Initiated 2026-10-01 · Price $232.28 (as of 2026-10-01 close, stockanalysis) · Mkt cap ~$66.8bn (287.7mm shares incl. August exchange shares) · Information technology / AI GPU cloud · Model: verified

Rating: Sell — Conviction: Medium

Nebius is executing well: revenue up 454% year on year, prepayments covering half the build, less debt than CoreWeave. But it is the same capital-hungry business: GPUs wear out in about five years, so by the time growth slows replacement capex absorbs most of EBITDA. At ~$67bn the market is paying for something close to the Bull path, and the Base case sits well below the price on both valuation methods.

Conviction tests (3a-v-c): T1 pass (Base midpoint −62% at +1pp, −57% at −1pp; Sell both) · T2 fail (Gordon $0.00 is a clear Sell; the exit-multiple method alone, $188.83 or −19%, reads closer to Hold) · T3 pass (FY0 facts from EDGAR, the 20-F, the Q2 6-K and the company's closing release; one est. input, the FY2023–24 NWC placeholders, does not feed value) · T4 pass (no pending deal, ruling or single-contract event outside the cases)

Model value range vs price
Bear $0.00Base $94.41Bull $364.70Price $232.28

Business overview

Nebius rents NVIDIA GPU clusters, storage and a managed inference platform (Token Factory) to AI labs, hyperscalers and enterprises. It files as the former Yandex N.V. (renamed). AI cloud was $574.9mm of Q2 2026's $582.3mm revenue; the rest is small; other former units are discontinued operations or minority stakes (carried at $1.6bn). Revenue went from $9.8mm (FY2023, restated) to $529.8mm (FY2025). FY2026 guidance is $3.0–3.4bn revenue and $7–9bn exit ARR, against $3.0bn ARR at June. Earnings turn on three things: how fast the 5 GW of contracted power becomes billed capacity (>1 GW a year from 2027), the price per MW on new deals ($20–25mm ACV in Q2), and how much of the $20–25bn 2026 capex customers prepay.

Competition

The thesis's main driver is GPU-hour pricing on new and renewing contracts. CoreWeave is the scale leader: trailing revenue $7.59bn against Nebius's $1.36bn and IREN's $0.71bn (stockanalysis, 2026-10-01). Oracle (OCI) and the hyperscalers bid for the same large AI-lab contracts; Microsoft is both a customer and a builder of its own capacity. Miner conversions (IREN) add lower-cost capacity. Pricing is still firm today: a Q2 capacity auction for Blackwell cleared 15% above prior highs (Quartr summary, August 2026). Pressure would show first in Hopper renewal pricing, then in prepayment terms. Oracle, the only comparable with mature earnings, trades at 15.9x lease-inclusive trailing EV/EBITDA; CoreWeave trades at 25.0x.

Bull case

  1. Capacity converts on schedule — the 5 GW of contracted power and $40bn of customer commitments become revenue as >1 GW a year comes online, with ARR at the top of the $7–9bn guide. Plays out if YE2026 ARR clears $8bn and 2027 deployment beats 1 GW. Model: rev_growth
  2. Pricing and operating leverage — AI-cloud adjusted EBITDA margin is already ~50%; Blackwell pricing at new highs and SBC shrinking relative to revenue lift GAAP margin into the mid-50s. Plays out if top-bin GPU supply stays tight through 2027. Model: ebitda_margin
  3. Prepaid contract paper earns a lower risk premium — prepayments covering 50–60% of capex and a 22-month payback justify a lower discount rate and a higher terminal multiple than a spot GPU renter. Plays out if new mega-deals keep carrying prepayments. Model: wacc, exit_ev_ebitda

Bear case

  1. Capex never stops — on a five-year GPU life, a ~$90–100bn fleet needs ~$18–20bn a year just to stand still, roughly D&A, so EBITDA overstates owner cash flow. Plays out if each GPU generation needs a full rebuy. Model: capex_pct_rev, exit_ev_ebitda
  2. Price compression — CoreWeave, Oracle, hyperscaler in-house capacity and miner conversions cap GPU-hour pricing, and Hopper capacity reprices at roll-off. Plays out if GPU supply loosens in 2027–28. Model: ebitda_margin, exit_ev_ebitda
  3. Commitments don't refill — $40bn of commitments is a fraction of what the 2027–29 build needs to earn back, and several customers are venture-funded labs. Plays out if new-deal signings slow or shrink after Microsoft and Meta. Model: rev_growth
  4. Funding dependence — $20bn+ a year of capex relies on converts ($4.3bn in H1 2026, $5.75bn in August), asset-backed debt and equity sales. Plays out if convert and ABS markets close. Model: wacc

Valuation & balance sheet

Metric (definition) Current Own history (range or 5y avg) Peers Source, as-of
EV/EBITDA (model basis: mkt cap − cash + debt + timing/investment adj.; GAAP EBITDA; leases excluded) n/m FY2025 (negative) · ~98x Base FY2026 · ~15x Base FY2027 n/a (business reset 2024) n/a on this basis Model; 6-K Q2 2026
EV/EBITDA (aggregator, lease-inclusive, trailing) 253.6x n/a ORCL 15.9x · CRWV 25.0x · IREN 463x (IREN distorted by early-ramp EBITDA) stockanalysis 2026-10-01
EV/Sales (aggregator, trailing) 48.3x n/a ORCL 7.6x · CRWV 12.5x · IREN 25.1x stockanalysis 2026-10-01
FCF (CFO − cash capex, H1 2026) −$3.6bn ($4.50bn − $8.13bn) n/a n/a 6-K Q2 2026

Model-implied value range (from model-summary.json; generic module, Gordon-growth DCF and exit-EV/EBITDA DCF, midpoints): Bear $0.00 · Base $94.41 · Bull $364.70 per share, i.e. implied returns of −100.0% / −59.4% / +57.0% vs $232.28. These ranges show how the bull and bear drivers translate into value; they are not price targets. The price sits about 51% of the way from Base to Bull, so the market is paying for roughly half the Bull path (~$50bn revenue by 2030, mid-50s margins). The exit method ($188.83) embeds 9x Y5 EBITDA, a ~40% de-rating from ~15x Base FY2027. Gordon ($0.00) embeds far less, because Base Y5 unlevered FCF is still negative once replacement capex runs at ~45% of revenue; the >10% gap between methods is itself a re-rating in the exit method. Both methods floor at $0.00 in Bear, a ~$15bn revenue plateau where FCF never turns positive, so a tail sensitivity is moot.

Balance sheet: roughly net-cash at the FY0 date (cash $13.8bn vs debt $13.5bn after the August converts), so net leverage is not meaningful today; the Base path runs 10.2x falling to 2.6x net debt/EBITDA as the build consumes cash. Coverage on FY2025 EBITDA is negative; Base path 7–13x. Liquidity: $8.04bn cash at 6/30 plus $5.75bn gross from August, plus a $775mm asset-backed facility (drawdown undisclosed). Nearest material maturity: converts due 2029–2032 (~$7.7bn carrying value after $800mm was exchanged for ~15.8mm shares), then $3.45bn 0.50% 2030s and $2.3bn 4.50% 2034s. Ratings: none.

Model note: verified (LibreOffice matched all 3,207 formula cells). Unverified inputs: historicals.nwc (FY2023–24 placeholders; do not feed projections). Assumptions without basis: none. No scenario CHECKs. Cash and debt are 6/30/26 plus the August converts (gross) and note exchange, so Y1 capex is H2 only (~$14.4bn). H1 GAAP EBITDA ($167.8mm) and the H1 NWC inflow ($3,990mm) are deducted in other_claims; non-marketable equity stakes ($1.62bn) are credited there at carrying value. NWC includes non-current deferred revenue and, unavoidably, capex payables (AP and accrued are combined). Revenue after 2026 is tied to capex at ~0.4x turns, a judgment from the Q2 deal economics. Older-note coupons are estimates. Beta 1.8 is a judgment (aggregator 1.46 spans the Yandex-era suspension). No EPS guidance; Base EPS stays negative through FY2030.

Scenario stress test

Reasoned from the bull/bear drivers above. The model column comes from the scenario overlays (Base case + shock).

Scenario Effect Mechanism Magnitude Model Δ value vs Base ($/sh)
S1 Fast equity crash − Volatile AI name, ~17% short interest: the multiple de-risks; contracts unaffected over weeks High −$38.20
S2 Slow bear / recession − Prepaid contracts hold, but new deals and the 2027 build stall, spot pricing falls, funding costs rise; transitory, so milder than the permanent Bear plateau High −$62.10
S3 Rapid rate shock − Back-end-loaded value and a converts/ABS-funded build; low coupons limit the interest channel Med −$25.43
S4 Slow rate grind − Same channel, grinding Med −$13.09
S5 Soft-landing cuts + Lower discount rate and cheaper funding; growth intact Med +$21.89
S6 Recession-driven cuts − AI-spend pullback and customer stress outweigh lower rates High −$42.07
S7a Credit liquidity shock − $20bn+/yr capex depends on open convert and ABS markets; ~$13.8bn cash softens it High −$34.83
S7b Slow default cycle − Venture-funded AI-lab counterparties plus dearer refinancing High −$41.79
S8 Stagflation − Power, lease and labor inflation against fixed-price contracts, plus higher rates Med −$21.55
S9a Dollar spike + USD-priced contracts, euro-cost sites (Finland, France, Iceland, UK): costs fall in dollars Low +$0.10
S9b Dollar slide − Mirror of S9a Low −$0.10
S10 Melt-up + Momentum AI-infrastructure name: melt-up and short covering re-rate it High +$45.84
S11 Energy supply shock − European power exposure, more than US-only peers; mostly a risk-off multiple effect Low −$7.94
S12 Mega-cap/AI derating − The core risk: an AI-capex unwind stalls new contracts, compresses GPU-hour pricing and de-rates the group High −$94.41

S12 equals the full Base − Bear gap. Currently active/on watch per the playbook: S3 partially active (the 2026-09-30 macro log shows the pace legs crossed too); S8, S10 and S11 on watch. S3 is live here mainly through the discount rate. S10 is the main way the price holds above Base.

Model value change vs Base, by scenario
S12 Mega-cap/AI derating−$94.41S2 Slow bear / recession−$62.10S6 Recession-driven cuts−$42.07S7b Slow default cycle−$41.79S1 Fast equity crash−$38.20S7a Credit liquidity shock−$34.83S3 Rapid rate shock−$25.43S8 Stagflation−$21.55S4 Slow rate grind−$13.09S11 Energy supply shock−$7.94S9b Dollar slide−$0.10S9a Dollar spike+$0.10S5 Soft-landing cuts+$21.89S10 Melt-up+$45.84

What would change the call

Upgrades if: YE2026 ARR reaches the $9bn top of guidance and 2027 guidance shows >1.5 GW of deployment with prepayments still covering ≥50% of capex, and evidence that replacement capex on the 2024–25 Hopper fleet runs well below D&A (e.g. older GPUs re-contracted rather than rebought). Downgrades if: n/a — already Sell. The call would firm if FY2026 revenue misses the $3.0bn low end, new-deal ACV per MW falls, or prepayment coverage drops.

Watch items

Sources