Nebius Group N.V. (NASDAQ: NBIS) — Everything They Build, They Sell — and Dilute

Independent fundamental equity research Report date: 2026-06-10 Security: Nebius Group N.V., Class A ordinary shares (NASDAQ: NBIS); foreign private issuer (Netherlands); files Form 20-F (annual) and 6-K (interim) in USD. Price context (2026-06-10): ~$220.12 · Market cap ~$55.9B · Enterprise value ~$56.6B (net cash) · ~220.4M shares · 52-wk range ~$35 → ~$279 · ~20% of float short.

⚡ Claude’s Take

This block is the author’s own independent, subjective opinion and general information only — not investment advice. Everything below it (the body of this article) is deliberately position-free and carries no recommendation or price target.

Verdict: HOLD / AVOID-chasing here (~$220). Accumulate-on-weakness only in roughly the $110–150 band (≈ 4–5× exit-2026 ARR, ≈ $35–45B EV). Not a short. Conviction: medium.

Tag: “Everything they build, they sell — and dilute.”

Nebius is a genuinely good operating story trading at a genuinely demanding price, and the gap between the two is the whole call. The demand is real and pre-sold — sold out every quarter, ARR up 50% sequentially to $1.9B, a $21B contracted backlog, NVIDIA/Meta/Microsoft all validating the platform, and a rare net-cash balance sheet among the neoclouds. That is why this is not a short despite an obvious bear case: ~20% of the float is already short, the backlog and net cash blunt the liquidity bear, and an NVIDIA-blessed, founder-led compounder with squeeze dynamics is a dangerous thing to be short of. But at ~$220 the market is underwriting the base-to-bull outcome with essentially no margin of safety: the business loses ~$600M+ at the GAAP operating line, full-year-2025 adjusted EBITDA was negative, the only reported “profit” is a non-cash mark on a private ClickHouse stake, and the company must fund $20–25B of 2026 capex against ~$9.3B of cash and ~$3B of revenue — a gap that will be closed with debt, customer prepayments, and equity that can dilute Class A holders 40–50%+ over the convert lives. You are paying a full price for a price-taking, commodity-leaning GPU-rental business at what looks like the euphoric phase of a capital cycle.

The framing is momentum-darling-meets-capital-cycle-peak: quality-of-growth is high, quality-of-earnings is poor, and the moat is a shared NVIDIA-supply-and-operational-cost edge without customer captivity — Greenwald’s weakest, least-durable advantage. I want this business, but I want it with a cushion — the $110–150 zone is where the dilution math and the cyclicality leave room to be wrong. Flips bullish if management funds the capex largely non-dilutively (prepayments + asset-backed financing against Meta/Microsoft credit) while core AI-cloud EBITDA margins hold ~45%+ as supply floods in. Flips bearish if a hyperscaler trims or cancels, GPU pricing rolls over for two consecutive quarters, or the ATM/equity becomes a primary funding source.

1. Executive Summary

Nebius Group is an “AI-native hyperscaler” — a neocloud that builds and rents GPU compute for AI training and inference. It is the reconstituted Western remnant of Yandex N.V.: after divesting all Russian assets in July 2024 (the largest corporate exit from Russia) for ~$5.2B in headline consideration, the company renamed itself, relisted on Nasdaq in October 2024, and pivoted essentially all of its capital and energy into AI infrastructure. The transformation has been violent and, operationally, spectacular. Nebius AI core revenue grew +841% year-over-year to $390M in Q1 2026; annualized run-rate revenue (ARR) reached $1.9B at end-March, up 50% sequentially; the company guides to $7–9B exit-2026 ARR and $3.0–3.4B of FY2026 group revenue; and it carries a ~$21.3B contracted backlog (RPO) anchored by Microsoft and Meta.

The investment question is not whether demand exists — it plainly does, and management says “everything we build, we sell.” The question is the quality of the business, the quality of the earnings, and the price. On business quality, Nebius has a real but shallow edge: privileged-but-shared NVIDIA supply (a $2B NVIDIA equity investment, Exemplar Cloud status, early Rubin/Vera access), genuine operational speed, and a vertically-integrated owned stack (>75% of power on owned sites). None of this constitutes a durable Greenwald franchise moat, because there is no customer captivity at the bare-metal layer where most revenue sits — raw GPU compute is fungible, and the better-capitalized hyperscalers can in-source. On earnings quality, the picture is poor: the continuing business lost ~$612M at the GAAP operating line in FY2025, full-year adjusted EBITDA was negative (−$65M), and the only reported net profit comes from non-cash fair-value marks on a private ClickHouse stake. The “operating cash flow” the company touts is manufactured by customer prepayments (a delivery liability), not earnings; FY2025 free cash flow was roughly −$3.66B.

The defining tension is the funding gap: $20–25B of 2026 capex against ~$9.3B of cash. Management intends to bridge it with prepayments, asset-backed financing collateralized on Meta/Microsoft contracts, corporate debt, and equity (a 25M-share ATM is authorized but unused). The dilution arithmetic — stacked convertibles, the NVIDIA warrant, ~16%-of-revenue stock-based comp, plus any ATM — is the spine of the bear case. The macro overlay is a classic Marathon capital cycle: record industry capex (~$725B of Big-Tech spend in 2026) is flooding supply into a price-taking rental layer, with power scarcity the only governor delaying mean reversion.

This memo takes no position and sets no price target (see the fenced Claude’s Take above for the one labeled exception). The body lays out the embedded expectations — what must be true to justify ~$56–57B of enterprise value — and the bear/base/bull scenarios that bracket it.

2. Business Overview

What Nebius does. Nebius operates a full-stack AI cloud: it procures NVIDIA GPUs, builds and operates data centers, and sells AI compute and related services. The Nebius AI cloud business is 98% of group revenue ($390M of $399M in Q1 2026). Management describes a layered platform spanning “from bare-metal to multi-tenancy to inference to agentic,” monetized through three mechanisms:

1. Bare-metal GPU clusters — large reserved clusters sold to hyperscalers and frontier labs (Microsoft, Meta). Management itself calls this “the most basic, not differentiated service.”
2. Managed multi-tenant cloud — GPU-hours sold both on-demand (pay-as-you-go) and as fixed reserved-capacity contracts. This was the original product and serves AI-native startups, software vendors, and enterprises.
3. Token Factory (launched November 2025) — managed inference monetized per-token, supporting open models (DeepSeek, Llama, GPT-OSS, Nemotron, Qwen). This is the fastest-growing layer, with agentic search (via the Tavily acquisition) the speculative next frontier.

Segments. Beyond the core, Nebius consolidates two immaterial-but-dilutive businesses: TripleTen (B2C edtech bootcamps; ~10% of revenue, tuition-funded) and Avride (autonomous vehicles and delivery robots; “limited contribution”). Together they are roughly $9M of quarterly revenue and a drag on group margin — the gap between Nebius AI’s 45% Q1’26 adjusted-EBITDA margin and the group’s 32% is essentially these two. The company also holds equity stakes in ClickHouse (a fast-growing analytics database; the source of the recent non-cash gains) and Toloka (data labeling, deconsolidated in 2025). Management has signaled an intent to find “strategic and financial partners” for Avride and TripleTen — i.e., to deconsolidate them as it did Toloka.

Revenue model and recurring-revenue framing. The “recurring revenue” label is partly overstated. Reserved-capacity contracts and Token Factory consumption are genuinely repeat/contracted, and contract durations are lengthening; but the bulk of the backlog is long-dated capacity-leasing to a handful of large counterparties, which behaves more like long-term infrastructure rental than like recurring SaaS. Crucially, management is explicit that software is not a separate revenue stream (“we look at the software as an enabler… not that we build software to generate a separate revenue stream”). The software widens the addressable customer base and improves unit economics through optimization — it is a cost/TAM play, not a high-margin software annuity.

Geography and capacity. Nebius is building globally: the UK (a £1.7B investment across four NVIDIA-powered sites), Finland (its anchor European site), and the US (Pennsylvania, a 1.2GW owned site; plus Missouri and Alabama for 2027). Contracted power exceeds 3.5GW, targeting ≥4GW by end-2026, with line-of-sight to 5GW by 2030 under the NVIDIA arrangement. Over 75% of contracted power is now on owned sites — a deliberate vertical-integration choice that improves economics but raises capital intensity.

Verdict. A capacity-leasing business with a software-optimization wrapper and two small non-core ventures. The economic engine is renting NVIDIA compute at scale; the “full-stack” story is real as a cost-and-reach advantage but does not convert the business into a software-margin or high-retention franchise. Calling it a “hyperscaler” flatters it — it is a fast-executing, well-funded challenger neocloud, one tier below the public leader (CoreWeave) and several tiers below the true hyperscalers in scale and self-funding capacity.

3. Industry Dynamics

Market structure and profit pools. AI compute is in a once-in-a-generation build-out. Combined Big-Tech 2026 capex guidance runs ~$635–725B (more than 2× 2024); CoreWeave alone guides $30–35B of 2026 capex (after $14.9B in 2025); Crusoe, IREN, Nscale, Lambda and others are all racing. The profit pool, however, is concentrated at NVIDIA, which sells the GPUs to everyone at ~75% gross margin, and only transiently at the neocloud rental layer while supply is short. The neoclouds are price-takers on both critical inputs — GPUs (NVIDIA controls allocation and cadence) and power (scarce, slow to build). The rental layer is the most commoditized link in the chain.

The real bottleneck is power, not GPUs. New power-connected data-center capacity takes 3–5 years; transformer, switchgear, and turbine shortages create 24–72-month delays. Microsoft has disclosed tens of billions of dollars of Azure demand it cannot fill for lack of power. This power scarcity is what is currently propping up GPU rental pricing and could extend the favorable phase of the cycle by two to four years. It is also why Nebius’s owned-power strategy (multi-GW contracted, >75% owned) is strategically sensible — securing power is a more durable advantage than securing GPUs.

Barriers to entry are capital/power barriers, not franchise barriers. Entry requires NVIDIA allocation, multi-gigawatt power contracts, land, and tens of billions of financing. These deter sub-scale entrants — but they do not deter the hyperscalers (AWS, Azure, GCP, Oracle OCI), who are vastly better capitalized and who in-source their own capacity. The structural threat to the neoclouds is not other neoclouds; it is their largest customers building their own.

Capital-cycle read (Marathon lens). This is a textbook mid-euphoria capital cycle: extraordinary returns (Nebius AI 45% adjusted-EBITDA in Q1’26) plus cheap, abundant capital are drawing a flood of supply. Marathon’s framework predicts mean reversion as supply catches demand — high returns attract capital, capital builds capacity, capacity compresses pricing and returns. The power constraint delays, but does not repeal, the overshoot. Morgan Stanley estimates >$400B of hyperscaler debt issuance to fund the build. When a sector is simultaneously the most loved, the most capital-hungry, and the most rapidly expanding supply base, forward returns on incremental capital are usually disappointing.

Historical analogues sharpen the warning. The closest precedents to a debt- and equity-funded land-grab in a commoditizing capacity layer are the 1999–2001 telecom/fiber build-out and the various crypto-mining capacity cycles. In both, demand was real and growing, the technology was transformative, and the early operators did earn extraordinary returns — and yet the capacity-owning middle layer was where the capital was destroyed, because supply overshot, pricing collapsed toward marginal cost, and the assets depreciated faster than they could be amortized against falling prices. Fiber owners went bankrupt while the internet they enabled flourished; the analogy is imperfect (GPUs depreciate and must be replaced, unlike dark fiber, which arguably makes the neocloud case worse on reinvestment) but the structural lesson holds: being right about the demand for a technology is not the same as earning a return on the capacity that serves it. The toll-takers and the application layer captured the durable value; the capacity renters were the cyclical casualties.

Where Nebius could escape the analogy. Three things would distinguish Nebius from the fiber casualties: (1) genuine, persistent power scarcity that caps supply for years; (2) a shift in revenue mix toward stickier inference/software that earns a franchise margin rather than a rental spread; and (3) contract structures (long-dated, prepaid, take-or-pay-like) that transfer cyclical risk to the customer. Nebius is pursuing all three, and the prepayment/owned-power strategy is genuinely better-positioned than a pure on-demand renter. But none is yet proven through a downturn.

Verdict — structurally mediocre-to-poor for the rental layer. High capital intensity, price-taking on the two key inputs, a commoditizing core service, and a supply wave that should compress ASPs. The industry is attractive only while power scarcity persists and demand outruns supply — which describes today, but is a cyclical, not structural, condition. The durable winners of this build-out are more likely to be the toll-takers (NVIDIA, power/grid, and possibly the application layer) than the capacity renters in the middle.

4. Competitive Position

Scale reality check. Nebius is a strong #2-tier player, not a leader. The public benchmark, CoreWeave (CRWV), is roughly 5× larger (Q1’26 revenue ~$2.08B vs Nebius $399M), with a ~$99.4B revenue backlog / ~$66.8B RPO versus Nebius’s $21.3B RPO. The hyperscalers are 10–100× larger and self-funding. Where Nebius distinguishes itself is the balance sheet — it carries net cash (~$9.3B cash, ~$9.2B convert principal) while CoreWeave carries ~$35B of debt and ~$2.1B of annual interest. Nebius is the better-capitalized challenger, but a challenger nonetheless.

Independent quality signal. SemiAnalysis’s ClusterMAX 2.0 rating places CoreWeave alone at Platinum and Nebius at Gold (alongside Oracle, Azure, Crusoe, Together AI). SemiAnalysis credits Nebius with the lowest absolute price and best short/medium-term terms among technically-competent clouds, plus <2-day cluster provisioning. Translation: Nebius competes on price and operational speed — a cost/operational edge, not a franchise.

The NVIDIA relationship — advantage or “backing everyone”? The $2B NVIDIA equity investment, multi-generation Exemplar Cloud status (latest on GB300 for training), early access to Rubin/Vera, and line-of-sight to 5GW are genuine and valuable — supply certainty is a real competitive advantage when GPUs are scarce. But NVIDIA also backs CoreWeave, Crusoe, Lambda, and Nscale. The advantage is shared and non-exclusive — a rentable supply edge, not a proprietary moat. It also carries a circular-financing taint (NVIDIA investing equity into a customer that buys NVIDIA GPUs).

Switching costs and captivity. This is the crux of the moat call. At the bare-metal/training layer — most of revenue — switching costs are low: GPU compute is fungible, and sophisticated AI customers routinely multi-source across 3–4 vendors. At the managed-cloud and Token Factory/inference layer, switching costs are moderate and rising (model deployment, fine-tuning, MLOps integration create stickiness), but this layer is small and unproven. Customer concentration — Meta and Microsoft dominate the backlog — is fragility, not captivity: these counterparties are themselves hyperscalers who can in-source and who hold pricing leverage. Management treats them as financing enablers (“this contract can unlock billions of capital at attractive rates”), not as a captive customer base.

Greenwald moat-type call. The candidate advantages map to: (1) a supply advantage (NVIDIA allocation) that is shared/non-exclusive; (2) a cost/efficiency advantage (vertical integration, owned power) that is real but replicable and does not lock customers in; (3) switching costs that are low at the dominant layer. The best fit is a supply/cost advantage without customer captivity — in Greenwald’s taxonomy, the weakest and least durable of the three genuine advantage types, and the one that does not survive scale competition from better-capitalized rivals. There is no economies-of-scale-plus-captivity (the only durable type), because there is no captivity.

Moat-to-financial-outcome test. If Nebius’s full-stack software vanished tomorrow, its bare-metal customers could rent equivalent GPUs from CoreWeave, Lambda, or a hyperscaler with limited disruption — their outcomes would not durably deteriorate. By this standard, a “moat” that fails that test is not a moat. The only place the claim ties to a defensible financial outcome is Token Factory/inference stickiness — which is small, early, and contested.

The vertical-integration economics — real, but a margin edge, not a moat. Management’s strongest structural argument is that owning the full stack (>75% of power on owned sites, in-house data-center design, in-house hardware-software optimization) yields a durable cost advantage: lower power and build cost per delivered GPU-hour, higher utilization, faster provisioning. This is credible and shows up in the SemiAnalysis “lowest absolute price” assessment and the 45% Q1’26 segment EBITDA margin. But a cost advantage in a commodity is a margin advantage, not a franchise — it lets Nebius earn a better spread for as long as it out-executes, and it can be competed away by any rival who matches the vertical integration (CoreWeave, the hyperscalers, and well-funded entrants are all building owned capacity). In Greenwald’s terms, a cost advantage without captivity erodes as competitors replicate the cost structure; it does not compound. It is worth paying something for — but not the durable-franchise multiple the bull case implies.

The inference-stickiness thesis is the only path to a real moat — and it is unproven. If Token Factory and the agentic layer (Tavily) embed customer workloads — models deployed, fine-tuned, integrated into production pipelines with MLOps tooling — switching costs rise and Nebius could earn franchise economics on a growing share of revenue. This is the bull’s best structural argument and management’s clear intent. The evidence is thin: inference is the fastest-growing layer but small; the agentic monetization model is “still to be defined”; and the hyperscalers (with Bedrock, Vertex, Azure AI) are racing to commoditize managed inference. The race is between Nebius building stickiness and the industry commoditizing the layer.

Verdict — no durable competitive advantage. Nebius has a genuine cost/supply edge plus operational excellence in a crowded, commoditizing market. The differentiation is real but shallow and concentrated in an unproven software layer. The single most important disconfirming question is whether Token Factory/inference can build genuine switching costs before hyperscalers commoditize inference and a supply normalization compresses ASPs.

5. Growth History and Forward Opportunities

The numbers are extraordinary. On a continuing-operations basis (continuing Nebius, ex-Russia), revenue went from roughly $9.8M (FY2023) → $91.5M (FY2024) → $529.8M (FY2025), and group revenue hit $399M in Q1 2026 alone (+684% YoY, +75% QoQ). The core Nebius AI segment grew +841% YoY in Q1’26. ARR reached $1.9B at end-March 2026 (from $1.25B a quarter earlier, +50% sequentially). Management guides to $7–9B exit-2026 ARR and $3.0–3.4B FY2026 group revenue. The sales pipeline grew 3.5× quarter-over-quarter in Q1’26 (a record, and notably excluding the Meta/hyperscaler strategic deals), and management says “4+ customers compete for every GPU we bring online.”

Quality of growth — real demand, but supply- and capital-gated. Management is explicit that of its three growth levers — utilization, pricing, capacity — only capacity is binding (“we sold out our capacity”; pricing and utilization are “not limiting”). Growth is therefore a direct function of how fast capital + power + GPU allocation convert into live megawatts. This is high-quality in the sense that it is pre-sold and demand-backed, but lower-quality in that it is capital-cycle-dependent and its returns hinge on pricing holding as supply floods in. It is growth bought with capital, not earned through a widening franchise.

The RPO→revenue gap is an execution bet. Reconciling a $21.3B backlog and a $7–9B exit-ARR target against ~$1.6B of FY2025 revenue depends on back-half-weighted 2026/2027 capacity landing on schedule. Capacity delivery is explicitly back-half-2026 weighted (a Q3 step-up), with Pennsylvania not lit until ~end-2027 and Missouri/Alabama in 1H’27. Media reports of delays at the Vineland, NJ (Microsoft) site — which management denies — illustrate the execution fragility. Any power, turbine, or construction slip directly defers the ARR ramp.

Forward opportunities. (1) Inference/Token Factory — the fastest-growing layer, with better margin optionality and the side benefit of extending GPU useful life (older chips cascade from training to inference). (2) Agentic — via Tavily; monetization model “still to be defined.” (3) Geographic expansion — UK, US, Finland. (4) Inference-talent M&A — Eigen AI (rated #1-speed inference by NVIDIA) and Clarifai.

Where growth disappoints. (a) Pricing — today’s gains are cyclical; 2027 supply (Nebius + CoreWeave’s $35B + hyperscalers’ ~$700B) likely compresses ASPs, hitting both revenue and the 40% margin target. (b) Capacity slippage — the single biggest near-term risk given back-loaded delivery. © Hyperscaler in-sourcing — the $15B Meta option assumes a strong resale market. (d) Depreciation/obsolescence — if real GPU economic life is shorter than the depreciation schedule, the reinvestment treadmill is harsher than the headline growth suggests.

Verdict — spectacular but lower-quality growth. Pre-sold and demand-backed, yet capital-cycle-dependent, execution-fragile on capacity timing, and resting on pricing and margins that are partly cyclical. The growth rate is not in doubt for 2026; its durability and economic return beyond the current cycle are.

6. Financial Quality

Multi-year financials (continuing operations, USD millions; reconciled to FY2025 20-F filed 2026-04-30 and SEC XBRL):

n/m = not meaningful (sub-scale); n/d = not separately disclosed at this granularity. Q1’26 figures from the earnings release / call (2026-05-13).

The headline GAAP profit is entirely non-operating. FY2025 net income of +$82.5M is not earnings. It is the sum of a +$598.9M non-cash ClickHouse fair-value mark (ASC 321, based on a single observable funding-round price — flagged by the auditor as a critical audit matter), ~$55M of money-market gains, ~$27M of FX, and ~$73M of discontinued-operations gains (the Toloka disposal). Strip those and FY2025 was a ~$665–670M pre-tax loss. The actual operating line lost −$611.7M on $529.8M of revenue (≈ −116% operating margin), and the operating loss has widened every year (−$285.7M → −$399.6M → −$611.7M). The Q1’26 net income of $621M is the same mechanism repeating — a ClickHouse Series D mark-up at a reported ~$15B valuation forced another non-cash step-up. None of it is cash the cloud business earned.

Adjusted EBITDA flatters the model by erasing the depreciation wall. The 20-F’s own reconciliation shows full-year-2025 group adjusted EBITDA was a −$64.9M LOSS — the company did not generate positive adjusted EBITDA for FY2025. The widely-cited “32% group / 45% Nebius AI” Q1’26 margins are a one-quarter inflection, not a run-rate. The single largest add-back is D&A of $417.9M (+442% YoY) — and for a business whose core asset (GPUs) depreciates over a handful of years and must be perpetually replaced, adding back all depreciation is precisely the adjustment that overstates economic profitability. Honest run-rate economics sit much closer to the GAAP operating loss than to “32% EBITDA margin.”

The depreciation wall is ahead, not behind. Of ~$6,190M of gross PP&E at end-2025, ~$2,417M (39%) is “assets not yet in use” and therefore not yet depreciating. As that capacity is energized and the $20–25B of 2026 capex lands, D&A will rise by multiples, and reported operating losses (absent offsetting revenue) will widen before they narrow. A quality-of-earnings watch item: management indicated a move toward a 5-year GPU depreciable life for 2026 (the FY2025 20-F policy note reflects a 4-year life). Lengthening the life mechanically flatters EBIT and net income; it should be watched and normalized. (Note: Nebius’s 4-year basis is at the conservative end of the peer set — CoreWeave uses 6 years — so on a like-for-like basis Nebius’s reported losses are not understated relative to the loudest peer; a move to 5 years would narrow that conservatism.)

Cash flow — “operating cash” is customer financing. FY2025 continuing-ops operating cash flow of +$401.9M was manufactured by working capital: deferred revenue (customer prepayments) contributed +$1,565.8M, partly offset by a −$714.6M AR build and ~−$441M of other-asset/VAT movements. Strip the prepayment and underlying operating cash flow was deeply negative (~−$1.0 to −1.2B). The same prepayment engine produced the heavily-marketed $2.3B of Q1’26 operating cash flow. After −$4,066M of capex, FY2025 free cash flow was roughly −$3.66B, even with the prepayment tailwind. This is a large, structural cash consumer — not a cash generator.

Balance sheet — net cash today, but optically so. Gross convertible debt of ~$9.2B principal roughly equals ~$9.3B of cash post-Q1’26, but a meaningful slice of that cash (~$2.3B+) is customer prepayments — a liability to deliver future compute, not free cash — and the entire balance is earmarked for a capex program that exceeds it 2–3× within twelve months. Stock-based compensation runs ~$83M (~16% of revenue), high and dilutive. There are also disclosed material weaknesses in internal control over fixed assets and over TripleTen revenue recognition — a fixed-asset control weakness in a company mid-way through a $6B-and-growing PP&E build is exactly where capitalization, useful-life, and impairment errors would hide.

A unit-economics / cluster-ROIC sketch — the question adjusted EBITDA dodges. Management has not disclosed clean cohort-level returns, so this is necessarily an estimate, but the arithmetic frames the debate. Take a notional GPU cluster: industry rules of thumb put all-in capex (GPUs, networking, data-center fit-out) at roughly $30–40k per high-end GPU, and a reserved cluster can gross perhaps $1.5–2.5/GPU-hour. At ~70% utilization a GPU generates on the order of ~$10–15k of annual revenue, against which Nebius’s own 4-year depreciation implies ~$7.5–10k/yr of depreciation on the GPU alone — before power (a large cash cost), networking/facility depreciation, interest on the convertibles funding the build, and SG&A. The 45% “segment adjusted EBITDA margin” is struck above the depreciation and interest lines; after honest depreciation and the ~4–7% effective cost of the convertible debt, the after-tax return on the incremental capital is, on these assumptions, marginal to thin unless (a) utilization and pricing stay at today’s elevated levels for the full asset life and (b) the GPUs earn a meaningful “second life” in inference after their training prime. That is precisely the bull case — and precisely what a capital-cycle downturn would break. The honest conclusion: the cluster can be a good investment at peak pricing and a poor one at trough pricing, and adjusted EBITDA is structurally silent on which regime you are funding into. This is why the metric matters so much and why the prudent bias is to anchor on GAAP operating losses and cash burn rather than the adjusted figure.

Verdict — economics do not yet improve with scale; they are still being purchased. The continuing business loses money at the operating line and at the cash-flow line; the only “profit” is paper; the depreciation burden is still climbing; and the favorable unit economics management cites (45% segment EBITDA) sit above a depreciation wall that has not yet fully arrived. There is a plausible path to genuine operating leverage if ARR scales into the fixed-cost base and pricing holds — the Q1’26 inflection is real and encouraging — but as of today the financial-quality evidence is skeptical-to-negative.

7. Capital Allocation

The Yandex divestiture — a competent forced exit. Faced with a stranded, sanctioned Russian asset base, management converted it into ~$1.47B of clean cash ($1,283.2M + $184.2M per the cash-flow statement; the ~$5.2B headline included non-cash and assumed components) plus a Nasdaq relisting, booking a −$784.6M disposal loss. Salvaging a Western, investable company out of an impossible geopolitical situation was a capable piece of capital allocation under duress.

The pivot — an all-in bet at the cycle peak. Divestiture proceeds plus ~$4.2B of FY2025 convertibles plus ~$1.15B of FY2025 equity went almost entirely into GPUs and data centers (PP&E $128M → $5,553M in two years). This is a maximum-conviction bet on AI infrastructure made at the top of a >$700B industry capex cycle — precisely the configuration Marathon’s capital-cycle framework warns against. It may prove brilliant if demand and power scarcity persist; it is, unavoidably, pro-cyclical capital allocation into the most crowded trade in technology.

“Spin-and-retain” produces paper, not cash. Nebius incubates ventures, deconsolidates them to shed the cash drag, and retains minority upside: ClickHouse (marked ~$737M at Series C, stepped up again at the ~$15B Series D), Toloka (deconsolidated, $130.6M retained), Avride (Uber + Nebius committing up to $375M). The recurring non-cash marks on these stakes are the main thing keeping reported net income positive — but they generate little operating cash and the stakes are illiquid. This is value creation in an accounting sense; it is not (yet) distributable cash.

M&A — small and capability-driven. Tavily ($177.3M cash + $30.3M earnouts, Feb 2026), Eigen AI, and Clarifai are inference/agentic “acqui-hires,” not scale plays. Disciplined in size; sensible in rationale (accelerate the inference roadmap and acquire scarce talent).

The funding gap is the central capital-allocation question. FY2026 capex guidance of $20–25B against ~$9.3B of cash and ~$3B of revenue means the model is predicated on permanent capital-markets access. Even after raising >$6B YTD 2026 ($4.34B of March converts at 1.25%/2.625% cash coupons; NVIDIA’s ~$2.0B), Nebius still needs roughly $12–18B more in 2026 alone. The stated funding stack: customer prepayments (non-dilutive but front-loads counterparty risk and delivery obligations), asset-backed financing collateralized on Meta/Microsoft contract receivables (leveraging those counterparties’ credit), additional converts/corporate debt, and an authorized-but-unused 25M-share ATM plus strategic equity.

Dilution is structural and stacking. On a ~253M basic share base (219.5M Class A + 33.55M Class B): the FY2025 converts alone represent ~66M Class A (strikes ~$51 and ~$139); the March-2026 converts add ~25–30M more; the NVIDIA warrant of ~21.1M shares at $0.0001 is near-certain (~9.5% effective dilution already); plus M&A earnout shares, ~16%-of-revenue SBC, and any ATM. If the stock holds above the convert strikes, fully-diluted Class A could rise 40–50%+ over the convert lives. Convert effective interest rates of ~4–7% (versus the 1–3% cash coupons) also mean rising non-cash interest accretion through the income statement.

Insider signal — neutral-to-soft-negative. Every one of the trailing twelve Form 4s (Apr–Jun 2026) is a sale; zero open-market purchases. The signal is softened — the CEO’s was a mandatory RSU sell-to-cover, and others were under Rule 10b5-1 plans — but the complete absence of any insider buying as the stock ran from ~$100 to ~$240 is itself notable. Foreign-private-issuer status exempts insiders from Section 16 short-swing rules.

Governance. Founder/CEO Arkady Volozh controls ~52% of the vote on ~11% of the economics via Class B super-voting shares; Nebius is a “Controlled Company” under Nasdaq rules with the associated reduced minority protections.

Verdict — a capable forced exit followed by a pro-cyclical, serially-dilutive all-in bet. Management is clearly capable and the strategic moves (Yandex salvage, NVIDIA/Meta/Microsoft anchoring, owned-power vertical integration) are coherent. But the capital intensity, the funding gap, and the dilution arithmetic mean shareholder value depends as much on the terms and dilution of the next $12–18B as on operational execution. On the principle that capital allocation is the bridge between business value and shareholder value, the jury is out, and the bias is cautious.

8. Changes and Headwinds — Last Two Years

Nebius is barely 18 months old as a standalone Western entity, and the period is a near-total reinvention.

• Corporate transformation (Jul–Oct 2024): Yandex N.V. sold all Russian assets (July 2024); reorganized the non-Russian businesses as Nebius Group; resumed Nasdaq trading in October 2024 after a multi-year halt. Founder Volozh was removed from the EU sanctions list in March 2024 and renounced Russian citizenship in February 2026. The FY2025 20-F auditor states the prior going-concern doubt has been removed. The legacy-Russia overhang is materially reduced, though residual reputational/headline risk persists.
• Capacity ramp: contracted power went from ~2GW (late 2025) to >3.5GW (Q1’26 call), targeting ≥4GW for 2026, >75% on owned sites. Pennsylvania became the second owned GW-scale US site.
• Capex hikes: FY2026 capex guidance was raised from $16–20B (Q4’25 call) to $20–25B (Q1’26 call), explicitly to fund 2027 capacity — not component inflation (pegged at low-single-digit % of spend).
• Anchor contracts: Microsoft (up to ~$17.4B through 2031 per the 20-F; some press reports cite ~$19.4B, likely core vs. core-plus-optional); Meta ($27B/5yr = $12B dedicated from early 2027 + $15B optional capacity at Nebius’s discretion); NVIDIA ($2B equity + 5GW-by-2030 line of sight). The first Meta contract was fully delivered in Q1’26.
• M&A: Tavily, Eigen AI, Clarifai — inference/agentic capability acquisitions.
• Accounting/earnings-quality: the indicated move toward a 5-year GPU depreciable life for 2026 is an EBIT-flattering change to monitor.

Verdict. On balance these changes strengthen the growth and optionality case (a clean Western entity, NVIDIA-blessed, hyperscaler-anchored, owned full stack) while sharply raising the financing and execution bar. The thesis now lives or dies on funding $20–25B/yr of capex without ruinous dilution.

9. Risk Analysis

The two risks that actually move the thesis are the financing gap (High/High) and customer concentration (the 83%-of-receivables single-customer datapoint is the most underappreciated number in the file). The capital-cycle/pricing and demand-cycle risks are the medium-term governors; everything else modulates severity.

10. Valuation Discussion (Embedded Expectations)

No price target. Scenario and embedded-expectations framing only.

Anchors (reconciled to market data, 2026-06-10): price $220.12; market cap ~$55.9B; EV ~$56.6B (essentially net cash: ~$9.6B debt vs ~$9.4B cash); ~220.4M shares. Forward set: exit-2026 ARR guide $7–9B; FY2026 revenue $3.0–3.4B; group adjusted-EBITDA margin guide ~40%; RPO/backlog $21.3B (end-2025); management’s medium-term EBIT-margin aspiration ~20–30%.

Trailing multiples are meaningless; only forward is honest. TTM ratios (~65× P/S, ~65× EV/Revenue) are artefacts of triple-digit growth. The useful lenses: EV / FY26-revenue midpoint ($3.2B) ≈ 17.7×, and EV / exit-2026-ARR midpoint ($8B) ≈ 7.1×.

What the price embeds. To rationalize ~$57B of EV against the multiple a capital-intensive infrastructure business eventually earns — call it ~6–8× EV/EBITDA at maturity — Nebius needs roughly $7–9B of steady-state EBITDA. At a mature ~35–40% EBITDA margin, that implies ~$18–25B of revenue (~2.5–3× exit-2026 ARR), reached around 2028–29, with margins surviving the depreciation wall and the $20–25B/yr capex funded largely with non-dilutive prepayments and asset-backed debt. That is the triple condition embedded in today’s price: compounding × margin durability × non-dilutive funding. Each is individually plausible; the conjunction of all three is what the buyer at ~$220 is underwriting.

CoreWeave comp (the key public neocloud benchmark):

(CRWV figures: stockanalysis.com; CNBC 2026-05-07; Seeking Alpha; public market data.) On current-year revenue NBIS looks far richer (~17.7× vs ~6.7×), reflecting its earlier ramp; on forward ARR the gap narrows but NBIS still carries a premium. The premium is defensible on two grounds — NBIS’s net-cash balance sheet (vs CRWV’s $35B debt and ~$2.1B annual interest) and its owned-stack + inference-software optionality — and discountable on two — smaller scale and a far shorter backlog ($21.3B vs $99.4B).

Scenario EV ranges (illustrative; not targets):

• Bear (~$15–25B EV): ARR stalls at $8–12B as the capital cycle turns, GPU pricing softens, or a hyperscaler trims; dilution compounds; mature EBITDA margin compresses to ~20–25%. ~2–4× forward ARR. Material downside from current EV.
• Base (~$45–65B EV): ARR reaches ~$15–20B by 2028 at ~30–35% EBITDA; capex funded with a tolerable prepay/debt/modest-equity mix. ~5–7× trough-forward-ARR — roughly the current EV.
• Bull (~$90–130B+ EV): Nebius becomes a top-tier AI-native hyperscaler; ARR $25–35B+ by 2028–29 at 35–40%+ EBITDA; inference software lifts blended margins; owned-stack ROIC exceeds cost of capital; dilution contained — a CoreWeave-or-better multiple on a much larger base.

A reverse-DCF / capital-returns cross-check. Approach the same question from cash returns. Assume the bull’s trajectory: ARR scales to ~$20B of revenue by ~2029 at a genuine (post-depreciation) operating margin of ~20% — a 20-F-consistent medium-term aspiration — for ~$4B of EBIT, ~$3B after tax. But the business must keep spending to grow and to replace fast-depreciating GPUs: if maintenance-plus-growth capex consumes the bulk of operating cash flow through the build, free cash flow to equity stays modest or negative until growth slows. Discounting a free-cash-flow stream that only turns meaningfully positive in the early 2030s — and then haircutting for 40–50% potential share dilution along the way — it is difficult to justify ~$57B of equity value without assuming both that the terminal margin lands at the high end (≥25–30% sustained) and that the reinvestment treadmill eases (i.e., GPU economic life is genuinely 5+ years and pricing holds). Flip it around: the capital-returns lens (Marathon) asks what return the market expects on the ~$20–25B/yr of incremental capital. At today’s valuation the implied expected return on that capital is high and sustained — exactly the configuration that the capital cycle says mean-reverts as supply floods in. The reverse-DCF and the capital-cycle lens agree: the price requires the favorable regime to persist for most of the asset base’s life.

Where the market is right vs. wrong. Right that demand is real now and that Nebius is uniquely positioned among independents. Possibly wrong in extrapolating sold-out, rising-price conditions through a full capital cycle, and in under-weighting the dilution arithmetic of self-funding $20–25B/yr. No own-history valuation percentile is meaningful — the post-relisting trading history is too short.

11. Variant Perception

Consensus (very bullish). Average analyst target ~$238; BofA raised to $280 (2026-06-08); the rating distribution skews strongly to Buy. The Street treats Nebius as an AI-native hyperscaler in the early innings of a generational build-out, and the recent flow (UK £1.7B expansion, the Kao Data 10-year deal, the Meta contract) has reinforced that read.

Strongest bull case. A genuinely differentiated, owned-full-stack AI cloud compounding into a multi-hundred-billion-dollar TAM: a $21.3B backlog, NVIDIA equity plus supply priority, Meta/Microsoft anchors, a net-cash balance sheet (rare among neoclouds), founder-led, sold out every quarter with rising prices and lengthening contract durations, and an inference/Token-Factory software layer that should lift blended margins and stickiness over time.

Strongest bear case. Commodity GPU rental at a capital-cycle peak: ruinous capex and GPU depreciation, 83%-of-receivables single-customer concentration, circular NVIDIA financing, a structural funding gap that forces continuous dilution, and no proven durable moat — the software stack is asserted but not yet shown to defend pricing through a downturn. The reported “profit” is a non-cash venture mark; full-year adjusted EBITDA was negative.

The 3–5 assumptions that matter most: (1) AI-compute demand stays supply-constrained for the next 2–3 years (no hyperscaler capex pause); (2) GPU pricing/utilization holds as supply floods in (the capital-cycle question); (3) the $20–25B/yr capex is funded mostly non-dilutively; (4) the owned-stack/inference software delivers durable margin and ROIC above cost of capital, not just a transient demand premium; (5) the GPU depreciable life assumption survives the Rubin/Vera transition.

Falsification tests.

• Bull breaks if: a hyperscaler trims or cancels; GPU spot/contract pricing rolls over for two consecutive quarters; ATM/equity issuance becomes a primary funding source (dilution); or core-cloud adjusted-EBITDA margin compresses while utilization stays high (pricing pressure).
• Bear breaks if: ARR tracks toward or through the high end of the $7–9B exit-2026 guide; contract durations and ASPs keep extending into a normalizing supply environment; capex is funded via prepayments/asset-backed financing at attractive terms with minimal equity; and core AI-cloud EBITDA margin holds ~45%+ as scale builds.

12. Fact vs. Interpretation Table

13. Open Questions

1. Microsoft contract value — 20-F states “up to ~$17.4B through 2031”; some press cites ~$19.4B. Likely core vs. core-plus-optional; unreconciled.
2. GPU depreciable life — what is the FY2026 policy (4 vs 5 years), and what is the EBIT impact of any change? The FY2025 20-F note reflects 4 years.
3. The $15B Meta “optional” tranche — revenue-recognition and margin treatment are unconfirmed; management frames it as a financing backstop, not committed revenue.
4. True normalized unit economics — what is the GPU-cluster ROIC after honest (not adjusted-EBITDA) depreciation and cost of capital? Management has not disclosed a clean cohort-level return.
5. Funding mix of the next $12–18B — how much will be prepayments/ABS (non-dilutive) versus equity/ATM (dilutive)? This single variable dominates per-share outcomes.
6. Material weaknesses — timeline to remediate the fixed-asset and TripleTen revenue-recognition control weaknesses.
7. Inference/Token Factory traction — is it building measurable switching costs and retention, or is it still a small, demand-premium feature?

14. What Must Be True

Bull case — what must be true, and its falsification test. For the bull case to be correct, Nebius must (a) keep its capacity sold out at firm-or-rising prices through 2027 as industry supply surges, (b) fund the $20–25B/yr build largely without equity dilution via prepayments and asset-backed financing on Meta/Microsoft credit, and © convert the 45% one-quarter segment EBITDA margin into a durable, depreciation-bearing margin as ARR scales toward $7–9B and beyond. Falsification test: if GPU contract/spot pricing rolls over for two consecutive quarters, or the ATM/equity becomes a primary funding source, or core AI-cloud EBITDA margin compresses while utilization stays high, the bull thesis is broken.

Bear case — what must be true, and its falsification test. For the bear case to be correct, the neocloud capital cycle must turn — supply catching demand, ASPs compressing — and/or a major customer must trim, and/or the funding gap must force value-destructive dilution, exposing a commodity rental business with no durable moat and paper-only profits. Falsification test: if ARR tracks toward or through the high end of the $7–9B exit-2026 guide, contract durations and ASPs keep extending into a normalizing supply environment, the capex is funded at attractive non-dilutive terms, and core EBITDA margin holds ~45%+ as scale builds, the bear thesis is broken.

The two cases are not symmetric in evidence: the bull case rests on conditions observable today (sold-out capacity, rising prices, a huge backlog); the bear case rests on conditions that typically arrive later in a capital cycle (supply catch-up, pricing reversion, dilution). The disagreement is therefore largely about timing and durability, not about the current facts.

15. Source Appendix

See the Source Appendix below for the full enumerated source list with URLs and access dates. Primary sources include: Nebius Group FY2025 Form 20-F (SEC, filed 2026-04-30) and the FY2024 20-F; the Q1 2026, Q4 2025, and Q3 2025 earnings-call transcripts and the BofA/Morgan Stanley/UBS conference presentations; the Q1 2026 6-K (period end 2026-03-31); SEC XBRL company facts (CIK 0001513845); the SEC Form 4 insider corpus; and SEC Schedule 13G filings. Secondary sources include SemiAnalysis ClusterMAX 2.0, Datacenter Knowledge, CoreWeave public filings/coverage (stockanalysis.com, CNBC, Seeking Alpha), Tom’s Hardware and Man Group on AI capex, and GPU-depreciation analyses. Market data via public market-data sources (price, snapshot, curated news), reconciled to filings.

The body of this article carries no investment recommendation and no price target; the only position expressed is the clearly-labeled Claude’s Take block at the top, which is the author’s own subjective view. This is general information, not investment advice; do your own research.

APPENDIX A — Standard Diligence Questionnaire

Nebius Group N.V. (NASDAQ: NBIS) — Standard Diligence Questionnaire

Supplemental to the analysis above. Grounded in the same primary sources; Fact / Interpretation / Assumption labels used where it matters. Where a question does not map to the business model, the correct analog is given.

General

What thoughtful questions have other investors asked about this company? The recurring institutional questions (from the Q1’26 / Q4’25 calls and the BofA/Morgan Stanley conferences): (1) How will you fund $20–25B of capex without crushing dilution? — answered with prepayments, asset-backed financing on Meta/Microsoft credit, corporate debt, and an unused ATM. (2) How concentrated is the customer base, and is Meta/Microsoft a strength or a fragility? (3) How does the $27B Meta deal (the $15B optional tranche) actually work and book? (4) Is GPU pricing cyclical, and what happens to the 40–45% margin when supply floods in? (5) Are Token Factory/software a distinct, sticky revenue stream or just an enabler? — management says explicitly an enabler, not a separate revenue line. (6) Capacity timing — are sites (Vineland NJ, Pennsylvania) on schedule? Skeptics additionally focus on the non-cash nature of reported profits and the circular NVIDIA financing.

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? (Interpretation) There are no “earnings” to speak of — the continuing business loses ~$612M at the GAAP operating line. Revenue and pricing, however, are arguably at a cyclical high: the company is sold out, raising prices, with “4+ customers per GPU,” in the euphoric phase of an industry capital cycle. The risk is that today’s pricing and the 45% segment EBITDA margin are nearer a peak than a trough.

Driven by the external environment or internal actions? Both. The external AI-capex boom (~$725B of 2026 Big-Tech spend) drives demand; internal actions (securing power, GPU allocation, owned-stack build, capital raises) determine how much of it Nebius can capture. Growth is capacity-gated (internal execution + capital), not demand-gated.

How stable are revenues? Increasingly contracted (>$21B RPO, lengthening durations) but concentrated — Customer D was 83% of gross receivables at end-2025. Stability of contracted revenue is improving; stability of pricing on future capacity is the open question.

Outlook for products/services? Strong near-term: exit-2026 ARR guided to $7–9B, FY2026 revenue $3.0–3.4B. Inference (Token Factory) is the fastest-growing layer; agentic is nascent.

How big will this market be — growing, shrinking, domestic or international? Growing rapidly and global (US, UK, Finland, EU). The AI-compute TAM is large and expanding, but the rental-layer profit pool is contested and prone to capital-cycle mean reversion.

Business Quality & Competitive Moat

Is the industry getting more or less competitive? More. Hyperscalers, CoreWeave (~5× Nebius’s scale), Crusoe, Lambda, Nscale, and others are all adding capacity; ~$725B of 2026 industry capex floods supply into the rental layer.

How profitable is the business (ROIC, ROE)? (Interpretation) GAAP ROIC/ROE are negative on an operating basis; reported positive net income and the 14% TTM ROE are artefacts of non-cash equity-stake marks, not operating returns. True cluster-level ROIC after honest depreciation and cost of capital is undisclosed and is an open question.

How profitable is the industry — how many competitors, what barriers to entry? The profit pool sits with NVIDIA (~75% GPU gross margin) and transiently with neoclouds while supply is short. Barriers to entry are capital/power barriers (GPU allocation, multi-GW power, tens of billions of financing) — real against small entrants, ineffective against hyperscalers.

Can the business be easily understood? Yes at the surface (rent GPU compute) but the financials are complicated by discontinued operations (Russia), non-cash equity-stake marks, convertibles, and customer prepayments.

Can it be undermined by foreign low-cost labor? Not directly relevant — the competitive variables are GPU access, power cost, capital cost, and engineering talent, not labor arbitrage.

Do brands matter? Marginally. Reputation/reliability and SemiAnalysis-type ratings (Nebius = Gold; CoreWeave = Platinum) matter for enterprise trust, but AI compute is largely fungible — buyers multi-source on price and availability.

What is the nature of competition? Price, GPU availability/speed-to-provision, contract terms, and increasingly software/inference optimization. Today it is a scramble for scarce capacity; tomorrow it is more likely a price war as supply normalizes.

Customers’ switching costs? Low at the bare-metal/training layer (fungible compute, multi-vendor sourcing); moderate and rising at the managed-cloud/Token Factory/inference layer (model deployment, MLOps integration). The durable-moat question hinges on the latter, which is small today.

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? Yes — equity stakes (ClickHouse, Toloka, Avride) carry option value; the NVIDIA supply relationship and contracted-power positions are strategic assets not on the balance sheet.

Off-balance-sheet liabilities? Large purchase/capacity commitments and the obligation to deliver against ~$2.3B+ of customer prepayments (deferred revenue, on-balance-sheet but a delivery liability rather than free cash). Operating-lease and power-contract commitments are substantial.

How conservative is the accounting? Mixed. Conservative: 4-year GPU depreciable life (vs CoreWeave’s 6). Aggressive/optical: reliance on adjusted EBITDA that adds back the dominant D&A line; non-cash equity-stake marks driving reported net income; “operating cash flow” inflated by prepayments. Material weaknesses disclosed in fixed-asset and TripleTen revenue-recognition controls. Net: treat reported profitability skeptically.

How CapEx-hungry is the business? Extraordinarily — $20–25B of 2026 capex against ~$3B of revenue. This is among the most capital-intensive business models in public markets.

Capital Allocation & Management

How much FCF does the business generate, and how does management use it? It does not generate free cash flow — FY2025 FCF was roughly −$3.66B. All capital (divestiture proceeds, converts, equity, prepayments) is deployed into the GPU/data-center build.

Significant acquisitions recently? Small, capability-driven: Tavily (~$177M cash + $30M earnouts), Eigen AI, Clarifai (inference/agentic acqui-hires). The defining transaction was the divestiture of all Russian assets (2024).

Buying back shares? No — the company is a net issuer (converts, NVIDIA warrant, ATM authorization, SBC).

Issuing large amounts of new shares to insiders? SBC runs ~16% of revenue — high. The NVIDIA warrant (~21.1M shares) and stacked convertibles imply 40–50%+ potential Class A dilution over the convert lives.

Compensation policy of directors/management? Equity-heavy; insiders have been net sellers (all trailing-12 Form 4s are sales, mostly 10b5-1/sell-to-cover). Founder Volozh holds Class B super-voting control.

Motivations of management? (Interpretation) Founder-led, building an “AI-native hyperscaler” — strategic and growth-maximizing, with a credible execution record (the Yandex salvage). Aligned with long-term value via ownership, but the controlling-shareholder structure limits minority influence.

Valuation & Market Data

Is the stock an ADR, MLP, or K-1 issuer? No K-1/MLP. NBIS is the Class A ordinary share of a Netherlands company listed directly on Nasdaq; it is a foreign private issuer (files 20-F/6-K, not 10-K/10-Q), with the associated reduced disclosure cadence and Section-16 exemptions for insiders.

Dividend policy? None — no dividend; all capital reinvested.

How profitable is the business? Not profitable on an operating or cash basis; reported net income is non-cash marks (see above).

Is net income diverging from cash from operations? Yes, sharply and in both directions — reported net income is inflated by non-cash equity marks, while “operating cash flow” is inflated by customer prepayments. Neither reflects underlying operating profitability, which is negative.

Risks & Downside

What factors would cause the stock to decline? A hyperscaler trimming/cancelling; GPU pricing rollover; a dilutive equity raise; a capacity-delivery slip; an AI-capex pause; an impairment or markdown of an equity stake; or simple multiple compression from a demanding starting valuation.

Risk of a catastrophic loss? (Interpretation) Moderate and scenario-dependent. The net-cash balance sheet and $21B backlog reduce near-term solvency risk, but a capital-cycle downturn combined with the funding gap could force deeply dilutive raises or distressed asset sales. Customer concentration (83% of AR) is a single-point-of-failure risk.

Chance of a total loss? Low in the near term (net cash, real revenue, real assets), but not negligible over a multi-year horizon if the capital cycle turns hard while the company is mid-build and dependent on continuous external financing.

Recent News & Events

Has the business environment changed recently? Yes — favorably on demand (sold out, rising prices, pipeline +3.5× QoQ) and on validation (Meta $27B, Microsoft, NVIDIA $2B equity, UK £1.7B, Kao Data 10-year deal). The financing bar rose with the capex-guide hike to $20–25B.

Significant acquisitions? Tavily, Eigen AI, Clarifai (2026). Divestiture of Toloka (deconsolidated). The 2024 Russia divestiture remains the defining corporate event.

Change in accounting policies? An indicated move toward a 5-year GPU depreciable life for 2026 (FY2025 policy note reflects 4 years) — an EBIT-flattering change to monitor. Disclosed material weaknesses in fixed-asset and TripleTen revenue-recognition controls.

Recent changes — new markets, facilities, management? New US sites (Pennsylvania 1.2GW, Missouri, Alabama); UK expansion (4 NVIDIA sites); senior go-to-market hires (Americas, APAC, Middle East GMs). Founder Volozh renounced Russian citizenship (Feb 2026) and was removed from EU sanctions (Mar 2024).

APPENDIX B — Source Appendix

Nebius Group N.V. (NASDAQ: NBIS) — Source Appendix

Primary sources prioritized. Third-party data reconciled to primary filings. Access date 2026-06-10 unless noted.

Primary — SEC filings (CIK 0001513845; foreign private issuer)

Primary — Management transcripts (earnings calls & conference presentations)

Secondary — industry, competition, valuation

Market data (public sources, reconciled to filings)

Note on authority: for this US-listed foreign private issuer, the SEC 20-F/6-K filings and XBRL are primary. Public market-data and news aggregators are convenience/third-party sources used for orientation and reconciled to filings. Third-party sentiment scores and analyst targets are signals, not evidence, and never inform the valuation conclusion.