Nebius Group N.V. [NBIS] — Trade Construction & Risk Management

⚠️ SUPERSEDED IN PART — 2026-07-29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The Gate block and the Gate 4 expected-return arithmetic below are also superseded, by the named Criteria (each with a type: BINDING or MEASURED, returning PASS / FAIL / INDETERMINATE), the reverse-DCF implied-path test, and a separate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 5 of the Equity Underwriting Memo · As of 2026-07-27

POSITION VERDICT: RETIRED 2026-07-29 — the memo outputs an analysis, not a position.
                  See Nebius_Criteria_and_Valuation_2026-07-29.md for the current Criteria block.
The gate lines below are the 2026-07-27 record and are superseded.

GATES: 1 (Mechanism):              PASS  — improving fundamentals evidenced, and quantitatively corroborated
       2A (Estimate variant):      FAIL  — house is 3-7% BELOW Street on revenue; no positive variant exists
       2B (Duration/optionality):  FAIL  — NEW under v1.4.0; tested against all four evidence legs, fails leg 4
                                           (the price already embeds 93% of contracted power). See §2.2B.
       3 (Catalyst):               FAIL  — the Q2 print tests execution; the disagreement is about the terminal
       4 (Expected return):        FAIL  — E[R] +5.7% gross / +2.8% net vs a 4.7% cash hurdle (was +3.1%/+0.15%)
       5 (Feasibility):            PASS  — with a hard constraint: 105% realised vol, 161% implied vol
       6 (Momentum / tape):        PASS (conflicted) — with the 12-month trend, against the 1-month; named + mitigated

ENTRY (conversion trigger): $155  |  TARGET: $175 (Base DCF)  |  INVALIDATION: $132
TIME HORIZON: 18 months, anchored on the FY2026 exit-ARR print (Feb 2027) — not the Q2 print
SCENARIO-WEIGHTED E[R]: +5.7% gross (+3.1% dilution-adjusted, +2.8% net); still 1.9pp SHORT of the 4.7% hurdle
       p_bear cut 0.30 -> 0.27 on new transcript evidence of re-lease pricing (§2.2B leg 2). Gate 4 still fails.
       Simple: risking ~29% to make ~-5%  (to the Base target)  — which is precisely why the answer is no
SIZING: Conviction: LOW  |  Volatility: HIGH (105% realised)  |  Resulting size: 0% today; 1.5% on conversion
VEHICLE (on conversion only): outright equity, small. NOT options — see §6.
INVALIDATION TRIGGERS:
  - Weekly close below the 200-day moving average ($140.29 and rising) — kills the long-side watch entirely
  - Any equity raise via the ATM programme at or below $180 — signals the asset-backed financing template failed
  - Q2'26 ARR below $2.8bn, or FY26 revenue guidance cut below $3.0bn — Bear case is running

1. The four conclusions, kept separate

The failure mode this task exists to prevent is reasoning as one continuous narrative. So, deliberately disaggregated:

1.1 Fundamental conclusion

Improving, and improving fast, on evidence rather than narrative. Revenue +684% YoY to $399.0m in Q1'26; AI cloud ARR $1.25bn → $1.92bn in one quarter; gross margin 51% → 74%; adjusted EBITDA turned positive at +$129.5m group / 45% margin in the AI cloud unit; RPO $21.3bn → $33.6bn; $4.78bn of customer cash collected in advance; a first asset-backed facility priced at SOFR+250 covering >100% of the capex it funds. Every cost line fell sharply as a share of revenue. This is real operating leverage in filed numbers.

Against that: GAAP net income is uninformative (a $780.6m non-cash ClickHouse revaluation); the server useful life was extended from 4 to 5 years in the same quarter, flattering EBIT; customer concentration is material and undisclosed; and the FY2026 plan requires spending $20–25bn against $3.3bn of revenue.

1.2 Expectations conclusion

The market already expects this. House Base FY2026 revenue of $3,300m sits 2.9% below the Street's $3,400m; FY2027 $9,800m sits 6.7% below the Street's ~$10,500m. The Street rating is Buy across 18 analysts, with targets rising (Goldman $267→$286 on 2026-07-01; Northland to $410 on 2026-07-20).

ΔE = house − Street is negative and small. There is no positive expectations gap to monetise on the long side, and the negative gap is far too small — and pointed the wrong way relative to a Buy rating — to build a short on.

1.3 Valuation conclusion

Fairly valued at the centre; extraordinary variance around it. Base DCF $174.88 (−5.4%). Reverse DCF says today's $184.92 implies a 9.31x exit on Base 2031 EBITDA against the 9.0x the model uses — a 3% gap. Probability-weighted value $190.56. Scenario range $9 / $175 / $436.

Valuation therefore cannot originate a thesis here in either direction. It can only translate one, and §1.2 shows there is nothing to translate.

1.4 Portfolio conclusion

No position. Even if one believed the Base case with confidence, the scenario-weighted expected return of +3.1% does not beat holding T-bills at 4.7% — on a security with 105% annualised realised volatility, a 30%-probability near-total-loss scenario, and 0.70 correlation to a semis/AI-hardware cluster that would dominate any book built from the current screen. The correct action is to define the price and evidence at which this becomes attractive and wait for it.

2. Gate-by-gate

Gate 1 — Causal mechanism: PASS

A specific, named, evidenced mechanism that improves future fundamentals: capacity conversion. Nebius holds >3.5 GW of contracted power against ≤1 GW of connected power guided for year-end; each megawatt brought online converts contracted obligations into recognised revenue at a 74% pre-depreciation gross margin. This is not a hypothesis — Q1'26 is the mechanism visibly working (ARR +54% QoQ, cost of revenue 49%→26% of sales).

Quantitative corroboration standard (long-side mirror), applied honestly:

Gate 1 passes on the strength of the disclosed operating data, with the scorecard not adding much beyond the F-score. That is stated rather than dressed up.

Gate 2 — Variant vs. consensus: FAIL

This is the gate that decides the memo, and it fails unambiguously.

For a Long, Gate 2 requires the house to be materially above consensus on the metric the mechanism moves. The house is below. There is no positive variant. Gate 2 fails.

For a Short, the house's small negative revenue variant is nowhere near material enough (3–7%, well inside the Street's own $3.0–3.9bn dispersion), and it is contradicted by every quality signal — F-score 7/9, benign accruals, top-decile momentum, net-cash balance sheet. This is exactly the "expensive-but-not-a-short" tech trap the skill warns will recur, except NBIS is not even expensive on the house Base case. A short here would be a pure multiple-mean-reversion bet with no diversified quant architecture behind it — the specific error the NET build made and this process exists to prevent.

Where the house genuinely disagrees with the Street is the multiple, and the Task 3 bridge proves it: on a shared FY2027 revenue base, the Street's $258 target requires the EV/revenue multiple to expand from 4.85x to 6.76x, against a comparable (CoreWeave) at 2.87x. A multiple disagreement is a legitimate view; it is not a Gate-2 variant, because no operating event resolves it inside a normal horizon.

Gate 2B — Duration / optionality variant: FAIL (new test under v1.4.0)

v1.4.0 introduced a second route through Gate 2: a specific addressable opportunity that consensus is demonstrably not modelling. This is the only route by which NBIS could have passed Gate 2, since the house sits 3–7% below Street on the near-term numbers. It is tested here against the hard four-part evidence bar in references/trade-construction.md, and all four are required.

Gate 2B fails on leg 4, with leg 1 only partial. Per the reference's own language, "a duration variant supported by fewer than all four is narrative, not evidence, and fails the gate." Two of four is not close.

Why this matters more than the arithmetic. The temptation on a name like this is exactly the one v1.4.0 was written to enable and to discipline: management is visibly telegraphing a gigawatt-scale future, the independent corpus broadly confirms it, and it would be easy to call that a duration variant. It is not one, because TAM capture is not TAM expansion. Nebius is not opening a new addressable pool that the Street has missed; it is executing against a contracted pool the Street has already counted and the price has already capitalised. Distinguishing those two is the specific honesty rule in references/tam-sizing.md, and applying it here is what keeps a Watchlist a Watchlist.

Gate 2 therefore fails on both routes: 2A (no positive estimate variant) and 2B (no un-modelled opportunity).

Gate 3 — Catalyst that resolves this specific disagreement: FAIL

There is no shortage of dated catalysts (Task 6 lists eleven). The problem is that none of them resolves the actual disagreement.

The disagreement is about what a five-year-life GPU fleet is worth in 2031. No event inside eighteen months answers that. Per the gate's own language — "a catalyst that tests execution when the disagreement is actually about a multiple assumption does not satisfy this gate" — Gate 3 fails.

Gate 4 — Positive scenario-weighted expected return, net of costs: FAIL

v1.4.0 revision to the probabilities, and the one piece of new evidence behind it. The v1.3.0 memo named "a demonstrated re-lease of first-generation capacity at flat-or-better pricing" as a specific p_bear reducer and recorded that no such evidence existed. The generative transcript pass found it (Valuation §3A.1): the CRO, in Q1-2026 Q&A, states pricing is strong "across both old and new GPU generations", that Nebius "just raised prices again in the latest quarter", and that it is "still selling out across all chip types at the higher prices." That is a management assertion in Q&A rather than a disclosed cohort metric, and no issuer in this sector publishes one — so it is worth a 3-percentage-point cut to p_bear, not a re-rating.

v1.3.0:  E[R] = 0.30 × (−94.9%) + 0.45 × (−5.4%) + 0.25 × (+135.8%)  =  +3.05%
v1.4.0:  E[R] = 0.27 × (−94.9%) + 0.48 × (−5.4%) + 0.25 × (+135.8%)
              = −25.62%        +   (−2.61%)     +   +33.96%
              = +5.73%

Gate 4 still fails, by 1.9pp instead of 4.6pp. The gap has more than halved on one sentence of new evidence, which is itself the useful disclosure: this decision is extremely sensitive to p_bear, because the Bear is a near-total loss. At p_bear = 0.25 with the residual to Bull the net E[R] is +11.7%; at 0.20 it is +19.0%. A single disclosed re-lease metric would move this name from Watchlist to Long on arithmetic alone — and Gate 2 would still block it, which is the correct architecture and worth saying out loud.

Original v1.3.0 arithmetic, retained for calibration scoring:

E[R] = 0.30 × (−94.9%) + 0.45 × (−5.4%) + 0.25 × (+135.9%) = +3.07%

Gate 4 fails on both the optimistic and the conservative construction. The point-estimate risk/reward is worse still: risking 29% (to the $132 invalidation) to make −5% (to the $175 Base target). There is no formulation of this trade at $184.92 that produces an attractive payoff.

Sensitivity of the conclusion to the probabilities. Because the Bear is a near-total loss, the answer is highly sensitive to p_bear. For E[R] to reach 15% at the current price, p_bear must fall to roughly 20% with the residual going to Base. That is a real possibility — and it is exactly what a clean Q2 print plus a second large asset-backed financing would justify. It is why this is Watchlist rather than Avoid, and it is the precise mechanism by which the name converts.

Gate 5 — Implementation feasibility: PASS, with a binding volatility constraint

Mechanically feasible. But note what the volatility number does to sizing: a 1.5% position in a 105%-vol name contributes roughly the same book-level risk as a 6% position in a 25%-vol name. This is the constraint that caps the size even after conversion.

Gate 6 — Don't fight the tape without naming it: PASS (conflicted), with mitigation

Naming the conflict explicitly, as the gate requires: a long here is with the twelve-month trend and against the one-month and intermediate trend. This is the momentum-crash configuration Daniel & Moskowitz (2016) identify — a top-momentum name that has begun reversing violently is exactly where momentum strategies suffer their largest losses. The gate is a bar-raiser, not a veto, and it passes only because the memo (a) names the conflict here, in the decision output, and (b) applies a mitigation.

The stated mitigation, which is also the conversion trigger: do not initiate until either - price reaches $155 (where E[R] on unchanged scenarios rises to ≈ +23%, and where the entry sits comfortably above the rising 200dma), or - the trend re-establishes — a weekly close back above the 50-day MA with the 200dma intact — and in either case at one conviction tier below what the fundamentals alone would justify.

A weekly close below the 200dma removes the long-side watch entirely rather than making it cheaper: at that point the trend filter has flipped and the Bear scenario's 30% weight would need to rise, not fall.

3. Portfolio context — executing the portfolio-book contract

portfolio_book.json read at 2026-07-27. State: positions: [], watchlist: [NET], as_of: 2026-07-27.

1. Marginal correlation. The book holds no positions, so there is no realised pairwise correlation to compute against. That is not a pass by default — the relevant correlation is with what the book is about to hold. Computed from trailing-1-year daily returns (Alpaca SIP bars):

The correlation implication, stated plainly as required. Roughly 20 of the 47 scored long candidates on the 2026-07-27 screen belong to one semis / AI-hardware cluster, and MU is being underwritten concurrently. An NBIS long alongside an MU long would be correlated exposure to a single macro factor — the AI capital expenditure cycle — expressed through two tickers. Under the book's own rule ("two 3% positions at 0.8 correlation are a 6% position wearing two tickers"), any NBIS + MU combination must be sized as a single cluster against the 5% single-name maximum and the 25% sector-concentration limit, not as two independent positions. Concretely: if MU is initiated at 3%, the maximum NBIS could ever take on conversion is 2%, and the combined 5% would already be at the single-name cap for the cluster.

This is a genuine constraint on a book with zero positions and a screen output dominated by one theme, and it argues for taking the best expression of the AI-capex trade rather than several. On the evidence assembled here, NBIS is not that best expression at $184.92: it carries the group's highest forward multiple, its most extreme volatility, and its most binary terminal outcome.

2. Exposure math.

All within max_gross_exposure 100% and max_net_exposure 100%. No breach.

3. Sector concentration. Current AI-infrastructure cluster weight: 0%. On conversion at 1.5% (or 2% net of an MU position), cluster weight remains far inside the 25% max_sector_concentration. Not binding today; would become binding quickly if several screen candidates from the same cluster were initiated together.

4. Capital competition — the decisive test.

The book is not at its exposure limit, so no existing position would be displaced. The test is therefore purely against the cash hurdle — and NBIS fails it at the current price by 1.9 to 4.6 percentage points. The honest statement is the one the book skill invites: the book is 100% cash and the cash hurdle is currently winning. On this name, at this price, it should keep winning.

5. Book update. This decision writes an NBIS watchlist entry with its conversion trigger. Per the handoff instructions for this session, the entry is staged in PENDING_book_and_ledger_updates.json for the parent session to merge rather than written directly.

4. Entry / exit levels

Time horizon: 18 months, anchored on the FY2026 results / exit-ARR print (~February 2027), which is the first event that meaningfully recalibrates the 2027 base from which the terminal value is extrapolated. It is deliberately not the Q2 print — a Q2 beat would not change the house view, and saying so up front prevents the position from being talked into existence on a good quarter.

5. Position sizing logic

Conviction composite (weighted, scored honestly)

Per the skill's rule, a strong fundamental and balance-sheet score cannot compensate for near-zero Gate 2/3/4 scores; conviction is capped at Low regardless. And per the hard-gate rule, Low conviction with three failed gates is not a small position — it is no position.

Volatility tier

Realised volatility 104.5% annualised (252d), 137.1% (21d); beta 3.19 vs SPY (1.40 reported 5Y). Comfortably above the >45% threshold. HIGH.

Sizing grid (house max single-name weight 5%, per portfolio_book.json)

Grid output at Low conviction × High volatility: below minimum size → Watchlist. The grid and the gates agree independently, which is the useful check.

Size on conversion: if the conversion trigger fires and Gate 2/4 are re-satisfied (which requires either the price at $155 or a genuine positive variant emerging), conviction would rise to Medium at most, giving 1% at High volatility. Applying the Gate 6 momentum mitigation — one tier down if entering against the intermediate trend rather than after it re-establishes — the practical range is 1.0–1.5%, and capped at 2% in any scenario given the MU cluster interaction (§3).

Consensus positioning — stated, not skipped

This decision is against a Buy consensus that is currently strengthening: 18 analysts, average target $258.13 (+40% from spot), with Goldman raising to $286 on 2026-07-01 and Northland to $410 on 2026-07-20. Going against a strengthening consensus is real crowding and momentum risk.

Two things make this defensible rather than reckless: first, the disagreement is fully decomposed (Task 3 §3.2 — it is a multiple disagreement, proved on a shared revenue base, not a hand-wave); second, the decision is not to short into that consensus, it is to decline to pay up. Declining to buy against a bullish Street carries only opportunity cost. That asymmetry is why "Watchlist" is a materially safer way to be contrarian here than any active position would be.

6. Vehicle selection

Three ways to express a long view were compared. None is recommended today, because the decision is no position. The comparison determines what would be used on conversion.

IV-crush disclosure (required whenever a catalyst-dated option is discussed). Any structure held through the Q2 print (~2026-08-06, unconfirmed — one aggregator lists 2026-07-28) carries elevated pre-event implied volatility that typically collapses immediately after the event resolves, regardless of direction. The 21-Aug ATM straddle at 161.5% IV implies a ±33.3% move; if the actual move is smaller, both a correct long call and a correct long put can lose money to the combination of IV collapse and theta ($−0.62/day on the 21-Aug $185 call). This is not a one-sided way to express a view, and it is the second independent reason the options route is rejected.

Tactical catalyst trade: NONE RECOMMENDED. The skill's default is to produce a catalyst-dated contract alongside the structural position. Here there is no structural position, the house has no directional view inside the catalyst window, and the implied move exceeds the house expected move. Manufacturing a contract recommendation to satisfy a template would be precisely the negative-EV behaviour the gates exist to prevent. The recommendation ledger entry for this memo is therefore a structural_position record of type Watchlist with the scenario probabilities logged for calibration scoring — no contract.

7. Risk / reward statement

Scenario-weighted (the number that decides):

Simple version, for the front-of-report ticket: risking ~29% (spot to the $132 invalidation) to make ~−5% (spot to the $175 Base target). The point estimate and the scenario weighting agree: this is not a trade.

Flagged explicitly, per the gate's own instruction: a mildly positive E[R] with a severe tail is still a reason to pass — and here the tail is a 30%-probability 95% loss, which dominates everything else in the calculation. Note also the upside tail is capped by the convertible overhang: roughly 50m shares (≈20% of the base) of if-converted dilution re-engage on a large rally, so the Bull case is structurally harder to realise per share than the enterprise value suggests.

8. The seven questions every Long or Short must answer

Answered here even though the decision is Watchlist, because the answers are what would have to change.

1. What is the market wrong about? Nothing demonstrable. The market prices a 9.31x exit on my own Base 2031 EBITDA against the 9.0x I use. I have no evidence the market is wrong; I have a view that the distribution around that central case is far wider than the price compensates for.
2. What evidence supports the differentiated view? The differentiated view is about variance, not level: 167% of enterprise value in terminal value, a 5-year asset life against 5-year contracts, and a Bear case that requires no default or fraud — only that AI compute demand stops compounding.
3. Why will the disagreement resolve now rather than years from now? It will not, and that is Gate 3's failure. The earliest partial resolution is the FY2026 exit-ARR print in February 2027.
4. What prevents the company from growing into its valuation? Nothing prevents it — the Bull case is a live 25% possibility, and if it happens the stock more than doubles. That is precisely why this is not a short.
5. What would invalidate the view? A second and third large asset-backed financing at ≤SOFR+300 without equity issuance (would cut p_bear toward 20% and push E[R] to ~+19%); Q2 ARR above $3.2bn; a demonstrated re-lease of first-generation capacity at flat-or-better pricing. Update: the third of these is now partially evidenced — the CRO stated in Q1-2026 that pricing is strong across "both old and new GPU generations" and that prices were raised again with capacity still sold out. Weighted at 3pp of p_bear (§Gate 4) pending a disclosed metric. Note that even fully credited it would not change the decision, because Gate 2 fails on both 2A and 2B independently of expected return.
6. What is the adverse scenario? Missing a doubling. A 25% chance of +136% forgone is a real cost of waiting, and it should be stated as such rather than hidden — the conversion trigger at $155 may never print.
7. Why is taking no position superior to owning it? Because cash yields 4.7% with no variance, and the position's net expected return is roughly 0–3% with 105% annualised volatility and a 30% chance of losing nearly everything. The capital is better employed doing nothing, or waiting for the price where the same scenarios pay +23%.

8A. What the unpublished-scoping pass changed, and what it did not

9. What converts this to a Long

Both conditions, not either:

Resulting size on conversion: 1.0–1.5% (Medium conviction × High volatility, one tier down for the momentum conflict), hard-capped at 2% and treated as part of a single AI-capex cluster with any MU position.

What converts this to a Short: essentially nothing available today. It would require an evidenced deterioration mechanism (contract cancellation or renegotiation, a failed financing, disclosed utilisation falling), a genuine negative variant vs. Street, and momentum no longer strongly positive. Absent all three, shorting a 29%-short-interest, top-decile-momentum, net-cash company with $33.6bn of investment-grade backlog is the trap, not the trade.