Nebius Group [NBIS] — Criteria & Two-Horizon Valuation

Phase Space Research · 2026-07-29 · Framework: Criteria (references/criteria.md, 2026-07-29), superseding the gate framework v1.5.1 · Archetype: INFLECTION

This document supersedes the Gate block and the Gate 4 expected-return arithmetic in Nebius_Trade_Construction_2026-07-27.md and Nebius_Valuation_Analysis_2026-07-27.md. The research, capacity and mention-frequency work in those files stands unchanged and is not restated here.

There is no position verdict in this document. The memo scores every Criteria and blocks on none of them. Whether this analysis justifies a position is a question about a particular book, and two books answer it differently. The previous manifest carried investment_decision: "Watchlist"; that field has been retired.

0. What moved, and why

1. Data hygiene — three corrections, stated rather than silently applied

1.1 Share count — the scanner bug is fixed and the count is confirmed. The coverage scanner previously read XBRL tags at the wrong nesting level (F.get(tag) instead of F.get("us-gaap", {}).get(tag)), so its fallback never fired and NBIS — a live position — scored INDETERMINATE "no share count" despite publishing one. The bug is fixed in the current coverage_scan.py, and the 2026-07-28 rescan returns a share count.

The share count used in this document is 253,016,971 — us-gaap:CommonStockSharesOutstanding, as of 2025-12-31, from the FY2025 Form 20-F. It is a point-in-time outstanding count, which is the correct base for an enterprise value.

Alternatives from the same filing, and their effect on EV:

The if-converted count matters and is not used in the headline: ~50.7m shares of convertible dilution re-engage on a rally, which is a real cap on upside and is recorded under Downside Criteria rather than buried in the share count.

1.2 Revenue — the automated TTM is 209 days stale and understates by 40%. NBIS is a foreign private issuer. It files a 20-F annually and 6-K interim reports; it publishes no quarterly XBRL, so coverage_scan.ttm_revenue() fell back to the last fiscal year and returned $529.8m as of 2025-12-31 — 209 days old. Per valuation.md ("Assert recency"), that is reported rather than used.

Revenue used: $878m TTM through Q1'26, sourced to Form 6-K filed 2026-05-13 via Nebius_Valuation_Analysis_2026-07-27.md. The stale figure would have overstated EV/Sales at 81.4x against the correct 49.1x.

1.3 Net debt. $198m at Q1'26 (debt $8,450m + operating leases $1,046m − cash $9,298m), from the existing valuation document. The FY2025 20-F XBRL implies −$425m but predates the Q1'26 $6.3bn raise ($4.3bn convertible notes + $2.0bn NVIDIA equity), so the XBRL figure is not used.

2. The Criteria

2.1 Quality Criteria — INFLECTION standard

The archetype is declared first, because the standard differs and conflating them was calibration item B23. NBIS is INFLECTION: high gross margin, deeply negative operating margin, margin expanding fast, growth very high. Value sits almost entirely in the terminal period.

The gross-margin limb is the framework's own named defect case. criteria.md and valuation.md both cite NBIS by name: "NBIS and CAI publish no GrossProfit tag, so a quality check written as gross_margin > 0.50 evaluated nan > 0.50 and returned FAIL — silently rejecting two names for absent data." The 2026-07-28 rescan reproduces the correct behaviour: it returns INDETERMINATE with the note quality inputs missing: gross_margin, not FAIL.

INDETERMINATE is a statement about the tag, not about the business. The figure exists in the primary filing — 74% — and resolving it there turns the limb to PASS. That is the correct escalation path: a missing tag sends the question to a human, it does not reject the name.

F-score and gross-profitability LEVELS are demoted to context here, per criteria.md: they were derived on mature profitable firms and score a pre-inflection company worst exactly when the opportunity is largest. The existing valuation document's factor scorecard already reached this conclusion independently — it marked gross profitability (GP/A 0.029) "NOT MEANINGFUL" and marked down asset growth from "worst reading on the long board". Those judgements stand and are now the framework's default rather than an exception.

2.2 Liquidity Criteria

Unchanged from Nebius_Trade_Construction_2026-07-27.md. Equity liquidity is ample. Any options structure requires the actual chain pulled first — open interest and quoted size at the specific strikes and expiry — per the HCA precedent where the maximum open interest across an entire chain was 18 contracts.

3. Valuation Criteria — the two mandatory outputs

Terminal value is 167% of EV in the existing base-case DCF (the interim period is cash-flow negative), far above the 60% threshold, so the reverse DCF is mandatory as the primary long-horizon output and the forward DCF becomes supporting evidence.

3.1 The implied-path test

Solved with assets/reverse_dcf.py.

What the price requires: a 143.8% revenue CAGR for five years, holding the terminal margin at 20.4%, the exit multiple at 4.5x, WACC at 10% and the share count and net debt at the figures above.

That would take revenue from $878m to roughly $75.6bn by 2031.

The margin (demonstrated − required) cannot be computed for this name, and that is the finding.

NBIS's realised revenue path is $4,794m (2021) → $13.5m (2022) → $9.8m (2023) → $91.5m (2024) → $529.8m (2025). The company was reconstituted after divesting its Russian assets; the 2022 figure is a post-divestment stub. The mechanical 3-year CAGR of 239.8% is a base artifact, not a demonstrated capability, and subtracting the required path from it would produce a spuriously comfortable +96pp margin that means nothing. Demonstrated CAGR is therefore reported as NOT ESTIMABLE, and the Valuation Criteria result rests on §3.2 instead.

3.2 The exit multiple is UNIDENTIFIED — and this is the binding finding

valuation.md: "An exit multiple may only be drawn from a comparator set whose growth brackets the subject's growth at the exit year. If no such comparator exists, the multiple is UNIDENTIFIED and must be declared so rather than defaulted to a peer median."

A 1,433-name universe scan was run to supply that comparator set. Screening on NBIS's mechanical 239.8% growth returns exactly five names — the bare minimum — and they are these:

Four of the five are sub-$100m shells. The median of 4.5x is set by ABTS at $3m of market capitalisation and ANTA at $75m. Using it to value a $43bn AI-infrastructure business is precisely the defect valuation.md exists to prevent — and it is worse than the original defect, because the selection variable (239.8% growth) is itself an artifact of a corporate reorganisation.

The exit multiple for NBIS is UNIDENTIFIED. No comparator set both brackets its growth and resembles its business. The 4.5x figure is reported for audit only and is not used as an anchor.

Implied compression, stated as a number. NBIS is loss-making, so the comparison is made on sales: current EV/Sales 49.1x against the nominal exit multiple restated on sales (4.5x × 20.4% = 0.92x) — a −98.1% compression. That number is not credible; it is the arithmetic consequence of an unidentified multiple, and it is shown to make the identification failure visible rather than to hide it inside a point estimate.

3.3 Sensitivity — over the exit multiple, never over scenario probabilities

This replaces the retired instrument directly. The old Gate 4 ran its range across p_bear (0.20 → 0.40) while point-estimating the multiple — the exact failure valuation.md names on NTRA, where the memo conceded the flip point "has to be sought on the exit multiple, which is where the real judgement lives."

The required path never falls below ~67%/yr for five years anywhere in a defensible multiple range. That conclusion is robust to the identification failure, which is why the Valuation Criteria can still return a result: whatever the right multiple is, the required path is extreme.

Valuation Criteria: FAIL. The price requires materially more than the business has demonstrated, and the supporting argument — 3.5 GW of contracted power converting to revenue — is quantified in the existing research but is, on the memo's own analysis, already in the price: "NBIS must connect and monetise 3.26 GW by 2031 — 93% of everything it has contracted — merely to be worth $174.88." That is context, not a variant.

This is a FAIL, not a PASS WITH ARGUMENT, because the argument does not clear the bar criteria.md sets: a named product cycle, mix shift or pricing action that is evidenced and not already discounted.

3.4 The 12-month target

Built per valuation.md — near-term consensus, named product-cycle events, and a multiple anchored on the name's own history.

The own-multiple anchor is UNIDENTIFIED for NBIS, and this is declared rather than defaulted. Its trailing EV/Sales series over the last 442 sessions ranges from 0.5x to 360.8x with a median of 111.4x — but that series is measuring the revenue base, which moved 54-fold across the window (as-known revenue stepped $9.8m → $91.5m → $529.8m), not the multiple. Snapping to that median would produce a target of $3,320, which is nonsense, and reporting it would be the same category of error as the 4.5x exit multiple.

The one anchor that is documented and identified is the forward multiple against a genuine comparable — CoreWeave — which the existing valuation document builds in §5. On consensus FY2027E revenue of $10,500m:

12-month target: $118 – $227, with no defensible point estimate inside it. NBIS trades at 4.11x consensus FY2027E revenue, a 43% premium to CoreWeave's 2.87x. The target band is the multiple band; the midpoint is spot.

NTM revenue for reference: $7,542m — 5/12 × FY2026E consensus $3,400m + 7/12 × FY2027E consensus $10,500m, both from the existing comps table. EV/NTM revenue today: 5.72x.

Sanity band against the external professional. Street targets span $120 – $410, average $258 across 18 analysts. The house band's top ($227) sits 12% below the Street average. The divergence is explained explicitly: the Street is capitalising the contracted book at a higher forward multiple than CoreWeave receives, and this document declines to assert which is right — it reports that the entire disagreement is about the multiple and none of it is about the revenue.

Note the direction. Item B16 recorded 16 of 16 house targets below spot, a median 46.1% below Street. This target band straddles spot, and its upper bound is +34%. That is the behaviour valuation.md says to expect once the 5-year DCF-derived target is retired.

4. Downside Criteria — MEASURED, and it does not reject the name

Type: MEASURED. criteria.md is explicit about why: the principal declined to set a cap — "I don't know if I trust the bear case analyses… I saw too many examples of downside cases without much backbone or rigorousness being used to exclude good investments." A binding test resting on an untrusted estimate reproduces the defect where sizing maximised over a known-biased number. It therefore constrains nothing.

The most interesting fact about this bear case is that nobody is contesting it. Across six earnings calls, no participant on either side — management or sell-side — has said "useful life", "re-lease" or "interconnection" once. The residual-value question that dominates the downside is not being argued; it is not being asked.

Under the previous framework this −94.9% tail acted as a de facto veto through the E[R] arithmetic, because a near-total loss at 27% probability dominates any probability-weighted return. It no longer does. It is logged here with its cause and its probability, it is scored by ledger_scorer.py against realised drawdowns, and it earns promotion to BINDING only if it proves calibrated at ~10 and ~30 resolved records. The active protection in the interim is inverse-volatility sizing, which sizes a 109.5%-volatility name down automatically.

5. Momentum Criteria — MEASURED, entry timing only

Momentum governs when to enter a position the thesis already justifies. It never governs whether to own one. Any language in the superseded documents where momentum blocks, vetoes or conditions the name is deleted.

criteria.md names this as the type-discipline test case: "It was a blocking gate, was demoted to timing, and is the change most likely to be silently reversed under pressure."

The two horizons disagree, and that is the timing signal, not a conflict. 12-1 momentum is top-percentile; RSI-14 at 35.7 says the last fortnight has been weak. The reference long/short book's longs cluster oversold on RSI (EW 34.6, ISRG 40.4) while its shorts cluster overbought — short-horizon mean reversion running opposite to 12-month momentum, at 1/20th the horizon. Both are timing inputs and they are not in conflict: 12-1 says the name is in favour, RSI says today is a reasonable day to buy one.

The superseded trade-construction document conditioned entry on a "$155 / 50-dma reclaim rule". That is retained as timing guidance and explicitly not as an admission test.

6. What this document deliberately does not say

• No Long, Short, Watchlist or Avoid. The previous manifest's investment_decision: "Watchlist" is retired.
• No 5-year DCF-derived price target. The instrument that produced 16 of 16 targets below spot is gone.
• No E[R] versus a cash hurdle. Replaced by the implied-path test above and by slot competition in the strategy layer.
• No variant-versus-consensus admission test. Path B is dissolved, not renamed.
• No claim that the exit multiple is known. It is UNIDENTIFIED and is labelled so.

7. Provenance

Not refreshed for this update: no new filings were pulled beyond those named, no new consensus was purchased, and the catalyst calendar is unchanged. Prices are as at the 2026-07-28 close.