Download the investment report (PDF, 6 pages) Download the five-pillar analysis (PDF, 4 pages)
Frequently asked questions
Is Nebius's $40 billion backlog real?
The committed backlog near $40 billion is disclosed by management and underpins asset-backed financing, but the contract terms are conditional. Microsoft's $17.4 billion is subject to deployment and availability with per-tranche termination rights, and Meta's $15 billion unsold-capacity portion is a backstop, not firm revenue. Treat the backlog as contracted optionality, not guaranteed cash.
What is the funding watch on Nebius?
Nebius guides $20-25 billion of 2026 capex against about $3.0-3.4 billion of guided revenue, roughly 6-8x sales. The model works only while customer prepayments, debt markets, and equity markets stay open together. The watch flips negative on a dilutive raise below $200, a major customer retrenchment, or backlog conversion stalling; it flips positive when 2027 capacity converts at guided margins with prepayments intact.
What is Nebius's five-pillar score?
Nebius scores 70/100, Positive with a funding watch. Sunrise and leadership pass; moat and fortress are partial; free cash flow fails for now (Q2 2026 true FCF was roughly negative $3.4 billion on $5.66 billion of capex). The FCF pillar has the clearest upgrade path if 2027 capacity converts at guided margins.
The verdict
Nebius went from a $551 million ARR curiosity in Q3 2025 to $3.0 billion of ARR in Q2 2026, signing the two largest AI infrastructure contracts ever disclosed by an independent neocloud along the way. Microsoft committed about $17.4 billion with an option to $19.4 billion (September 2025). Meta committed $12 billion of dedicated capacity with a path to $27 billion total (March 2026). Committed backlog sits near $40 billion with potential above $46 billion, roughly 13x trailing-twelve-month revenue. That ratio is the bull case in one number.
Pricing power is the second proof point, and it is unusual. On October 1, 2026, Nebius raised on-demand GPU prices 17-21%, its second increase in three months, and it raised prices on H100s, a chip generation launched in 2022. Raising rental rates on four-year-old silicon means demand is outrunning supply across every generation, not just Blackwell. Management said on the Q2 2026 call (August 12, 2026) that the company was effectively sold out as quickly as new capacity came online.
The funding model is the third proof point and the structural edge over the pure debt-funded neocloud. About 70% of Q2 2026 deals included customer prepayments covering 50-60% of the associated capex, with more than $9 billion of prepayments expected in 2026. Operating cash flow was +$2.25 billion in the quarter on the back of those advances. Customers are financing the build.
The bear case is the bill and the concentration. Guided 2026 capex of $20-25 billion is 6-8x guided revenue of $3.0-3.4 billion. Two customers dominate the $40 billion backlog. The Microsoft contract, for all its size, is subject to deployment and availability with per-tranche termination rights if Nebius misses delivery dates; the filing never says "take-or-pay." Short interest sits near 21% of float. At $249.87, the stock trades at about 21.5x 2026 EV/Sales on consensus estimates, which leaves no room for a growth stumble: the shares fell 30% in September 2026 on valuation and concentration fears before rebounding.
DalalBytes score: 70/100. Verdict: Positive, with a funding watch. FY2029 targets (DalalBytes estimates): bear $120 (-52%), base $460 (+84%), bull $800 (+220%).
Why it is winning, with evidence
Three facts carry the bull case. First, the backlog: $40 billion committed and $46 billion-plus potential against a $68.5 billion market cap (October 6, 2026). No other independent neocloud has two hyperscaler anchor contracts of this size. Second, the pricing power: two price hikes in three months (May 2026: +29% on-demand, +51% spot; October 1, 2026: +17-21% across H100, H200, B200, B300), plus a first capacity auction that cleared 15% above the prior highest Blackwell price. Third, the prepayment funding: customers covering the majority of the capex they consume, which is why operating cash flow printed +$2.25 billion in Q2 2026 while GAAP showed a loss.
The unit economics support the story at the megawatt level. The Q2 2026 shareholder letter discloses $20-25 million of annual contract value per MW on the four landmark deals (each above $1 billion total contract value), and a $40-50 million per MW opportunity on auction and short-term capacity, with the first such deal signed in Q3 2026. Against the 41% group and 50% AI-cloud adjusted EBITDA margins reported in Q2 2026, the megawatts pay back in low single-digit years at contracted rates.
The inference layer is the emerging second engine. Token Factory inference workloads more than tripled in Q2 2026, and the October 1, 2026 acquisition of Inferize targets GPU idle time directly, which is what you buy when utilization, not demand, is the constraint. If the software layer becomes the reason customers stay rather than the megawatts, the moat pillar upgrades.
Financials: six quarters of compounding
All figures reported by the company unless noted. Revenue has compounded 39-75% sequentially for six straight quarters with no down quarter:
Q3 2025: $146.1M (+355% YoY) · Q4 2025: $227.7M · Q1 2026: $399.0M (+684% YoY) · Q2 2026: $582.3M (+454% YoY, +46% QoQ). ARR exited June 2026 at $3.0B (+598% YoY, +56% QoQ). Adjusted EBITDA inflected from a $21M loss a year earlier to +$236.2M at a 41% margin in Q2 2026 (AI cloud alone: 50%). GAAP diluted loss was $0.68/share in Q2 2026 (consensus -$0.36, a miss), driven by depreciation on the GPU fleet and stock comp; adjusted net loss narrowed to -$33.2M from -$91.5M a year earlier.
Two margin definitions, stated plainly because both will be quoted. The company does not report gross profit; depreciation sits on its own line. Computed from reported revenue and cost of revenues in the 6-K exhibits, the reported-definition margin series is smooth and rising: 70.6% (Q3 2025), 69.9% (Q4 2025), 74.0% (Q1 2026), 77.1% (Q2 2026). Including depreciation, the "full" margin is far lower: 20.9% in Q1 2026 and 32.5% in Q2 2026. The bear case will quote the latter. Both are legitimate; the first measures operating leverage on delivered compute, the second measures the true cost including the silicon wearing out. A widely circulated "~21% Q1 margin" figure came from stuffing the $212.0M quarterly D&A charge into cost of goods; it was a methodology error, not a business event.
Cash and funding at June 30, 2026 (reported): $8.04B cash and equivalents (up from $3.68B at December 31, 2025), $5.98B of deferred revenue (contracted cash already collected), $8.50B of non-current debt. Since then: $5.75B of convertible notes closed August 24, 2026 (2030 notes, ~$313.46 conversion price), a $2.8B ATM executed at ~$224 average with 12.3M shares remaining at June 30, and a $775M asset-backed facility in July 2026. H1 2026 capex was $8.13B; Q2 alone was $5.66B. Next earnings: Q3 2026 on November 10, 2026.
The contracts, read honestly
Microsoft (Form 6-K, September 8, 2025): dedicated GPU capacity in tranches from the new Vineland, New Jersey data center over a five-year term, deploying during 2025 and 2026. Total contract value about $17.4 billion through 2031, explicitly "subject to the satisfaction of the deployment and availability" of the capacity, with an option for additional services taking it to about $19.4 billion. If Nebius misses agreed delivery dates past a grace period and cannot supply alternative capacity, Microsoft may terminate that GPU service tranche (per-tranche, not whole-agreement). Either party may terminate for material breach with a 60-day cure. The agreement carries service-level commitments and liquidated damages for late delivery, and Microsoft's obligations commence only once Nebius confirms it has secured the capex financing. The filing never uses the phrase "take-or-pay." It is a large conditional commitment, not guaranteed cash.
Meta (Form 6-K, signed March 13, filed March 16, 2026): an Infrastructure Service Agreement structured as a series of five-year orders totaling up to about $27 billion. Two parts: $12 billion of dedicated GPU cluster orders, deploying in tranches from early 2027 and described as one of the first large-scale deployments of NVIDIA's Vera Rubin platform; and up to $15 billion of "unsold capacity" under which Nebius first tries to sell the capacity to third-party AI cloud customers, and Meta is obligated to purchase whatever remains unsold for five years from each deployment. That second part is a backstop obligation, not a pure option and not firm take-or-pay: Meta pays only on capacity Nebius fails to sell elsewhere.
Composition and conversion: management describes about $40 billion of committed backlog "that we can borrow against" (Q2 2026 call), which is the borrowing base for the asset-backed program. About 36% of a $37.5 billion backlog is expected to convert to revenue over the 24 months ending June 2028, roughly $13.5 billion (Motley Fool, October 5, 2026). No renewal terms, repricing clauses, or expiration schedules beyond 2031 are disclosed anywhere; whether the $15B Meta backstop sits inside the $40B "committed" figure is not disclosed. Two customers are the large majority of the backlog. That concentration is the single largest risk in the file.
Competition: the larger rival, the structural threat, the second tier
CoreWeave (NASDAQ: CRWV) is the larger, leveraged rival: Q2 2026 revenue $2.58B (+112% YoY) vs Nebius $582M (+454%), backlog $104B vs ~$40B, 2026 revenue guide $12.4-13.2B vs $3.0-3.4B. But the balance sheets invert the story: CoreWeave carries $35B+ of total debt and lost $626M in Q2 on $640M of interest expense, while Nebius is roughly net-cash-neutral with prepayment funding. CoreWeave's market cap (~$48B, October 7, 2026) now trails Nebius ($68.5B) despite 5x the revenue: the market is voting for funding quality over scale. CoreWeave's edges: closer NVIDIA relationship, first-mover with AI labs, first Vera Rubin bring-up. Nebius's edges: owned data centers (above 75% of capacity), prepayment model, faster growth, cleaner balance sheet.
The hyperscalers (AWS, Azure, GCP) are simultaneously the biggest customers and the existential risk. They own the end-customer relationship, deeper balance sheets, and their own silicon programs. The bull evidence is that they cannot build fast enough and would rather rent: Microsoft's $17.4B outsourcing decision is the exhibit. The bear articulation (Jim Chanos): neoclouds are debt-fueled intermediaries vulnerable to being cut off when hyperscaler capacity catches up. The counter (William Blair, September 30, 2026): hyperscaler contracts let neoclouds convert sub-investment-grade credit into investment-grade financing, structurally lowering borrowing costs.
The second tier competes for the same power and GPU allocation: Lambda (private, reportedly eyeing a late-2026 IPO) undercuts on published price (B200 at $6.69/hr vs Nebius $8.50) and signed a $35B Anthropic deal; Crusoe holds 4.9 GW contracted with a 40 GW-plus pipeline; Nscale signed a $45B Anthropic deal; IREN is pivoting bitcoin-mining power to AI cloud. None match Nebius's combination of scale, backlog, and funding. Nebius wins on funding model, owned infrastructure, pricing power, and growth rate; it loses structurally on GPU allocation dependence (NVIDIA holds 9.3% and the spigot), balance-sheet depth vs hyperscalers, customer concentration, and no proprietary silicon or end-customer franchise.
Leadership and capital allocation: the Yandex operator
Arkady Volozh co-founded Yandex in 1997 and built it to a ~$30B peak valuation with ~$9B in annual earnings (2023), then executed the $5.4B Russia divestiture (closed July 15, 2024) and rebuilt the remnant into a neocloud in under two years. The bench is Yandex infrastructure veterans: CFO Dado Alonso, COO Nave Ophir, CTO Danila Shtan, Chief Infrastructure Officer Andrey Korolenko. This is a team that has built and operated hyperscale infrastructure before, the scarcest input in the neocloud land grab besides power.
The capital allocation record since the split is aggressive but disciplined in form: customer prepayments covering 50-60% of capex; asset-backed facilities against contracted cash flows; the $5.75B converts with a $313 conversion premium, raising duration without near-term dilution; the $2.8B ATM executed at ~$224 near the highs rather than into weakness; tuck-in acquisitions (Eigen AI, Tavily, ClarifAI, Inferize) aimed at utilization and inference margins rather than empire-building. The criticism writes itself too: the share count keeps rising, and every funding layer adds claims ahead of equity. Insider sales in early October 2026 (Volozh $8.85M for RSU tax withholding; COO Ophir $113.38M under a 10b5-1 plan) are optically unhelpful against 21% short interest, though the tax-related ones are mechanical.
Technology in brief
The fleet spans H100 (backbone), H200, B200, B300, and GB200-class systems, with Vera Rubin NVL72 units in validation as of August 2026. No total GPU count or generation mix is disclosed in any filing; connected megawatts are the proxy: 170 MW at end-2025, 800 MW-1 GW targeted for end-2026, contracted power target raised to 5 GW on the Q2 call, with more than 1 GW of new build planned for 2027. Footprint: Mantsala, Finland (75 MW); Kansas City, Missouri (5 MW initial, expandable to 40 MW); Vineland, New Jersey (Microsoft); a Pennsylvania AI factory campus securing up to 1.2 GW; Paris colocation; plus UK, Israel, and Middle East sites. Above 75% of capacity is owned rather than colocated.
Above the metal sit three software layers: Nebius AI Cloud ("Aether" 3.0, launched Q3 2025) with managed Kubernetes or Slurm, custom racks, InfiniBand to 3.2 Tbit/s per host, and enterprise compliance; Token Factory, a managed inference platform with 60+ pre-optimized open-weight models whose workloads more than tripled in Q2 2026; and the acquired optimization stack aimed at cold starts and idle capacity. The engineering thesis: raw GPU rental commoditizes, but a full stack converting weights to production inference at lower cost per token retains customers and lifts utilization, the binding constraint on returns. (A full infrastructure teardown is a separate forthcoming piece.)
Risks
- Customer concentration. Microsoft and Meta dominate the $40B+ backlog. A capex retrenchment by either, or a Vineland delivery failure triggering termination rights, hits revenue and the funding stack at once.
- GPU cycle risk. If AI compute demand disappoints, rental prices fall, utilization drops, and the $20-25B capex program becomes stranded fixed cost. Burry, Chanos, and Dalio have all warned on AI capex.
- Funding intensity. $20-25B of annual capex is 6-8x revenue. The model needs prepayments, debt markets, and equity markets to stay open simultaneously.
- Dilution. The ATM still has 12.3M shares available (June 30, 2026) and the converts convert near $313. Shareholders fund the growth they are buying.
- Contract conditionality. Neither landmark contract is unconditional: Microsoft's is subject to deployment and availability with per-tranche termination; Meta's $15B portion is a backstop on unsold capacity. The backlog is contracted optionality, not guaranteed cash.
- NVIDIA dependence. 9.3% owner, the sole GPU source that matters, and the allocation spigot. Alignment today, leverage tomorrow.
- Execution. Vineland tranches, 5 GW contracted, and 1 GW-plus of 2027 new build must all deliver on schedule.
- Geopolitical perception. Legally severed from Russia since July 2024, but the Yandex lineage still screens out some mandates and invites headline risk.
- Valuation. About 21.5x 2026 EV/Sales on consensus leaves no room for a growth stumble; September 2026 delivered a 30% drawdown on exactly this fear.
Valuation: three lenses
Lens 1: relative. At $249.87, Nebius trades at about 21.5x 2026 EV/Sales and 7.3x 2027 EV/Sales on TIKR consensus ($3.29B and $9.73B). CoreWeave trades near 6x forward sales with $35B+ of debt and slower growth; the Nebius premium is payment for 4x the growth rate and prepayment funding over pure debt. Hyperscalers trade 6-9x sales with far lower growth. The neocloud premium is the market paying for 100%-plus growth and compute scarcity.
Lens 2: the compression math. BNP Paribas (Outperform, $399 price target, September 24, 2026) sees about $22B of ARR exiting 2027. If that ARR converts anywhere near guided, the 2027 multiple compresses hard from here, which is the entire bull case in arithmetic: you are buying 2027-2028 revenue at a single-digit multiple while the market still prices 2026. The risk is symmetric: if the ARR does not convert, the 21.5x multiple has a long way down.
Lens 3: scenarios. FY2029 targets (DalalBytes estimates). Bear $120 (-52%): an AI capex air pocket, growth stalls, FY2029 revenue ~$12B at a distressed 4x EV/Sales with heavier dilution; the multiple compresses and the funding stack transmits the pain to equity. Base $460 (+84%): FY2029 revenue ~$22B (extends TIKR's $15.23B 2028 estimate at ~45%), 7x EV/Sales as scale compresses the multiple while the growth premium over CoreWeave holds, ~315M shares after ATM, converts, and SBC; the 2027 capacity converts at guided margins with prepayments intact. Bull $800 (+220%): the $22B ARR exiting 2027 converts, 5 GW delivers on schedule, FY2029 revenue ~$30B at a sustained 10x scarcity multiple, inference (Token Factory) lifts margins and the moat pillar upgrades. The bear case is a funding crunch; the base case is the backlog converting on schedule; the bull case is Nebius becoming the independent AI cloud the hyperscalers cannot do without.
The growth rate and the multiple are the story translated into numbers: mid-40s revenue growth into 2029 at a high-single-digit multiple is the base because the backlog and pricing power support it; the wide spread reflects a genuinely binary funding model.
What would change the verdict
- To the upside: 2027 capacity converting at guided margins with prepayment coverage intact; a third landmark customer diversifying concentration; Token Factory and inference lifting utilization enough to upgrade the moat; quarterly true free cash flow turning positive. Any two upgrade the verdict toward the mid-70s.
- To the downside: a dilutive equity raise below $200; a Microsoft or Meta retrenchment or delivery failure triggering termination rights; backlog conversion stalling (watch the 36%-over-24-months marker); GPU rental price deflation; or 2027 capex guidance rising without contracted revenue attached. Any one puts the funding watch on negative.
Sources
Nebius Group Forms 6-K with Q3 2025 (November 12, 2025), Q4 2025 (February 12, 2026), Q1 2026 (May 20, 2026), Q2 2026 (August 12, 2026) results exhibits; Microsoft commercial agreement 6-K (September 8, 2025); Meta Infrastructure Service Agreement 6-K (signed March 13, filed March 16, 2026); Q2 2026 earnings call transcript (August 12, 2026); NVIDIA Schedule 13G (July 20, 2026); SEC Form 4 filings (October 2026). Reporting and estimates: TIKR, BNP Paribas (September 24, 2026), William Blair (September 30, 2026), Motley Fool, SiliconANGLE, Reuters, Zacks, tradingview.com, stockanalysis.com, finnhub.io. Market data as of October 6-7, 2026. All targets and scores are DalalBytes research estimates.
Disclosure
This report is for educational purposes only and is not investment advice, a recommendation, or an offer to buy or sell any security. All targets and scores are DalalBytes research estimates based on public information available as of October 7, 2026. Investing involves risk, including loss of principal. The author may hold positions in securities mentioned. Readers should conduct their own due diligence and consult a qualified financial advisor before making investment decisions.
Research and opinion, not investment advice. Do your own due diligence before investing.
© 2026 DalalBytes Research · For educational purposes only. Not investment advice.