClickHouse Inc. based on observable price changes from ClickHouse's January 2026 Series D preferred stock financing, which raised $400 million at a reported valuation of approximately $15 billion", 6K-Q2-MDA), after $597.4m on the same asset in H1 2025. Strip that and the H1 2026 loss from operations is $303.9m, wider than 2025's $231.5m. Q2, with no ClickHouse mark, printed a $190.4m net loss.

The balance sheet at 2026-06-30 (6K-Q2-FS)

Shares outstanding 2026-06-30: 271,855,218 (238,400,165 Class A + 33,455,053 Class B), excluding 50,185,726 Class A in treasury (6K-Q2-PR). Add NVIDIA's 21,065,936 pre-funded warrant shares at a $0.0001 exercise price, which the company itself counts as outstanding for EPS (6K-Q2-FS Note 14): ~292.9m economic shares.

❗ Cash generation is customer prepayment, not earnings. Operating cash flow of $4,504.1m in H1 2026 is explained in one sentence: "The increase was primarily driven by advance payments received from customers, resulting in an increase in deferred revenue of $4,395.0 million" (6K-Q2-MDA). The deferred-revenue balance rose from $1,577.5m to $5,975.2m over the same six months. Against that, capex was $8,130.3m and investing outflow $8,288.8m. Free cash flow on the reported figures is −$3.6bn in six months, and it is negative by far more than that once the prepayments are read as the liability they are.

❗ The convertible stack is carried at a third of its value. Six series are outstanding — 2.00% due June 2029, 3.00% due June 2031, 1.00% due September 2030, 2.75% due September 2032, 1.25% due March 2031 and 2.63% due March 2033. Carrying debt is $8,545.7m; the note discloses total fair value of convertible debt of $20,823.7m (6K-Q2-FS Note 12). The $12.3bn gap is the embedded equity option, and it is the dilution the balance sheet does not show.

⚠️ Coupons are cheap because the equity is doing the work. 1.00%–3.00% money is not a credit market judgement about Nebius; it is the price of a call on a stock that has moved from roughly $30 to $236 (PX-2026-09-22). If the equity stops compounding, this financing channel closes and the company falls back on asset-backed paper — the July 2026 facility was $775m at SOFR + 250bp, "equivalent to a mid-single-digit percentage" (CALL-Q2-26, Dado Alonso).

Survives or dies: Nebius does not die on the balance sheet in the next eighteen months. It has $8.0bn of cash, $6.0bn of customer money already banked, nothing material due before June 2029, and almost no corporate-level debt. It dies — or rather, the equity does — if the AI cloud's marginal price per megawatt falls before $20–25bn of 2026 capex has been depreciated against it, because the whole model is build ahead of contract, in management's own words: "We build capacity ahead of contracts" (CALL-Q2-26, Arkady Volozh). The maturity wall is not the clock; the depreciation schedule is.

Capital allocation and returns (five years is not available; three is)

Nebius in its present form is two years old — the Yandex divestment completed in 2024 and the comparatives before it are a different company. A five-year table would be dishonest, so:

(6K-Q2-MDA cash-flow section; 6K-Q2-FS.)

• Share count goes one way. 219,465,088 Class A outstanding at 2025-12-31 → 238,400,165 at 2026-06-30, plus 21.1m NVIDIA warrant shares issued in March, plus 2,743,655 unvested Eigen AI consideration shares, plus whatever the $20.8bn fair value of the converts eventually becomes. Nothing offsets it: no buyback exists.
• The ATM was used hard and fast. On 2026-05-13 the CFO said "We have not utilized this program to date, but we are evaluating change on it" (CALL-Q1-26). By 2026-08-12: "We issued 12.7 million Class A shares at a weighted average price of $224 per share, generating gross proceeds of approximately $2.8 billion. As of June 30, 12.3 million shares remain available under the program" (CALL-Q2-26). Half a 25m-share programme inside one quarter. He then described it as "a flexible funding tool, and it remains an option in our arsenal rather than a commitment."
• Returns on invested capital are not yet a meaningful number. Operating income is negative on $13.0bn of net PP&E; ROIC is negative and incremental ROIC is undefined. The company's own proxy for the return is the deal yield it quotes: "$20 million to $25 million per megawatt, with upfront payments that cover 50%, 60% of the associated CapEx" on midterm contracts, and "$40 million to $50 million per megawatt range, and sometimes above" on short-duration deals (CALL-Q2-26, Volozh). Those are revenue-per-megawatt figures, not returns, and the difference matters: a two-year payback on a five-year-depreciated asset is the whole bull case and none of it is disclosed at the contract level.
• Goodwill impairments in the window: none. $605.6m of goodwill was only created in H1 2026.

Capital-allocation verdict: ⚠️ neutral, un-scoreable on returns. Nebius is not destroying capital in the way the word usually means — it is spending every dollar it can raise on an asset base whose return nobody outside the company can compute, funded by shareholders and customers rather than by earnings. That is a legitimate strategy for a land-grab and an unverifiable one from outside. Call it neutral and put the burden on S9's monitoring triggers.

S3 — Filings, gate-minimum

Read: the FY2025 20-F risk-factor summary, Item 15 (controls) and the auditor's reports; the Q2 2026 MD&A and financial statements in full; the Q2 and Q4 press releases; the 2026-02-12 auditor-change 6-K and its Ex-16.1; the 2026-09-08 Palantir 6-K; both 2026 transcripts, Q&A-first per SKILL §3.1.

The three biggest risks management itself discloses

All three are the company's own words, located in the 20-F's Summary of Risk Factors (20F-2025, Item 3.D):

1. Financing dependence. "Our core business is capital-intensive and currently not profitable, and our ability to continue to operate and to grow will depend in large part on our ability to raise additional equity or debt financing, either through the public or private markets or other third-party sources." It is the second bullet on the list, above competition and above demand.
2. Demand may not arrive for capacity already built. "If there is insufficient customer demand to utilize the compute capacity we build, we would not realize the expected returns on our expansion efforts and our businesses, future operating results and financial condition would be adversely affected" — and, separately, "A slowdown, deferral or reprioritization of AI-related customer spending, or the development of excess industry capacity, could materially adversely affect our revenue growth, results of operations, financial condition and share price."
3. Internal control. "We have identified two material weaknesses in our internal control over financial reporting. If we are unable to remediate these material weaknesses, or if we identify additional material weaknesses in the future or otherwise fail to establish and maintain an effective system of internal controls, we may not be able to accurately or timely report our financial condition or results of operations, which may adversely affect our business."

A fourth is worth naming because it is the depreciation question in management's own language: "Our results of operations may be adversely affected if we are not able to accurately estimate the value and useful lives of our long-term infrastructure assets or to amortize them over the periods we anticipate."

❗ Internal control: adverse opinion two years running, and the auditor has been changed

• Management's own conclusion: "Based on that evaluation, we concluded that as of December 31, 2025, our internal control over financial reporting was not effective" (20F-2025, Item 15).
• The auditor's: "In our opinion, because of the effect of the material weaknesses described below on the achievement of the objectives of the control criteria, the Company has not maintained effective internal control over financial reporting as of December 31, 2025" (20F-2025, Reanda's ICFR report). The financial statements themselves carry an unqualified opinion.
• Two years, not one: "Reanda's reports dated April 30, 2025 and April 30, 2026 expressed an adverse opinion on the Company's internal control over financial reporting" (20F-2025, Item 16F).
• The two weaknesses open at 2025-12-31 are, verbatim (20F-2025):
• "Our controls related to fixed assets were not adequately designed and were not operating effectively. Specifically, we did not fully implement and ensure the effectiveness of the relevant controls and procedures over depreciation start dates, and timely reconciliation around the asset count process."
• "We did not adequately and timely implement and maintain effective information technology general controls and have not consistently documented the execution of business process controls supporting revenue recognition in respect of our TripleTen business unit (representing approximately 10% of total revenues)."
• Three weaknesses existed at 2024-12-31; two were remediated in 2025 and two remain, so the count fell from three to two while the asset base it governs grew from $5.6bn to $13.0bn of net PP&E. Management "expect[s] to complete this project by the end of 2026."
• The auditor changed on 2026-02-12: the board dismissed Reanda Audit & Assurance B.V. (a small Amsterdam firm) and will put Deloitte & Touche LLP to the 2026 AGM for FY2026 (6K-AUDITOR). The company states there was no disagreement, and Reanda's Ex-16.1 letter agrees with the disclosure except that it "[is] not in a position to agree or disagree" with the statement that Deloitte was not consulted. That exception is boilerplate for an outgoing auditor; the change itself is the right direction of travel and is noted here, not held against the name.

⚠️ Customer concentration, from the issuer's own note

"As of December 31, 2024 and 2025, the Group's maximum exposure to credit risk related to significant customer concentrations, measured at the gross fair value of accounts receivable was $ 6.6 ( 59 %) attributable to Customer A in 2024 and $ 597.0 ( 83 %) attributable to Customer D in 2025, and no significant customer concentration existed as of December 31, 2023" (20F-2025, concentration note). The cluster file ai-capex-2026-09-20.md already carried this figure; it is confirmed here from the primary document. Customer D is not named in the filing; the $27bn Meta contract and the Microsoft contract are the two candidates management discusses on the calls, and this dossier does not guess between them.

Earnings quality

1. The non-GAAP bridge. Adjusted EBITDA adds back nine items, and the two that matter are D&A ($471.7m in H1) and share-based compensation ($137.8m in H1). Adjusted EBITDA of $365.7m is arrived at from a loss from operations of $303.9m — the bridge is bigger than the result. Adjusted net loss is disclosed alongside and is the more honest line: −$133.5m in H1 2026 and −$446.7m in FY2025, against reported net income of +$430.8m and +$29.0m (6K-Q2-PR, 6K-Q4-PR). ⚠️ Flag 1.
2. Share-based compensation was $137.8m in H1 2026, 14% of revenue and 11% of total operating costs, up from 8% a year earlier. Of the Q2 figure, $74.9m is the Eigen AI acquisition, plus $41.0m of "post-combination compensation expense" recognised in product development — that is $115.9m of deal consideration routed through the income statement and then added back to adjusted EBITDA (6K-Q2-MDA). ⚠️ Flag 2.
3. Cash conversion. Operating cash flow / net income is meaningless here — both numerators are dominated by non-operating items. The usable version: OCF $4,504.1m against a $303.9m operating loss, with $4,395.0m of it explained by the deferred-revenue increase. Cash conversion is, precisely, prepayment. ⚠️ Flag 3.
4. Working capital. Receivables fell from $720.3m to $288.6m while revenue rose 529%. That is the good direction and it is the mirror of the prepayment model — customers pay before delivery. No flag. Worth watching: a receivables balance that starts rising again is the first sign the prepayment terms have stopped being available.
5. Costs moved onto the balance sheet. ❗ The useful life of server and network equipment was extended from four years to five in Q1 2026: "Starting in the first quarter of 2026, 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 and partially offset the increase in depreciation expense during the period" (6K-Q2-MDA). The company does not quantify the offset. The material weakness still open at the same date is in controls over depreciation start dates. The estimate that most moves reported profit sits inside the control area the auditor says is not effective. ⚠️ Flag 4, and the most important line in this dossier.
6. Sector equivalents: n/a.

EQ: 4 flags — adj. EBITDA bridge exceeds the operating loss; SBC 14% of revenue with $115.9m of deal consideration expensed and added back; OCF is customer prepayment ($4,395.0m of $4,504.1m); server life 4y→5y in Q1 2026, unquantified, inside an open material weakness on depreciation start dates.

Transcripts, read Q&A-first (SKILL §3.1)

Both 2026 calls are IR-moderated. There is no live analyst queue: the Head of Investor Relations reads questions submitted through a portal and hands each to a named executive, often re-framing it on the way. Sell-side names are attached to the questions (Goldman Sachs, Morgan Stanley, Citi, Wolfe, BNP Paribas, Bank of America, Baird, Cantor, Northland, Water Tower, Arete), but nobody gets a follow-up. This is a disclosure-quality fact about the name and it makes every dodge below cheaper to execute than it would be on an open call.

Guidance changes (step 2 of §3.1)

• Q1 2026: 2026 capex raised to $20–25bn "which is up from our prior range of $16 billion to $20 billion" (Volozh). Contracted-power target raised from "more than 3 gigawatts" to "at least 4 gigawatts". ARR, revenue and margin guidance reiterated.
• Q2 2026: all financial guidance reaffirmed — "annualized run rate revenue of $7 billion to $9 billion, group revenue of between $3 billion and $3.4 billion, group adjusted EBITDA margin of approximately 40%, and capital expenditures of $20 billion to $25 billion" (Alonso). Contracted-power target raised again, to 5 GW. Connected power held at 800MW–1GW.
• Q2 2026 new disclosure: ARR "$3 billion at the end of June, up 598% and increased 56% (sic) [ 58% ] from $1.9 billion at the end of March"; prepayments "will bring in more than $9 billion of upfront funding this year"; first asset-backed facility $775m at SOFR + 250bp; "more than $40 billion of committed backlog" to borrow against.

What analysts asked about (step 3) — in order of weight: 2027 pricing and when a 2027 outlook arrives; how the capex is funded and whether debt is still the right instrument; the Vineland, New Jersey site and the adjourned planning hearing; the gap between connected power and revenue; customer concentration against Meta and Microsoft; the margin step-down guided for Q2; whether new entrants selling compute at premium prices change the market. The questions are consistently about financing and timing, not about demand — which is itself the market's read.

Dodges (step 4) — three, and two of them are the same topic:

• Q2 2026 · what the $40–50m/MW price actually is · Stefan Slowinski (BNP Paribas). The question, as IR put it: "specifically around our $40 million per megawatt monetization in Q3. Is this reflective of initial pricing from selling Vera Rubin capacity?" IR then re-framed the hand-off — "Andrey, do you want to talk about the ramp of Vera Rubin capacity?" — and Andrey Korolenko answered entirely about deployment timing ("we expect to start deploying the Vera Rubins late this year or early next year"). Whether the premium price is Vera Rubin capacity was never addressed. Locator: CALL-Q2-26, Korolenko's turn following the Slowinski question.
• Q2 2026 · the ACV-per-megawatt threshold · Tyler Radke (Citi). Asked "Is there a threshold of ACV per megawatt you are waiting for on the longer-dated deals to sign into those?" Marc Boroditsky answered "what we strive to do is to optimize across customer type, price, payment structure, duration and deal size, rather than strictly looking at a single variable or a compiled metric … So a bit wider than purely ACV per megawatt." No threshold given. Locator: CALL-Q2-26.
• Q1 2026 · the share of contracts repricing · Alex Duval (Goldman Sachs). Asked "is there a way for us to think about the share of older shorter-term contracts that could benefit from this pricing dynamic?" Boroditsky answered with qualitative strength — "We just raised prices again in the latest quarter, and we are still selling out across all chip types at the higher prices" — and with duration, ACV and prepayment trends, but never the share. Locator: CALL-Q1-26.

The thread to pull: contract-level unit economics were declined in two consecutive quarters, by the same executive in two of the three cases. Nebius will give a price range per megawatt and a payback characterisation ("less than 2-year payback period", CALL-Q2-26) but not the mix, the threshold, or which capacity earns which price. That is the single number the whole thesis turns on, and it is the one number management will not decompose. Carried to the scorecard's ⚠️ line, with an S9 trigger.

A near-dodge that is not scored as one: asked when a 2027 outlook arrives, Alonso gave a pricing baseline and said "we will provide a formal guidance later this year" — a deferral with a date, not an evasion.

S5 — Ownership

13F register at 2026-06-30 (13F-Q2-3S): 739 filers, 184,177,832 shares reported against 271,855,218 outstanding. Top holders:

⚠️ The top of the register is not conviction money. Goldman Sachs, Susquehanna, Jane Street, Citadel and Bank of America — five of the top eight — are dealers and market makers, and Goldman alone added 28.7m shares in the quarter Nebius issued $4.3bn of converts. A convertible issue creates exactly this: delta hedges booked as long equity. Reading the 13F as "institutions are buying" would be a misread. The genuinely discretionary names in the table are Orbis Allan Gray (trimming), Fred Alger (trimming) and two new entries, Lone Pine and Situational Awareness.

⚠️ Latest-quarter 13F ingestion lags (README warning, carried): ten prior holders have no 2026-06-30 filing ingested yet — Janus Henderson (6.4m), Value Aligned Research Advisors (6.2m), Accel Leaders 4 (4.6m), Morgan Stanley (3.7m), UBS Group (3.6m), Capital Research Global Investors (3.4m), Clear Street (3.1m), Accel Leaders 3 (2.3m), Geode (2.1m), UBS AM (1.6m). None of these is an exit; the script checked and says so.

Insiders (threespread.py insiders NBIS --days 365): 104 Form 4 filings in twelve months and 15 Form 144s since 2026-03-27. The ones inspected are small and RSU-driven — the most recent, a director's trust selling 6,364 shares for $1,350,377.16 at roughly $212, against 6,364 RSUs acquired 2025-01-03 (144-2026-09-15). Nothing here reads as a position being exited; it reads as vesting. Insider ownership percentage is not in this dossier: as an FPI Nebius files no DEF 14A, and the 20-F's beneficial-ownership table was not read at gate-minimum depth. Gap.

Founder control. Volozh holds Class B shares; the 20-F risk summary says so plainly — "The concentration of voting power with our founding shareholder limits the ability of our minority shareholders to influence corporate matters, including the election of directors" — and Nebius "is a 'Controlled Company' within the meaning of the Nasdaq Stock Market rules" (20F-2025).

Short interest (SI-FINRA): peaked at 61,068,467 shares (2026-07-15), fell to 45,272,901 at 2026-08-31 — a 26% cover in six weeks, through and after the Q2 print. Days to cover 2.0–3.5 throughout; this is not a squeeze setup. (Equibles also returned a model estimate of ~43.3m for the 2026-09-15 settlement, published 2026-09-24; it is an estimate, not a FINRA figure, and is not used.)

Shelf / ATM / buyback. ATM: 25,000,000 Class A shares authorised, 12.7m sold in Q2 2026 at a weighted average $224 for ~$2.8bn gross, 12.3m remaining (CALL-Q2-26). No buyback exists. The F-3ASR shelf (File No. 333-286932) is automatic and effectively unlimited — 6K-Q2-MDA incorporates the quarter's results into it by reference. A shelf hangs over this name permanently, and the ATM showed in one quarter that it will be used.

S7 — valuation table (FY2028), and its assumptions

At $236.12 (PX-2026-09-22) on ~292.9m economic shares (271.9m outstanding + 21.1m NVIDIA warrant shares), market value is ~$69.2bn. Net debt on carrying values is ~$0.5bn ($8,545.7m debt less $8,042.1m cash), so enterprise value is also ~$69.7bn — but the converts' fair value of $20,823.7m says the honest capital structure is nearer $69.2bn of equity plus $12.3bn of in-the-money optionality the share count does not yet carry.

There are no earnings and will not be for years, so the model runs on EV/EBITDA against 2026 guidance ($3.0–3.4bn revenue, ~40% adjusted EBITDA margin → $1.2–1.36bn of adjusted EBITDA, which is ~51× EV/adj. EBITDA on this year's guide).

Assumptions, stated so they can be attacked: revenue compounds off the $7–9bn exit-2026 ARR guide (base takes the midpoint and grows it ~35%/yr; bear assumes pricing halves as 2027 capacity lands; bull assumes the $40–50m/MW short-duration pricing generalises). Net debt of $14bn at FY2028 assumes another ~$12bn of asset-backed and corporate debt on top of today's converts and no further equity beyond the remaining ATM — which is conservative on dilution and optimistic on financing mix; every dollar raised as equity instead moves the share count up and the per-share numbers down. Management's guide is taken as given here because S8's verdict is reliable; if that changes, so does the base case. The 8×/12×/16× multiples are judgement, not observation: there is no stable comparable for a business this young.

What would have to be true for $236 to be right: FY2028 revenue near $14bn at a 40%+ adjusted EBITDA margin and a 12×+ multiple surviving the point at which the 2026–27 capex starts depreciating in full — i.e. the market is paying today for the base case landing with no financing accident and no pricing break in between. The spread between bear and bull is 26×, which is what a levered, pre-earnings, prepayment-funded infrastructure build should look like on paper.

No buy price and no position size (owner decision 2026-09-23).

S8 — Management credibility

The record is short by construction: Nebius has existed in its current form since the 2024 Yandex divestment, and only two earnings calls are reachable from this environment (ROIC.ai holds Q1 and Q2 2026 and nothing earlier; the founder's quarterly shareholder letters, where the ARR targets are set, are on the IR site and outside the egress policy). Five rows are scoreable and one is open.

Are the promisers still there? Yes — Volozh (founder/CEO), Alonso (CFO), Korolenko and Boroditsky all appear on both calls, and no Item 5.02-equivalent departure appears in the 6-K stream read for this pass. No row is departed. The 20-F does flag the team as unfinished: "We are continuing to build out our senior management team, and our success will depend on both hiring and retaining current and future key management" (20F-2025).

Dodge log rolled up: 3 dodges across the two calls; contract-level unit economics dodged in both quarters (the ACV-per-megawatt threshold, the share of contracts repricing, and what the $40–50m/MW price is attached to). Management guides the aggregate reliably and declines the decomposition consistently. Neither call took a live follow-up.

Verdict: ✅ reliable — with a thin record and a disclosure caveat. Five scored rows: two beats, two mets, one withdrawn. Nothing was missed; the one withdrawn row is a capex raise, which is spending more rather than delivering less, and the two beats are real (a margin guide beaten by nine points and a power target exceeded two quarters early). Against that: six quarters is not a cycle, the only downturn this team has managed at Nebius is the one it hasn't had yet, and the promises that get kept are the ones management chose to make — aggregate power, aggregate revenue, aggregate margin — while the per-contract economics behind them have been declined twice. The S7 base case therefore takes the guide as given, and the moment a first guide is cut rather than raised, this verdict is stale.

S9 — Monitoring

• Next earnings: Q3 2026 results, expected mid-November 2026 — derived from the filing cadence (Q3 2025 released 2025-11-12, Q4 2025 on 2026-02-12, Q1 2026 on 2026-05-13, Q2 2026 on 2026-08-12), not from an IR confirmation. ROIC.ai lists no scheduled call. Derived, not sourced.
• Filing feed: the Tier 0 EDGAR sweep covers cik 1513845. Note that as an FPI, Nebius's substantive disclosure arrives as 6-K exhibits rather than 10-Qs, so the sweep sees the wrapper — the exhibit index has to be opened to know what a 6-K contains. The 2026-02-12 auditor change arrived this way and would have been easy to miss.
• Leading indicator between quarters: the deferred-revenue balance (currently $5,975.2m). It is disclosed every quarter, it is the mechanism that funds the build, and it turns before revenue does. Prepayment coverage falling below the "50%, 60% of the associated CapEx" management quoted means either the customers changed or the terms did. The receivables line is the same signal from the other side.
• The trigger that converts this name to a buy is the first quarter in which capacity deployed in 2026 shows up as revenue at a disclosed price per megawatt — i.e. Q4 2026 or Q1 2027, when the contracts Volozh described at $20–25m/MW and the auction and short-duration deals at $40–50m/MW come online together. Until then the price-per-megawatt range is management's characterisation of its own book and has been declined twice under questioning. A quarter that prints revenue consistent with the upper half of that range, with the ARR guide met and no new equity beyond the remaining 12.3m ATM shares, settles the argument the whole thesis rests on.
• Kill criteria (specific, testable):
• The prepayment engine stalls. Deferred revenue flat or falling quarter-on-quarter while capex continues. The model is financed by customers; when they stop paying up front, the equity pays instead.
• A first guide gets cut. The FY2026 ARR range ($7–9bn) or revenue range ($3.0–3.4bn) is lowered rather than reaffirmed. S8's verdict is built on the absence of this.
• Useful lives are extended again, or the fixed-asset material weakness is still open in the FY2026 20-F. Either one means the number that sets reported profit is being managed inside a control environment the auditor has called ineffective for three consecutive years.
• Customer D's receivable share stays above ~70% in the FY2026 20-F concentration note without a second name appearing beside it. 83% of receivables from one counterparty is a single point of failure dressed as a backlog.
• The converts stop being cheap. A new issue priced above 5%, or asset-backed paper above SOFR + 400bp, means the market has re-rated the build — and at $20–25bn a year of capex, the cost of the marginal dollar is the business.