First or fastest mover in a massive TAM — satisfied on speed, not on primacy.
The market for AI compute is measured in hundreds of billions of dollars a year and is supply-constrained: the five largest hyperscalers alone are spending more than $700 billion on capital expenditure in 2026 and still buy capacity from third parties. Nebius was not first — CoreWeave reached scale earlier — but it is among the fastest. Run-rate revenue grew nearly sevenfold in a year, contracted power guidance for the end of 2026 has been raised to 5 gigawatts, connected power will reach 0.8–1 gigawatt, and management plans to add more than a gigawatt a year from 2027 across sites in the United States, the Nordics, the United Kingdom, France, Spain and Israel. The runway is long; the field is crowded and includes the company's own largest customers.

Network effects and data flywheel — partially satisfied.
Renting compute has no inherent network effect, and that is the central weakness of every company in this category. Nebius is building the closest available substitutes. Token Factory serves open models by the token and production inference volumes tripled in the latest quarter; Tavily, the agent-search service, has 2.5 million developers, up from 1 million in February; and the cloud platform adds identity, key management, cost controls and an AI operations agent with each release. These create switching costs and a developer funnel that pure capacity lessors lack, and operating data from one of the larger GPU fleets outside the hyperscalers feeds back into scheduling and hardware design. It is a real but still thin flywheel — most revenue today comes from large reserved-capacity contracts, not from the platform layer.

Disruptive technology — satisfied at the system level.
The differentiation is vertical integration inherited from two decades of building search infrastructure: in-house server and rack design, proprietary cloud software, and data centers engineered for power efficiency. The evidence is in the numbers rather than the brochure — a core adjusted EBITDA margin of 49.7%, up from 24% two quarters earlier, contract values of $20–25 million per megawatt a year against roughly $12 million at the start of 2026, and the first Nvidia Vera Rubin systems already received. The dependency is equally clear: the product is Nvidia silicon, Nvidia is a shareholder and warrant holder, and allocation of each new generation is outside the company's control.

AI-disruption-resistance — two anchors satisfied.
Nebius sits on the AI-infrastructure anchor — it sells the compute that models are trained and served on — and on the physical anchor, since the value chain ends in powered, cooled buildings full of hardware that no software agent can intermediate. No AI-disruption risk flag is required. The score is held at 7.0 because the moat is execution speed and integration rather than a structural monopoly: compute is ultimately a commodity, the customers with the largest contracts are also building their own capacity, and at least four well-funded rivals are pursuing the same contracts.

Trait 1 — Missionary vision (20%) — 9/10
Arkady Volozh co-founded Yandex in 1997 and spent twenty-five years proving that an independent technology company could out-engineer a global giant on its home ground. Nebius is the same ambition restated for the AI era: an independent, full-stack AI cloud built by engineers, positioned between the hyperscalers and the capacity lessors. The vision is specific enough to guide capital — own the hardware design, own the software, serve AI builders directly — and it explains why the group kept its data-center engineers and let go of everything else in 2024.

Trait 2 — Radical long-termism & skin in the game (25%) — 9/10
Volozh holds roughly 11% of the economics and about 52% of the votes through ten-vote Class B shares, the structural control the framework looks for. Behaviour matches structure: the company is spending seven times its revenue on capacity this year, and management has said it is deliberately holding back some capacity from long-term contracts because it expects pricing to keep rising. Two caveats. On a fully diluted basis his voting power falls below 50%, so control is strong rather than absolute; and he was subject to European Union sanctions from 2022 until they were lifted in March 2024, a history that is resolved but belongs in any honest account.

Trait 3 — Product & customer obsession (20%) — 8/10
Shareholder letters read like release notes: a new cloud platform version each quarter, storage read speeds up 30%, a shared file system three times faster, accuracy-parity benchmarks for each newly hosted open model, keyless pay-per-search for autonomous agents. Acquired teams from Eigen AI and Clarifai were folded into the inference product within months. The deduction is for mix — the largest contracts are bare capacity for Microsoft and Meta, where the product layer matters least.

Trait 4 — Execution velocity (20%) — 9/10
Run-rate revenue of about $430 million in June 2025 became $3.0 billion in June 2026. Every tranche owed to Microsoft has been delivered on schedule, including a switch of the power source at one site to fuel cells without a timeline slip. Four contracts of more than $1 billion each were signed in a single quarter, contracted-power guidance was raised from more than 4 to 5 gigawatts, and full-year guidance has been reaffirmed or raised at every report in 2026.

Trait 5 — Capital efficiency & financial discipline (10%) — 6/10
The unit economics are improving fast — payback on new capacity under two years, customer prepayments above $9 billion expected in 2026, and a first asset-backed facility at 2.5 points over the benchmark rate. But the funding mix is still heavily dilutive: about $15 billion of convertible notes across four issues in fifteen months, $2.8 billion of shares sold at the market in one quarter, Nvidia warrants, and trailing free cash flow of minus $5.9 billion. Depreciation already exceeds adjusted EBITDA, so GAAP profitability is some distance away. Discipline is visible in the terms; it is not yet visible in the share count.

Trait 6 — Talent magnetism & organisational scaling (5%) — 8/10
The founding asset was a team of roughly a thousand infrastructure engineers who had already built and run one of the largest non-American cloud and search platforms, and who moved as a group. Few competitors began with that depth. The leadership bench around the founder is experienced, and the organisation has scaled across a dozen countries in two years. Dependence on a single founder remains high.

Valuation — FLAG
On the framework's primary basis — the most recent full fiscal year, 2025 revenue of $529.8 million — the price-to-sales ratio is 115.6x, a number that mostly measures how fast the company is changing. On trailing twelve-month revenue of $1.36 billion it is 45.1x; on the midpoint of 2026 guidance ($3.2 billion) it is 19.1x; and on the midpoint of the year-end run-rate guide ($8 billion) it is 7.7x. None of these is near the sub-5x entry threshold, and this is not an asset-light model that earns an exemption. The headline market capitalisation also understates what is being paid. It uses 274 million basic shares; in-the-money convertibles, Nvidia warrants, options and restricted stock take the fully diluted count to roughly 370 million (an estimate), which puts the diluted market capitalisation near $83 billion and every multiple above about a third higher.

Revenue and margin trajectory
Revenue was $582.3 million in the second quarter of 2026, up 454% year over year and 46% sequentially, with the core AI cloud at $574.9 million. Run-rate revenue reached $3.0 billion at the end of June and is guided to $7–9 billion by December, which makes 2027 revenue of $9–10 billion plausible from contracts already signed. Reported gross margin is 77% and group adjusted EBITDA margin 41%, but both exclude depreciation, which is the real cost of this business: trailing operating income is about minus $500 million and the GAAP net loss was $190.4 million in the quarter. The margin that matters — EBITDA less depreciation on a fleet with a five-to-six-year life — is still negative and will stay so while the fleet is growing faster than revenue.

Balance sheet and path to profitability
Cash was $8.0 billion at the end of June, before $5.75 billion of convertible notes issued in August; total debt is about $16 billion on a pro-forma basis (an estimate), almost all of it low-coupon convertibles maturing from 2029 to 2034. Trailing free cash flow is minus $5.9 billion and 2026 capex is guided to $20–25 billion, funded by roughly $9 billion of customer prepayments, the convertibles, share sales and a first $775 million asset-backed facility. The path to profitability is credible — contracted cash flows, sub-two-year payback, investment-grade counterparties — and the company can stop building and turn cash-positive if it chooses. It does not pass the framework's test of a balance sheet that avoids dilution; dilution is the funding model.

Continuous Dilution
About $15 billion of convertibles in fifteen months, $2.8 billion of shares sold at the market in a single quarter and 21 million Nvidia warrants mean the fully diluted share count is already about a third above the basic count, with more to come while capex runs at several times revenue. Even if the business compounds as hoped, a material part of the value created accrues to new capital rather than to existing shares.

Customers Who Are Also Competitors
Microsoft and Meta account for the large majority of contracted backlog; at the end of 2025 only about $1 billion of $21 billion in remaining performance obligations came from outside big technology companies. Both are building their own capacity at far larger scale and are buying from Nebius because they are short of it today. When their own supply catches up, renewals will be negotiated from a position of strength — and Microsoft retains the right to reject tranches that miss specification.

Commoditisation and Price Deflation in Compute
Current pricing reflects scarcity. Each new accelerator generation lowers the cost of a unit of compute, and tens of gigawatts of capacity are under construction across the industry. A fleet bought at today's prices and depreciated over five to six years is exposed if rental rates for older hardware fall faster than the depreciation schedule assumes.

Funding-Window Risk
The model depends on capital markets staying open to convertible issuance and asset-backed borrowing at low spreads. A credit event elsewhere in the AI-infrastructure chain, or a sustained rise in long-term interest rates, would raise the cost of every gigawatt still to be built and could force growth to slow at exactly the moment the market is valuing it.

Nvidia Dependence
The product is Nvidia hardware, Nvidia is an investor and warrant holder, and early access to each generation is a competitive necessity rather than a right. A change in allocation policy, or customers shifting workloads to custom accelerators that Nebius does not offer, would weaken the proposition.

Key-Person and Governance
Strategy, culture and control rest with one founder whose voting majority is no longer absolute on a fully diluted basis. The group also carries non-core assets — a majority economic interest in Toloka, about a quarter of ClickHouse, the Avride autonomous-vehicle unit and the TripleTen education business — which add optionality but also losses and management distraction.