Nebius Group N.V. (NBIS)
Initiation Report

Nebius has converted AI-infrastructure demand into a $3.0bn June ARR run-rate and major customer commitments. The investment question is whether powered capacity can come online fast enough—and at sufficient utilization—to absorb Q2 capex of $5.657bn, interest and depreciation without excessive dilution.

US GAAP is the reporting anchor; adjusted EBITDA retains the issuer definition.

The AI-factory thesis is a conversion problem. Large contracts improve demand visibility and prepayments help fund hardware, but revenue and GAAP earnings depend on when powered capacity is commissioned and utilized.

Falsification is economic: ARR can grow while GAAP EPS and cash returns deteriorate if commissioning or utilization lags the asset base.

The twin centers on Nebius AI Cloud and preserves separate routes for facilities, products, NVIDIA supply, direct customers, AI labs, geographies, Avride, TripleTen and the Toloka investment. Every accepted entity and relationship is shown; topology is not financial weighting.

Nebius ended June at $3.0bn ARR and raised its 2026 contracted-power target to 5GW. Contracted power defines the build ceiling; energized racks, delivered GPUs, networking and customer acceptance determine when capacity begins generating revenue. [E1] [E2]

Large commitments create demand visibility but not immediate revenue. Most Q2 wins are tied to late-2026 capacity and contribute mainly in 2027; Q2 agreements averaged more than $20m annual contract value per MW, with customer prepayments covering 50–60% of associated capex in about 70% of deals. [E2] [E4]

Nebius AI Cloud generated $574.9m of Q2 revenue. Aether, AI Studio, Token Factory and Echo extend the stack above raw GPU capacity; Avride and TripleTen remain distinct options. Utilization and higher-value service mix—not product count—drive unit economics. [E1] [E2]

Product breadth supports customer utility; it is not a substitute for utilization, revenue recognition or gross-margin evidence.

Q2 capex was about $5.7bn. June cash was $8.042bn and gross debt $8.546bn; deferred revenue of $5.975bn is a contract liability, not debt. D&A reached $259.7m after the server/network useful-life policy moved from four to five years. [E1] [E2] [E5]

The ownership diagram explains who funds each layer; it does not prove a margin or capex advantage until actual contract economics are disclosed.

The issuer Q2 release is the source of record for current actuals. GAAP and non-GAAP measures, continuing operations and the five-year server/network useful life remain explicit.

Accounting discipline: † issuer-defined non-GAAP. Adjusted EBITDA is not combined with GAAP EPS, and deferred revenue is not debt.

End-June ARR annualizes the final month of AI Cloud revenue. It does not by itself produce a Group revenue forecast, because scope, contract start, commissioning, acceptance and non-cloud revenue remain separate.

Withholding an unsupported point estimate is more decision-useful than presenting an arithmetically precise but scope-inconsistent forecast.

Management reiterated its FY2026 outlook; the numerical $3.0–3.4bn revenue range was disclosed in the Q1 shareholder letter. H1 actual revenue was $981.3m. The sensitivity below shows the Q4 revenue required under three hypothetical Q3 Group-revenue observations.

Forward status: mechanical consistency test only. The report does not promote a Q3 revenue, net-income or EPS forecast.

Deal economics are improving on management's estimates, but the asset base is expanding rapidly. Cash, debt, contract liabilities, capex, D&A and prepayments must therefore remain separate analytical objects.

The five-year useful life reduces annual depreciation relative to the prior four-year policy; it does not change cash capex or prove utilization.

The market snapshot below is mechanical, not intrinsic value. It combines the 12 August share price with 271.855m basic shares outstanding as of 30 June, reported cash and debt, and end-June AI Cloud ARR; no multiple is selected as fair value.

Risks and catalysts enter a scenario only after the affected capacity, accounting line, timing and observable evidence are defined. Commitments and price reactions are not treated as realized revenue or causality.

SEC filings and issuer financial materials are the source of record for current actuals. Market data are supporting observations. Forward scenarios and formulas are analyst calculations.

Key limitations

• No Q3 point forecast is presented because timing and Group-versus-AI-Cloud scope do not reconcile.
• No approved annual capacity schedule or valuation assembly was available at the evidence cut.
• Contracted power, commitments, prepayments and ARR are not recognized revenue.
• Issuer visuals are credited at point of use; analytical charts and the digital twin are report-created.
• This report is information for discussion, not investment advice or an offer to transact.