Nebius Group
Demand is arriving. The buildout must earn its return.

Nasdaq NBIS · Nebius Group N.V. · Reporting currency USD · Research reviewed 21 October 2026. Financial baseline: unaudited Q2/H1 ended 30 June 2026. This page is a dated investigation, not a live quote service. Requested by the reader as the next company after Sivers; no unverified social-source attribution is implied.

Investment view

Operating momentum is compelling; a purchase valuation remains unproven. Nebius is a way to investigate the economics of supplying AI compute rather than choosing which AI application wins. The difficult question is whether scarce capacity today produces attractive returns after financing, hardware replacement and dilution.

Our stance is research watchlist, pending a current enterprise-value reconciliation and a credible fleet-level cash-return model. Large contracts reduce demand uncertainty, but introduce delivery obligations. A good customer and a large order do not eliminate construction risk or establish that shares are cheap.

What the business sells

Think of Nebius as a specialist AI utility with a software layer. It assembles GPU computing capacity, networking, storage and operating tools into infrastructure that customers can use to train models and run them in production. Training builds or improves a model; inference uses that model to answer a request. The company describes this platform in its investor materials.

The economic chain is power and sites → installed GPU clusters → usable computing service → customer billing → cash after reinvestment. Every arrow matters. Contracting electricity does not mean a data center is energized; installing GPUs does not prove utilization; invoicing does not settle the lifetime return on the equipment.

Our assessment: differentiation must show up in reliable performance, software convenience, utilization and customer retention. Merely buying the same chips as competitors is a weak long-term moat. The group also contains Avride and TripleTen and investments including ClickHouse and Toloka; valuing these requires separate evidence rather than treating the whole group as pure AI cloud. Group description

Financial evidence

USD millions; identical-period comparisons. Figures are consolidated, not the standalone AI cloud segment. Source: 12 August results, pages 1–2 and 6–8.

Calculated Q2 revenue growth: 454.0%. Calculated consolidated adjusted EBITDA margin: 40.6%. Q2 operating cash flow less the displayed capital purchases: −$3,411.3m; H1: −$3,626.2m. This subtraction is an analytical cash-flow proxy, excluding other investing flows, not a company-defined free-cash-flow measure.

At June end, cash was $8,042.1m and current plus non-current debt $8,545.7m; these are historical balance-sheet figures. Results, page 6

Why the growth is happening

Management's Q2 shareholder letter reports AI cloud revenue of approximately $575m, an AI cloud adjusted EBITDA margin of 50%, and an exit annualized revenue run-rate of $3.0bn. Its ARR definition takes the final month's AI cloud revenue and multiplies it by twelve. It is neither recognized annual revenue nor guaranteed contracted recurring revenue.

The letter reports four major cloud agreements averaging over $1bn in total contract value, including Reflection and Cohere; much of the associated capacity contributes primarily during 2027. It describes prepayments on about 70% of Q2 deals and forecasts over $9bn of customer prepayments in 2026. These are management disclosures and expectations, not independently audited fleet economics. It also raises year-end contracted-power guidance to 5GW. Contracted power must be distinguished from installed and revenue-generating capacity.

Our interpretation: prepayments are valuable financing, but cash received before delivery comes with an obligation to provide service. A surge in operating cash flow can therefore coexist with a large funding requirement. Track recognized revenue and delivery costs alongside cash collections, and distinguish customer-funded construction from cash that shareholders can distribute.

Customer commitments: read the terms

Sources: Microsoft 6-K and Meta announcement filed with the SEC. The Meta headline of up to $27bn is not one year's revenue. Neither agreement's total value can be added to ARR as if it were annual sales. Do not count third-party sales and the Meta backstop on the same capacity twice.

For underwriting, ask which milestones trigger billing, who bears cost overruns, what availability penalties apply and how concentrated delivery obligations are. A customer contract can help secure financing while still leaving the supplier exposed to execution losses.

Bull case and bear case

Our variant question is not whether AI needs compute. It is whether Nebius earns more on each incremental dollar invested than investors already expect, over an entire hardware cycle. That requires an expectations baseline as well as an operating thesis.

Financing and dilution

The 24 August closing release reports $5.75bn of new convertible notes: $3.45bn at 0.50% due 2030 and $2.30bn at 4.50% due 2034. Separately, $800m of older notes were exchanged for approximately 15.8m Class A shares. This combines fresh financing, retired debt and equity dilution; gross proceeds are not free equity value.

The 11 March NVIDIA transaction filing describes a $2bn pre-funded warrant covering 21,065,936 Class A shares, with a nominal exercise price. A warrant is economically relevant even before ordinary-share issuance. Reconcile its exercise status before deciding whether to add underlying shares to an updated count.

The results report 271,855,218 ordinary shares outstanding at 30 June, excluding treasury shares. Results, page 2 Our required bridge starts with that June share count, then reconciles subsequent ordinary issuances, debt exchanges, warrant exercises, employee awards and remaining convertible exposure. Avoid counting the same underlying shares twice. A June balance sheet plus an October price would mix dates; adding August proceeds to June cash would also ignore subsequent spending, fees and operating changes.

Valuation: what can be said responsibly

Our base case (23 Oct 2026, est.): about $97 a share. It assumes 2028 revenue of $15bn (consensus is $12.3bn for 2027), a 20% operating margin after depreciation, a 21% tax rate and 20 times after-tax operating profit. That gives an enterprise value of about $47bn in 2028, or $37.8bn today at a 12% yearly discount. Net debt is about $3.2bn: June debt of $8.55bn, plus $5.75bn of new convertibles, minus $0.8bn exchanged, less cash of about $10.3bn (June's $8.04bn, plus the raise, minus an estimated $3.5bn spent in Q3). Shares are 308.7m (271.9m in June, plus 15.8m from the exchange and 21.1m from the Nvidia warrant), with another 15% of dilution assumed to fund 2027–28 capex, so 355m. Every input here is our arithmetic, not company guidance. The price is most sensitive to the margin: at 30% the value is about $150, and at 10% about $45.

Fair value (23 Oct 2026): the mean of Morningstar's $150 (12 Aug), the S&P Global consensus of $283.6 (22 analysts, 22 Oct) and our $97.4 is $177.0. That is 11.2% above the 22 Oct close of $159.17, so the meter is green. LSEG's consensus was not captured because IBKR was not pulled this round.

Use enterprise value = equity value + debt + relevant lease obligations − available cash, with a consistent treatment of leases in both value and earnings. For convertibles, model the debt or converted-equity outcome consistently; adding full debt and all conversion shares simultaneously can double count the claim. Customer prepayments require separate delivery and working-capital modeling; they are not interchangeable with ordinary debt.

An illustrative sensitivity, not a forecast: on a hypothetical $10bn of annual revenue, each five percentage points of sustainable free-cash-flow margin changes annual cash generation by $500m. At an assumed 10% cash yield, that changes indicated enterprise value by $5bn. Neither input is an estimate for NBIS. The example shows why a revenue multiple alone cannot settle the investment decision.

The next model should separate existing-fleet cash generation, growth construction, replacement capital, customer advances and financing. Compare a delay case, a normal case and a strong-utilization case, then divide by dilution consistent with each case. Give non-core holdings an explicitly sourced value or leave them outside the core estimate.

Analyst views and evidence limits

Broker targets are opinions with dates, assumptions and coverage limits. Public aggregators surfaced different coverage counts and consensus figures during discovery. Original current broker notes were not verified, so this report does not publish a consensus, broker target or implied upside as validated evidence. This section is unverified, rather than a claim that no analysts cover NBIS.

The useful comparison, once original notes are available, is their assumed capacity timing, GPU utilization, capex funding and share dilution. A high target says little without that bridge. Company statements and our own interpretation are separately identified throughout this memo.

Recent developments and catalysts

The partnership, acquisition and pricing announcements are evidence of product and distribution activity. They do not independently establish incremental earnings. The next review checkpoint is 4 November 2026, an editorial reminder, not an announced earnings date.

What would change our view

We would become more constructive with on-time billable capacity, repeat customer demand and a demonstrated path to cash generation after equipment refresh, all at a valuation that compensates for financing risk. We would reduce conviction if deployment delays persist, pricing deteriorates faster than costs, or repeated funding leaves diluted per-share economics weaker despite revenue growth.

Before a buy conclusion, resolve current enterprise value, contract concentration, cash committed to construction, hardware useful-life sensitivity, conversion terms and a proper discounted cash-flow or cash-yield comparison. The operating evidence warrants attention. It does not yet establish a margin of safety.

Sources and information history

Original issuer downloads and retrieval hashes are preserved locally. The public source manifest records which downloads succeeded and which SEC archives were blocked locally but readable via the research browser. This is a bounded review of the linked results, agreements and releases through 21 October; it is not a claim to have audited every filing or contract.