Nebius Group N.V. (NASDAQ: NBIS), initiation of coverage, Nov 13, 2026
Nebius rents scarce AI computers at peak prices that we expect to fade
Nebius builds liquid-cooled halls of NVIDIA chips and rents them to Microsoft, Meta and AI labs. The $128.64 price needs $17.05m of yearly revenue a megawatt for good; we expect $15.0m once supply catches up.
- Our rating
- Sell
- 12-month target
- $96.91
- Price, Nov 12, 2026
- $128.64
- Expected return, with dividend
- −25%
By Penmark Team
What Nebius is now
A fast-growing AI cloud that rents out NVIDIA chips by the megawatt
Nebius buys NVIDIA GPUs, wires them into liquid-cooled clusters and rents out the computing power. AI Cloud brought in $973.8m of 3Q26’s $980.8m revenue, though about $325m was one-off sales.
Customers paid upfront on over 90% of 3Q26 deals, and signed contracts reached $38.7bn. Quarterly sales grew 571% from a year earlier.
Consolidated revenue, continuing operations. Source: company filings, Penmark Research estimates.
Source: company filings, Penmark Research estimates.
What you are really buying
At $128.64, the price needs today’s scarce rental rates to last
Our discounted cash flow value is $45.45 a share. The price needs $17.05m of revenue a megawatt from 2032 to 2040, 13.7% above our $15.0m.
Three disagreements explain $51.96 of the $83.19 gap: rental yield, chip spending and build pace. The other $31.23 is a residual we put down to discount rates and single-year multiples.
Base value is our 15-year discounted cash flow (DCF). Each step moves one input to the market’s view. Source: Penmark Research estimates.
Where we disagree: rental rates
Scarcity prices above $25m a megawatt fade to $15.0m as supply arrives
New 3Q26 deals priced above $25m a megawatt a year, and GPU-hour prices rose over 20% in the quarter. Nebius’s CEO expects the shortage to ease in one to two years, and 200+ rival GPU clouds are building.
We expect $16.0m in 2027, a $17.5m peak in 2028 and $15.0m from 2032. At the market’s $15.75m, value would be $75.97 a share.
Yield is annual revenue per billed IT megawatt; 2026 is the 4Q26 run rate. Source: company filings, Penmark Research estimates.
Where we disagree: spending
$88bn of building over three years keeps cash flow negative until 2030
We expect capex of $27.31bn in 2026, $31.35bn in 2027 and $29.57bn in 2028; management guides 2027 to $40–45bn. Each megawatt of chips costs $30–32m, plus $12m for owned halls.
Free cash flow before customer advances is −$30.59bn in 2026 and −$26.64bn in 2027, and turns positive only in 2030, at $3.07bn. Advances save interest, not the cost of the build.
Free cash flow is operating cash flow less capex, excluding customer prepayments. Source: company filings, Penmark Research estimates.
Where we disagree: build pace
Building faster adds little value when new capacity just earns its cost
Management wants more than 1 GW a year, but colocation delays leave us at 720 MW of connected power at the end of 2026, then 1,000 MW a year.
At our yields new capacity earns about its 10.0% cost of capital. Adding 150 MW a year adds just $0.89 a share, and 300 MW either way moves value only between $43.58 and $47.22.
Values from our discounted cash flow model at base yields. Source: Penmark Research estimates.
The odds
Most of our value sits in a bull case we give a one-in-four chance
Our bad case, a GPU glut that cuts yield to $11m a megawatt, leaves nothing for shareholders. Our middle case is worth $45.45, 64.7% below the price.
Our good case, with software holding $17.0m a megawatt, is worth $265.54 and carries 76% of our weighted value. Weighting the three and adding 12 months gives our $96.91 target.
- Good, 25%880 MW at the end of 2026 and 1,300 MW more a year, yields of $17.5m to $19.5m a megawatt and $17.0m for the long run.
- Middle, 45%720 MW at the end of 2026 and 1,000 MW more a year, with yields of $16.0m in 2027 fading to $15.0m from 2032.
- Bad, 30%A GPU glut from 2028 cuts yields to $11m a megawatt and building stops in 2029, leaving no value after debt.
Circle size shows how likely we think each case is. The target weights all three and adds 12 months at the cost of equity. Source: Penmark Research.
What would change our mind
Lasting software-led pricing would prove us wrong
Our conviction is low. A Hold needs a 12-month target of $115.78, which a long-run yield of about $15.9m a megawatt would give.
The next tests are hyperscaler spending plans in January, the 4Q26 results in February and the $3.5bn bridge loan due in October 2027.
A Microsoft pullback from GPU clouds would cost $3.97 a share; more outsourcing would add $2.98.
We expect 4Q26 revenue of $1,169.3m and 720 MW connected, against a goal of more than $7bn of annual run-rate revenue.
Delayed colocation sites coming online would add $5.97 a share; another slip would cost $8.96.
Debt costing 150bp more would cut $10.45 a share.
Inside the full report
What each page covers after the summary on pages 1 and 2. Each opens the full report at that page.
See all 26 pages
- 3Margins reach 57.4% by FY28E, yet ROCE reaches only 7.4% on $76bn PP&E; we sit +3.2% above FY27E revenue consensus
- 4At 7.7x FY27E EV/EBITDA, NBIS trades at a 76% premium to CoreWeave (4.4x) despite lower margin and an unproven GW build
- 5Six exhibits link Nebius’s ARR-driven rally to $88.23bn of capex, fading yields and a $96.91 target
- 6Three operating debates explain $51.96/sh of the $83.19/sh gap between our $45.45 base DCF and $128.64
- 7Q3 advances topped 100% of deal capex, yet Q3 operating cash flow was −$273m
- 8AI Cloud generates 99.3% of Q3 revenue ($973.8m), but D&A ($826.4m) keeps 9M operating income negative
- 9Every 100 MW of connected power yields 73.6 MW of billed IT load and $1.18bn of FY27E revenue
- 10Steady-state 1 MW leaves $3.62m pre-tax profit after $5.00m GPU replacement and $0.53m shell depreciation
- 11Specialist neocloud revenue is estimated to scale from ~$48bn to ~$300bn by 2030E, while silicon captures upstream rent
- 12Nebius out-engineers smaller neoclouds (813 R&D staff), but trails CoreWeave on scale and hyperscalers on funding cost
- 13Dual-class Class B shares give CEO Volozh 51.6% voting control on a 9.8% economic stake
- 14Deal pricing, capex funding and energized MW move NBIS (+34% Q2 / −50% since Aug) — not GAAP EPS
- 15Post-2029 GPU rental yields fade from >$25m/MW scarcity peaks to $15.0m/MW, cutting $30.52/sh vs market views
- 16$88.23bn of FY26E–28E capex keeps core FCF negative until FY30E; advances save interest, not principal
- 17Adding ±300 MW/yr shifts base DCF by just +$1.77 / −$1.87/sh ($43.58–$47.22) unless yields stay elevated
- 189M26 GAAP net income (+$393.3m) and OCF (+$4.23bn) reflect ClickHouse gains and advances, masking −$19.03bn core FCF
- 19Revenue reaches $12.63bn in FY27E (+3.2% vs consensus) and $54.81bn in FY32E on $31.35bn FY27E capex
- 2015-year unlevered DCF yields $45.45/sh base value; EV/FY27E EBITDA cross-check gives $68.68/sh
- 21Justifying $128.64 requires $17.05m/MW terminal yield (+13.7%) or an 8.03% WACC (0.93x beta)
- 22Scenario weighting ($86.84/sh today) rolls at 11.60% Ke to our $96.91 target (−24.7%, Sell)
- 232027 signposts skew negative (−$3.65/sh net); software-led $17m/MW yield ($266 bull) is our main upside risk
- 24~51% 13F institutional ownership, 50.9m shares short and a $276.20 pre-print Street mean
- 25Annual model: revenue scales from $529.8m (FY25A) to $54.81bn (FY32E) and $76.94bn (FY40E)
- 26Core FCF ex-advances troughs at −$30.59bn (FY26E) and turns positive at +$3.07bn in FY30E
- 274Q26E reaches $1,169.3m sales (+413% YoY); net debt/EBITDA falls from 13.3x (FY26E) to 5.6x (FY27E)
- 28Reference: $826.4m 9M26 D&A and $196.5m SBC turn $831.5m Adj. EBITDA into a $253.5m EBIT loss
Penmark Research. This summary draws on our initiation report on Nebius Group N.V., published Nov 13, 2026. It is not investment advice. Analyst certification, rating definitions and important disclosures are in the full report.