Nebius Group N.V. [NBIS] — Valuation Analysis

⚠️ SUPERSEDED IN PART — 2026-07-29

The position verdict in this document is retired. Under the current framework (references/criteria.md, 2026-07-29) the memo outputs an analysis, not a position. Whether an analysis justifies a position is a question about a particular book, and two books answer it differently.

The Gate block and the Gate 4 expected-return arithmetic below are also superseded, by the named Criteria (each with a type: BINDING or MEASURED, returning PASS / FAIL / INDETERMINATE), the reverse-DCF implied-path test, and a separate 12-month target.

→ Current analysis: Nebius_Criteria_and_Valuation_2026-07-29.md

Everything else here — the research, the evidence, the mechanism work — stands. Residual references to "Watchlist" in the prose below are the historical record of the 2026-07-27 assessment and are left intact deliberately.

Task 3 of the Equity Underwriting Memo · As of 2026-07-27

Current price: $184.92 (Alpaca SIP daily bar, 2026-07-27 intraday) 12-month price target (Base case DCF): $175 · Probability-weighted value: $190.56 Scenario range: $9 (Bear) / $175 (Base) / $436 (Bull) Scenario-weighted E[R]: +3.1% — against a 4.7% cash hurdle

All valuation outputs below were built as live formulas in Nebius_Financial_Model_2026-07-27.xlsx and independently verified by evaluating the workbook's formula graph (formulas library) and cross-checking every result against a separately written Python implementation. Both agree to the cent. Verification log is in §8.

1. Valuation approach, and why the usual approach fails here

Nebius cannot be sensibly valued on any near-term multiple. Trailing revenue is $878m (TTM through Q1'26) against a $47.5bn enterprise value — 54x sales — and there is no GAAP profit to capitalise. Nor can it be valued on a "normalised" basis, because nothing about the current year is normal: capex guidance is $20–25bn against $3.3bn of revenue.

So the valuation is a DCF, and the DCF has one honest structural property that must be stated before any number:

In the Base case, 167% of enterprise value sits in the terminal value. The explicit 2026–2031 period contributes negative $30.2bn of present value. Every dollar of equity value in this name is a claim on cash flows that begin after 2030.

That is not a modelling flaw to be engineered away; it is the actual economic shape of a company spending $22bn/yr to build assets that earn later. But it means the valuation's precision is illusory and its sensitivity is the real output. A one-turn move in the terminal multiple is worth ±$33/share — 18% of the current price. Anyone who quotes a DCF price target on this name to two decimal places without saying that is selling false confidence.

Method weights

Explicitly not used: EV/EBITDA on current-year figures (meaningless at 38x); P/E (no earnings); P/B (book value is a 12-month-old snapshot of an asset base that doubled since); DCF perpetuity growth (a 5-year-life asset base makes a Gordon-growth terminal economically incoherent — an exit multiple with an asset-salvage floor is the correct structure and is what the model uses).

2. Assumptions and the base-rate check

2.1 Cost of capital

Cash taxes are modelled at 0% through 2028 (NOL shelter), 12% in 2029, 15% thereafter.

2.2 Depreciation — modelled from a capex vintage waterfall, not a ratio

80% of capex on a 5-year life (servers/network — the life the company itself moved to in Q1'26), 20% on a 25-year life (buildings, land, power infrastructure), half-year convention on new capex, plus the existing $7,131.7m PP&E over a remaining 4.5 years. This is the most consequential mechanical assumption in the model: Base-case D&A runs from $3.4bn in 2026 to $9.0bn in 2031, and it is what keeps EBIT negative until 2031 even as adjusted EBITDA reaches $14.5bn.

Flagged, not buried: the company extended server/network life from 4 to 5 years starting Q1'26. I have adopted the 5-year life because it is the company's stated policy and matches peers — but this reduces modelled depreciation by roughly 20% of the equipment charge, and every reported EBIT/net-income comparison across the 2025/2026 boundary is flattered by it. If the correct life is 4 years (a defensible view given Vera Rubin arrives in H2'26), Base-case 2031 D&A rises by ~$2.2bn and the Base value per share falls by roughly $20.

2.3 The three scenarios

Scenario definitions in words:

• Bull — the contracted book converts fully, the uncontracted 2027+ capacity re-lets at rising prices, and the asset-backed financing template (July 2026) repeats across the $40bn+ commitment book so equity dilution stays minimal. Nebius becomes the third meaningful AI-cloud platform behind the hyperscalers.
• Base — the company delivers roughly what it has guided to, hits $8bn exit ARR, and grows into a $29bn revenue business by 2031 at 50% EBITDA margins. Capacity re-lets at flat-to-modestly-lower prices as supply arrives. Financing is available but the equity funds ~35% of the gap.
• Bear — the contracted book is delivered and nothing more. Revenue plateaus around $8.5–9bn as the Meta and Microsoft terms mature with no incremental demand at economic prices; management cuts uncommitted 2026 capex to $11bn once demand cracks; the GPU fleet depreciates faster than it re-lets. Note that this Bear does not require a default, a fraud, or a contract cancellation. It requires only that the AI compute market stops growing at the pace the buildout assumes. That is why it carries 30% probability.

2.4 Base-rate check (required — Chan, Karceski & Lakonishok 2003)

Reference class: firms above $1bn of revenue sustaining >20% revenue growth for five or more consecutive years. The empirical base rate is roughly 3% of firms; growth persistence beyond chance is close to nonexistent, and the modal error in financial models is extrapolating the recent growth rate.

The prior being overridden, stated explicitly: even the Bear case assumes growth that the historical record says is uncommon. The Base case assumes a top-decile-or-better outcome. This memo does not assume that away.

Where the override is defensible: 2026–2028. At 2026-03-31 the company held $33.6bn of remaining performance obligations — 1.5x its entire asset base — from Microsoft and Meta, with $4,778.1m of customer cash already collected. Contracted revenue is not a forecast; it is an obligation. The near years of the forecast rest on signed paper, not extrapolation.

Where the override is NOT defensible: 2029–2031. There is no contractual support for those years. The initial Microsoft and Meta terms run roughly through 2030–2032, and whether that capacity re-lets profitably after them is genuinely unknowable today. This is why the terminal multiple in the Bear is 7.0x rather than the Base's 9.0x, and it is why the memo's conclusion turns on the width of the distribution rather than its midpoint.

Margin base rate, in contrast, is not the aggressive assumption. Q1'26 delivered a 74% gross margin (pre-D&A) and a 45% adjusted EBITDA margin in the AI cloud business. Every forecast margin in the model sits at or below what has already been reported. If this valuation is wrong on the upside, it will not be because of margins.

3. Street & positioning context

3.1 Consensus estimates — and an honest data gap

Alpha Vantage EARNINGS_ESTIMATES returned {"symbol":"NBIS","estimates":[]} — no coverage for this foreign private issuer. The estimate-revision counts that endpoint normally provides are therefore not available for this name, and the corresponding scorecard row below is marked not computable rather than silently dropped. Consensus revenue/EPS is web-aggregator sourced (stockanalysis.com, 18 analysts, retrieved 2026-07-27), and the count differs across sources (Benzinga cites 17) — that variance is disclosed, not hidden.

This is the single most important finding in the memo. The house forecast is within 3% of Street on 2026 and within 7% on 2027 — and below, not above. There is no positive variant view on the numbers. Whatever disagreement exists between the house view and the Street is therefore not about fundamentals; it can only be about the multiple and the terminal.

3.2 Consensus rating and price target — and the required bridge

The house target of $175 is 32% below the Street average of $258. The required decomposition — is that gap about the numbers, or about the multiple?

Take the shared FY2027E revenue base of $9,800m (house) and reverse-engineer the EV/revenue multiple implied at each price point, holding the revenue base constant so the multiple is the only variable:

The gap is entirely a multiple disagreement, and the proof is above. House and Street agree on 2027 revenue within 7%. The Street's $258 requires the FY2027 EV/revenue multiple to expand from 4.85x to 6.76x — a 39% re-rating — at a moment when the closest listed comparable trades at 2.87x and has fallen 41% in twelve months. The house target requires the multiple to compress mildly, to 4.59x, which is still a 60% premium to CoreWeave.

This produces a specific, checkable claim: twelve months from now, either NBIS's EV/forward-revenue multiple expanded toward 6.8x (Street right) or it held/compressed toward 4.6x (house right). That is falsifiable, and it is logged in the recommendation ledger as the exact framing to score.

Honest reasons the Street might be right, not cherry-picked: 1. Sell-side targets skew structurally bullish industry-wide — a useful prior, but nowhere near enough to explain a 39% multiple gap on its own. 2. The recent target raises (Goldman 2026-07-01, Northland 2026-07-20) cluster immediately after the Meta deal and the NVIDIA stake — consistent with extrapolating recent news rather than independently stress-testing the terminal. 3. The genuine risk to the house view, stated plainly: if the premium to CoreWeave is structural rather than a reversion candidate — justified by the net-cash balance sheet, the owned-power position, NVIDIA's 9.3% stake, and the demonstrated ability to fund contracted deployments with asset-backed debt — then the house's peer-multiple anchor will systematically miss for as long as that regime holds. NBIS genuinely is a better-financed business than CoreWeave. A premium is warranted; the question is only how much, and I have no strong evidence that 4.6x is right and 6.8x is wrong beyond the DCF, which is itself dominated by an unobservable terminal.

3.3 Short interest and days-to-cover — read as signal, and then corrected

The Asquith/Pathak/Ritter reading — high and rising short interest predicts underperformance — would score this as a strong headwind. I think that reading is wrong here, and the reason is arithmetic:

NBIS's Q1'26 diluted share count (309.0m, if-converted) exceeds its basic count (258.3m) by 50.7m shares of convertible-note dilution. Short interest is 61.0m shares. The company raised $4.3bn of convertible senior notes in Q1'26 alone, and convertible arbitrage desks hedge that issuance by shorting the underlying delta. Two independent facts corroborate: (a) the entire neocloud peer group carries 24–29% short interest despite wildly different fundamentals and price trajectories, which is a capital-structure signature, not a fundamental one; and (b) Susquehanna and Citadel appear among the largest institutional holders — market makers and multi-strategy firms, exactly the profile that runs convert arb.

Conclusion: the majority of NBIS's short interest is a hedging artefact. Reading it as bearish conviction would be a serious misread. It does, however, mean two real things: (i) squeeze risk on upside catalysts is genuine, because delta-hedgers must buy as the stock rises through conversion prices; and (ii) the borrow is already heavily used, which matters if anyone ever wanted the short side.

3.4 Institutional ownership composition

Institutional ownership 52.9%; insider 16.3%; 752 institutions filing 13D/G or 13F holding 119.3m shares (fintel.io/secform4.com, 2026-07-27). Largest holders: BlackRock (index/mandate capital), Fred Alger Management, Susquehanna International Group (market-maker/convert arb), Orbis Allan Gray, Citadel Advisors (multi-strategy). Most recent quarter: institutions bought 774k shares and sold 4.4m — net selling.

Read honestly: this is not a high-conviction active shareholder register. Two of the five largest holders are almost certainly holding hedged inventory rather than directional risk; one is an index mandate. The genuinely active, valuation-driven money is Orbis Allan Gray and Fred Alger. The one unambiguous conviction holder is NVIDIA at ~9.3% — which is also the company's principal supplier, so the signal is real but not disinterested.

3.5 Options market read

Alpaca options snapshots, 2026-07-27 18:43 UTC.

Realised volatility, computed from Alpaca daily bars: 137.1% (21d) / 117.9% (63d) / 104.5% (252d).

Implied sits above realised across the entire curve — a large, positive variance risk premium, exactly the condition under which buying options is systematically negative-expected-value (Coval & Shumway 2001; Bakshi & Kapadia 2003). Skew is close to flat: the 18-delta Sep put marks 147.2% IV against the 33-delta Sep call at 139.9% — a mild put bid, far less pronounced than a typical equity index, consistent with two-sided demand in a momentum name.

The load-bearing number for Task 5: the options market prices a ±33.3% move by 21-Aug. The house 12-month Base view is −5.4%. There is no window in which the house expected move exceeds the implied move. Naked long premium is ruled out on arithmetic, not preference.

3A. Transcript Mention-Frequency table (REQUIRED — references/mention-frequency.md)

History window: 6 quarters, Q4-2024 through Q1-2026. Nebius has held only seven earnings calls as a standalone company (the Yandex divestiture closed July 2024; the first Nebius report was Q3-2024). Six of the seven were retrieved; Q3-2024 could not be obtained. Pre-Q3-2024 calls are Yandex N.V. and are not comparable. Sources: InsiderMonkey (24Q4–25Q3) and stockanalysis.com (25Q4, 26Q1), with a cross-source control check on "revenue" and "customers" showing no level shift at the boundary. Chart: 36_mention_frequency.png. Counts: data/mention_freq_NBIS.json.

3A.1 The one count that required reading the text, not the number

"pricing" going 1 → 6 → 7 is directionless as a count. The passages resolve it, and they matter more than any other single sentence in this refresh:

Prepared remarks, Q1-2026: "On pricing, strong market demand is translating into pricing gains in our latest deals."

Q&A, Q1-2026, CRO Marc Boroditsky answering Goldman Sachs on older-generation GPU pricing: "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."

The v1.3.0 memo listed "a demonstrated re-lease of first-generation capacity at flat-or-better pricing" as one of the specific things that would cut p_bear. This is the first direct evidence on it, and it points the bullish way. Weighting, stated honestly: it is a management assertion in Q&A, not a disclosed cohort metric, and no issuer in this sector publishes one. It is carried into Task 5 as a 3-percentage-point reduction in p_bear, not as a thesis change.

In prose. Emerging: GW, MW, contracted, Meta, Microsoft, pricing, margin, financing, capex, cash, token, Pennsylvania — the vocabulary of a capital-formation and capacity company. Decaying: Blackwell, H200, Toloka, Avride, Israel, Kansas City, Finland, enterprise, data centers, and — most notably — ARR. Never present at all: RPO, backlog, useful life, re-lease, interconnection, permit, grid.

4. FACTOR & ANOMALY SCORECARD

Screen context, carried honestly. NBIS ranked 45th of 47 scored long candidates on the 2026-07-27 screen (reports/screens/Screen_2026-07-27.md), with a composite of −2.87 — the third-worst on the long board. The screen's computed factors were: 12-1 momentum +368%, 1-month −22.6%, 64% of 52-week high, +31% vs 200dma, GP/A 0.03, accruals −0.04, asset growth +250%, F-score 7/9. The screen itself flagged that "the rank is driven almost entirely by +250% asset growth and 0.03 GP/A" and instructed the memo to adjudicate that directly.

That adjudication is §4.2 below. Neither repeating the penalty nor waving it away is acceptable; the table below shows where I have and have not overridden the screen.

4.1 The scorecard — read for a LONG position

4.2 THE ADJUDICATION: is +250% asset growth the business model, or empire-building?

This is the question the screen explicitly handed to this memo. My answer is partly the business model, and the screen's penalty should be roughly halved — but not removed, and I can say exactly which parts survive.

What Cooper, Gulen & Schill (2008) actually documented. The asset-growth effect is one of the most robust cross-sectional anomalies in the record, and the authors and subsequent literature attribute it to three distinguishable channels: 1. Empire-building / agency — managers investing beyond the opportunity set, typically via acquisitions. 2. Financing-driven over-investment — firms issuing equity or debt into a receptive market and deploying it at below cost of capital, with the market slow to mark the return down. 3. Over-extrapolation — investors extrapolating the growth that the asset build implies, and being disappointed.

The signal's predictive power is strongest where investment is discretionary and uncontracted. So the correct question is not "is 250% a big number" — it obviously is — but which of these three channels is actually operating.

Channel 1 (empire-building): does not apply, and I can evidence that. - Goodwill went from $0 at 2025-12-31 to $163.3m at 2026-03-31. The entire 2026 M&A programme (Tavily, Eigen AI, Clarifai) is $163m of talent-and-technology acquisition against $22bn of planned capex. This is not an acquisitive asset build, which is the archetype the anomaly was documented on. - The asset growth is overwhelmingly PP&E: $846.7m → $5,553.3m → $7,131.7m across FY24 / FY25 / Q1'26. These are GPUs, data centres, land and power — not acquired balance sheets.

Channel 2 (uncontracted over-investment): substantially rebutted, with numbers. - RPO / total assets = 1.72x at 2025-12-31 and 1.51x at 2026-03-31. Every dollar of the asset base is covered one-and-a-half times over by signed, multi-year obligations from Microsoft and Meta — two investment-grade counterparties. - $4,778.1m of deferred revenue at 2026-03-31 — cash the customers have already paid for capacity that does not yet exist. Q1'26 operating cash flow of +$2,258.0m on $399.0m of revenue is explained almost entirely by a $3,198.0m deferred-revenue inflow. Customers are pre-funding the buildout. That is the opposite of the archetype. - The 2026-07-17 senior secured facility demonstrates that the contracted deployments can be financed on their own cash flows: "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 repeatable across the $40bn+ commitment book, the capex is largely self-funding rather than shareholder-funded.

Channel 3 (over-extrapolation): fully alive, and this is what survives. - The anomaly is not only about management behaviour; it is about price. NBIS is +254% over twelve months and +121% year-to-date. That is precisely the extrapolation the asset-growth effect punishes, and no amount of contractual backing neutralises it. - Contracted is not the same as economic. RPO tells you revenue will arrive; it tells you nothing about whether the assets earn their cost of capital. My own Bear case is exactly this: the contracted book is delivered in full, revenue reaches $8.5–9bn, and the equity is still worth $9. Contracts do not save you from spending $26–78bn on assets with a five-year life. - An accounting flag rides alongside it. The Q1'26 extension of server life from four to five years reduces depreciation on the very asset base whose growth is in question. It is legitimate and peer-consistent, but it means the "these assets haven't had a year to earn yet" defence is being made slightly easier by an accounting choice rather than purely by timing.

Verdict on asset growth: marked down from "worst reading on the long board / primary driver of a −2.87 composite" to a NEUTRAL-TO-MILD HEADWIND. Two of the three channels through which the anomaly operates are rebutted by disclosed, sourced facts. The third — investor over-extrapolation — is not, and the trailing return says it is live. The screen's mechanical penalty was directionally correct in flagging the name for scrutiny and quantitatively wrong in treating it as the dominant negative.

Verdict on gross profitability: marked down from "very weak" to NOT MEANINGFUL. GP/A of 0.029 divides FY2025's gross profit by a year-end asset base that tripled during that same year — the denominator ends the period at 3.5x where it started. This is a pure timing artefact. Computing it the way Novy-Marx intended (a stable franchise's profitability against its productive base), on the Q1'26 run-rate against the Q1'26 asset base, gives 0.053 and rising sharply, with a 74% gross margin underneath it. Novy-Marx's signal identifies durably profitable franchises; it does not usefully rank a company mid-buildout. I decline to score this row either way and say so, rather than reporting a number I do not believe carries information.

What I am NOT doing: I am not using this adjudication to convert a factor headwind into a factor tailwind. A company with 0.05 GP/A and 250% asset growth is not a quality factor long by any construction. The honest outcome is that two of the ten rows become uninformative rather than negative — which materially changes the composite rank but does not create a positive quantitative case.

4.3 Synthesis — do the anomalies support, contradict, or split on the fundamental view?

They split, and they split in an unusually informative way.

• The momentum block is internally contradictory. Top-decile 12-1 momentum (strong tailwind for a long) sits against 64.5% of the 52-week high and a −27.9% month (headwind). This is the classic momentum-crash setup Daniel & Moskowitz (2016) describe: a high-momentum name that has begun to reverse hard is where momentum strategies suffer their worst drawdowns. It is not a clean "buy the winner" signal, and it is not a clean "the trend has broken" signal either. It is genuinely ambiguous, and Gate 6 must treat it as such.
• The quality block is benign but low-information. F-score 7/9 is good. Accruals are negative (favourable) but for reasons that are accounting artefacts on both sides of the ratio. Gross profitability is not computable in a way that means anything. Nothing here corroborates a deteriorating-fundamentals story — which matters, because it means any short case on this name would need to be carried entirely by narrative, with the quantitative record actively disagreeing. That closes off the short side.
• The valuation/expectations block is where the real answer is. The scorecard does not, and cannot, tell you whether $47.5bn is the right enterprise value for a company that will be worth either $110bn or $2.6bn depending on how the AI capex cycle resolves.

Net: the anomaly evidence is not a reason to be long, and not a reason to be short. It removes the screen's apparent verdict (the composite of −2.87 overstates the case, as adjudicated above), and it removes the short side (clean earnings quality, strong momentum, F-score 7/9). What it leaves standing is the fundamental question, which §5–§7 answer.

5. DCF output

5.1 Base case free cash flow

5.2 Sensitivity — Base-case value per share

The grid spans $60 to $307 — a 5x range — on assumptions that are all individually defensible. Note where the current price of $184.92 sits: between the 9x and 10x columns at the house WACC. The market is inside the plausible range, not outside it.

5.3 Reverse DCF — what $184.92 already assumes

Working backwards from the market enterprise value of $47,538m and the house Base explicit-period cash flows:

The current price discounts the house Base case almost exactly — a 3% gap. This is the finding that decides the memo. There is no valuation gap to monetise in either direction. The stock is not cheap and it is not expensive relative to a defensible central forecast; it is fairly priced for the Base case and priced with no compensation for the width of the distribution around it.

6. Comparable company analysis

IREN and APLD are excluded from the multiple conclusions rather than estimated: neither has usable published FY2026/27 consensus revenue, and inventing one to fill a cell would be exactly the kind of false precision this process is meant to avoid.

Three conclusions:

1. NBIS trades at a 58% premium to CoreWeave on FY2027E EV/revenue (4.53x vs 2.87x). At CoreWeave's multiple, NBIS is worth $117/share.
2. The premium is partly earned. NBIS has a net-flat balance sheet against CoreWeave's $32.9bn of net debt, a higher gross margin, an NVIDIA equity stake, and 75%+ owned rather than leased power. If you had to own one of these two through a financing squeeze, it would be NBIS without hesitation. A premium is right; 58% is the question.
3. The "the whole group is re-rating together, so it's a sector-regime question" defence does not apply here — it applies in reverse. CoreWeave is −41% and Oracle −50% over twelve months while NBIS is +254%. The group is de-rating; NBIS is the outlier that re-rated. Its premium is idiosyncratic, which makes it both more defensible (it is attributable to specific, identifiable facts) and more fragile (those facts are already known and priced).

Implied value per share at various FY2027E EV/revenue multiples (on the house $9,800m):

(Table computed on the house FY27E revenue of $9,800m; the model's Comps tab uses the Street $10,500m and therefore shows $117 / $163 / $185 / $225 / $286 — the difference is the 7% revenue delta, disclosed rather than reconciled away.)

6A. Implied penetration in physical units (REQUIRED — references/tam-sizing.md)

Conventional multiples do the least work on this name (EV/TTM revenue 54.1x is meaningless mid-buildout), so the required statement is made in megawatts. Full build in Research Document §7A.4.

The statement. At today's price Nebius must connect and monetise 3.26 GW by 2031 — 93% of everything it has contracted — merely to be worth $174.88, which is 5.4% below where it trades. The market is already paying for the whole contracted book to be built, energised and sold out at current rates. There is essentially no un-priced capacity left to underwrite as optionality.

This is the number that decides Gate 2B in Task 5: the capacity the transcript signal points at (§3A) is the same capacity the price already discounts, so it is context, not a variant.

7. Valuation conclusion

The methods do not disagree about the centre. They disagree about the width. DCF Base $175, reverse DCF $185, current price $185, comps at the current multiple $185, probability-weighted $191 — a cluster within 9%. Meanwhile the honest scenario range is $9 to $436, a 48-fold spread, and the Street's own targets span 3.4x.

That is the finding. NBIS is fairly valued on any central estimate and carries an extraordinary variance around it. A valuation-driven Long requires a gap that does not exist here; a valuation-driven Short would be precisely the error this process exists to prevent — and would additionally be fighting top-decile momentum, a 7/9 F-score, benign accruals, and a company with net cash and $33.6bn of contracted backlog.

The decision therefore cannot be made in the valuation section. It is made in Task 5, on the gates.

8. Verification log

The Excel workbook Nebius_Financial_Model_2026-07-27.xlsx was built entirely with live formulas (yellow cells are inputs; blue cells are formulas). Microsoft Excel on this machine would not open a workbook under automation, so the workbook's formula graph was instead evaluated independently with the formulas library and every output cross-checked against a separately written Python implementation (data/valuation.py). Two real bugs were found and fixed by this process before it agreed:

1. The Scenarios-tab sensitivity grid referenced the SBC row instead of the EBITDA row for the terminal value (off-by-one in the DCF block's row arithmetic). Fixed.
2. A label in column A began with =, which Excel and the parser both read as a malformed formula. Fixed — and this is very likely why Excel refused to open the first build at all.

One reconciliation to disclose. The Python implementation additionally charges a financing-dilution adjustment — the value transferred to new shareholders if the cumulative funding gap is met partly by equity issued below intrinsic value. Applying it lowers the Bull case from $436.13 to $416.72, the probability-weighted value from $190.56 to $185.71, and E[R] from +3.05% to +0.4%. The Excel workbook is the model of record and excludes it; both figures fail the 4.7% cash hurdle, so the conclusion is unchanged either way. Reported here rather than picking the flattering one.