Nebius (NBIS) stock analysis: full report, October 22, 2026

What the business is

Nebius chiefly rents GPU and CPU computing to AI developers and businesses. It connects those chips into clusters with networking and software, then bills for capacity and model requests. A GPU is a chip built to handle many calculations at once, which makes it central to training and running AI models. Token Factory charges by token, a small piece of text an AI model processes or produces, including requests to open-source models whose underlying design or code is publicly available. Toloka, Avride and TripleTen serve other markets; Toloka became an equity investment in 2025. Revenue was $529.8 million in 2025, but three customers supplied 59% of Q2 2026 revenue.

• An H100 GPU-hour was listed at $4.50. An H100 is a high-performance Nvidia chip designed for AI computing.
• Interruptible H100 spot capacity was listed at $0.67 per GPU-hour on October 22, 2026.
• Nebius spent $4.066 billion on property and equipment in 2025 before it could bill for that capacity.
• Strategic-customer advances totalled $982.5 million in 2025, funding deployment while creating future service obligations.

The challenges

• Turning power into billable capacity. Contracted power earns nothing until sites are connected, equipped, tested and serving customers. Nebius reported 170 MW of active power at December 2025. The stake is converting electricity and construction into customer-ready service.

• Earning enough from each GPU. Utilization alone does not prove attractive returns. Nebius spent $4.066 billion on property and equipment in 2025 against $529.8 million of revenue. If pricing or paid usage disappoints, each deployed GPU carries a heavier burden.

• Funding the buildout. Nebius’s real free cash flow was negative $3,747.3 million in 2025. Customer advances totalled $982.5 million in 2025, but they create future service obligations; debt adds repayment claims.

• Converting concentrated commitments. Three customers supplied 59% of Q2 2026 revenue. Delayed delivery, weaker collections or customer bargaining power could therefore hit sales and cash flow disproportionately.

The numbers

Revenue surged, but operating margin remained negative in 2025. Nebius’s real free cash flow stayed deeply negative as capital spending rose; net cash in 2024 became net debt in 2025, while real free cash flow per share worsened from negative $4.03 in 2024 to negative $15.13 in 2025. EV-to-revenue is 34.8x against a negative 7.9% free cash flow yield. P/E is not meaningful, and dividend and buyback yields are both 0.0%. Diluted EPS was $0.04 in 2025 despite a $611.7 million operating loss.

How it earns, and what can hurt it (US$ millions)

Operating margin — Equipment lost in transit cost $43.6 million; without that loss, the operating margin would be -107.2%.

What one share gets (US$ millions; per-share figures in US$; shares in millions)

EPS, diluted — With reported tax unchanged, removing the $598.9 million investment gain changes earnings per share from $0.04 to a $2.43 loss.

The verdict

Nebius could reach $536.43 per share in October 2029, a 213% return, if it commissions capacity, converts demand and funds the buildout without excessive dilution.

Management guides to $3.0–$3.4 billion of revenue and $7–$9 billion of ARR in 2026. ARR means annualized run-rate revenue, calculated as the latest month’s Nebius AI Cloud revenue multiplied by twelve. It also guides to 800–1,000 MW of connected power, a 40% adjusted EBITDA margin for Nebius AI Cloud and 5 GW of contracted power by year-end 2026.

The weighted forecast reaches $277.86 in October 2027, a 62% return, and $536.43 in October 2029. It assigns 40% to the bear case, 45% to the base case and 15% to the bull case. The October 2029 range is $91.15 to $1,405.50. Bear, base and bull EBITDA assumptions are $3.5 billion, $8.5 billion and $12.5 billion.

By December 2026, connected power and customer-ready capacity must match guidance. Early 2027 tests delivery and billing under the major customer agreement. Through 2027–2029, paid utilization, pricing, collections and cash flow per share decide whether the forecast holds.

Nebius is an AI infrastructure bet whose value depends on turning power commitments into paid, cash-generating GPU capacity.

Nebius Group N.V. is chiefly a business that rents the machinery and software needed to build and use artificial intelligence. Money comes in when customers reserve computing capacity or send model requests, and goes out to buy, power, connect and staff the machinery that makes those services possible.

1.1 — What it is and what it sells

Nebius mainly sells an integrated environment for artificial-intelligence work. Its main product, Nebius AI Cloud, combines rented computing, storage and software. Nebius Compute offers CPUs, which handle general-purpose tasks, and GPUs, which perform many calculations at once and suit AI workloads.

AI customers need more than a powerful chip. Training a model means repeatedly showing it examples and adjusting its internal settings. GPUs can perform many of those calculations simultaneously, while a large model may require many GPUs working together in a cluster. InfiniBand, a high-speed networking system, lets those machines exchange information quickly, so expensive processors spend less time waiting. Nebius therefore sells a coordinated system of processors, storage, networking, cooling and software, rather than a bare server.

Token Factory sells a more managed service. It lets a customer send requests to open-source AI models through an API, a standard way for one piece of software to ask another for a service. The customer’s application might send a question and receive an answer without operating the GPU servers underneath. Token Factory measures this work in tokens, small pieces of text used for processing, so the customer pays for model requests and responses rather than GPU-hours. Customers can use shared infrastructure or pay for dedicated capacity with more control over performance, configuration and location.

The rest of the group sells different products. Toloka prepares and evaluates data for AI developers, including human judgments about which model answers are better. Since May 2025, Nebius has held Toloka as an equity stake rather than combining Toloka’s sales and expenses line by line. Avride develops autonomous-vehicle and delivery-robot technology. TripleTen sells online technology courses and reskilling programs.

The segment rows show the relative scale of Nebius, Avride and TripleTen and the losses still associated with Avride’s development work.

Results by segment, as the company reports them (US$ millions)

That mix identifies the products. The next question is how customers pay for them and what Nebius must spend to deliver them.

1.2 — How it makes money

Nebius earns computing revenue by charging for access to capacity and resources. Customers can use on-demand service as needed, or reserve capacity for a longer commitment and more assured access. At the pricing page checked for this essay, an H100 GPU-hour was listed at $4.50 on demand. Storage and other resources carry their own charges.

Token Factory uses tokens as its meter because Nebius operates the model-serving machinery for the customer. Toloka charges for data projects and expert work, while TripleTen collects tuition from students. Those receipts belong to those businesses: owning part of Toloka does not turn every Toloka sale into Nebius revenue.

The main computing business requires large cash outlays before the first request arrives. Nebius buys GPU servers and networking equipment and develops or secures data-center capacity. It then pays for electricity, cooling, space, maintenance, engineers and customer support. NVIDIA, a major processor maker, is an important supplier because its chips underpin Nebius’s advertised GPU offering.

Adjusted EBITDA is not the same as profit or cash left over. The Nebius AI Cloud business reported $59.0 million of adjusted EBITDA in 2025, but that measure excludes depreciation, interest, taxes and specified adjustments. Nebius spent $4.066 billion on property and equipment in 2025, mainly GPUs and data-center hardware. A server purchase consumes cash immediately, while depreciation spreads its accounting cost across the server’s working life.

Customer timing also changes the picture. Strategic-customer advances totalled $982.5 million in 2025 and helped produce $401.9 million of operating cash inflow. An advance is cash received for a service Nebius still owes. The company records it as deferred revenue until it supplies that service. If a customer prepays for future computing, Nebius has the cash now but must still operate the equipment and deliver the promised capacity later.

So a dollar can arrive before Nebius has earned it, while other dollars leave long before the equipment generates revenue. The economics therefore depend on both the customer and the scale of capacity being built for that customer.

1.3 — Who pays

Nebius serves individual developers, businesses and large enterprises. A developer can begin through the online console or a model API, while a larger customer can negotiate capacity, support and deployment requirements. Customers stay when Nebius provides available capacity and reliable performance, but they can compare other providers.

The largest customers are hyperscalers, giant technology companies with enormous computing platforms of their own. Microsoft, a large technology company, contracted on September 8, 2025, for capacity valued at up to $17,392.9 million. Meta, another large technology company, signed an initial agreement dated November 1, 2025, covering approximately $2,880.7 million. These arrangements can anchor demand for years, but they also require Nebius to build capacity, secure financing and meet delivery commitments.

Meta expanded the relationship on March 16, 2026. The agreement includes approximately $12 billion of dedicated capacity and up to $15 billion of additional capacity over five years. The dedicated capacity is scheduled to begin delivery in early 2027. The additional capacity is first offered to other customers, with Meta buying what remains unsold. It is therefore a demand backstop, not $15 billion of revenue already delivered.

Capacity is scarce because Nebius needs specialised GPUs, suitable buildings, reliable electricity and fast networking at the same time. NVIDIA and other suppliers control important components, while utilities and regulators affect access to power and grid connections. Large customers can negotiate because they bring volume, but suppliers can also negotiate because replacement equipment is difficult to obtain. Prepayments may help Nebius finance construction, while making delivery obligations more demanding.

The other businesses have different payers. AI developers and companies buy Toloka’s data services. Uber, a ride and delivery platform, is a commercial counterparty for Avride’s vehicles and robots. Individuals buy TripleTen’s courses. The regional rows show where Nebius reports its customer revenue.

Revenue by region, as the company reports it (US$ millions)

Nebius is therefore a capacity business with concentrated, demanding customers: it must turn contracts into working infrastructure before it can turn them into earned revenue.

Nebius Group N.V. is building a business that rents the computing power AI companies need, then wraps that power in software and support. Its chance depends on a simple test: can it turn scarce electricity and costly GPUs into a reliable service that customers keep buying as supply grows?

2.1 — The market it sells into

Nebius sells access to powerful computers, mainly GPUs, plus storage, networking and software that manages the work.

A GPU, or graphics processing unit, performs many calculations at once, which makes it useful for teaching an AI model from examples. Instead of buying GPUs and building facilities around them, a customer rents access through Nebius. A small team can pay for actual use through the cloud console. A larger customer may reserve a cluster for months and pay for guaranteed access. Computing is priced in GPU-hours, while storage is charged separately because keeping data available also consumes equipment and electricity.

AI work has two main forms. Training teaches a model from examples; inference uses the trained model to produce an answer, image or other result. Medical-imaging company xAID, for example, uses Nebius to train models on CT scans, with one training pass lasting five days. Training creates a large, concentrated demand for machines, while inference can continue whenever users call the finished model. Nebius therefore needs both large clusters and dependable availability.

The potential market is large. A 2030 forecast puts the AI-cloud market at $267 billion. Specialist providers are forecast to take 20%, implying approximately $53.4 billion for them collectively. That is a measure of potential, not a precise addressable market, because forecasts differ over which infrastructure, platform and managed services they include.

The opportunity is real, but its value depends on how quickly Nebius can supply working capacity.

2.2 — Where the market is heading

Nebius’s opportunity depends on converting power commitments into customer-ready computers before capacity becomes abundant and prices weaken.

Training and inference create different rhythms of demand. Training may occupy a large group of machines for hours or days. Inference may involve smaller jobs that continue as people use an AI product. Nebius reported that production inference workloads on its Token Factory platform more than tripled during the second quarter of 2026. That shows growing activity on one service, though it does not by itself show equal growth in revenue.

Electricity is only the first ingredient. Nebius reported approximately 170 MW of active power at December 31, 2025. Active power means electricity being consumed by IT equipment. By August 2026, the company had raised its contracted-power target to 5 GW. Contracted power is a future supply pipeline, not operating capacity.

Between those two points lie construction, electrical installation, cooling, GPU delivery, high-speed networking, testing and customer acceptance. Only after those steps can capacity earn revenue. A megawatt therefore does not tell us how many GPUs are running, how many customers are paying or how much profit the service creates.

The economics can become harder even while demand rises. Better chips and more efficient models may reduce the computing needed for each task, while cheaper AI may stimulate more use. Meanwhile, a forecast for October 2026 said more than 80% of neocloud providers could cease independent operations by 2030 because of cost pressure and weak differentiation. Nebius must therefore measure delivery speed, prices and billable use, not just its power pipeline.

The company competes in two different markets at once: for scarce infrastructure and for customers who want dependable AI services.

2.3 — The rivals

Nebius faces specialist AI-cloud providers, giant general-purpose clouds and companies that supply the buildings in which computing takes place.

A neocloud is a specialist provider built mainly for demanding AI workloads. Nebius and CoreWeave, its closest specialist rival, combine GPUs, networking, storage and management software into an AI-focused service. A hyperscaler such as Microsoft Azure or Amazon Web Services offers AI alongside databases, security, business software and ordinary computing. A data-center landlord, such as Applied Digital, may provide powered space while another company operates the AI infrastructure inside it.

These roles overlap. Microsoft, Amazon, Google and Oracle can offer AI capacity to customers already using their broader cloud systems. Their scale also gives them more resources for infrastructure, engineering and customer acquisition. CoreWeave provides a closer specialist comparison because it focuses on making large AI clusters ready for production.

Nebius also supplies dedicated GPU infrastructure to Microsoft under a five-year agreement at Vineland, New Jersey. Nebius said in August 2026 that it had delivered all capacity tranches under that contract. The arrangement demonstrates that Nebius can deliver substantial infrastructure, but it does not show that Microsoft’s downstream customers are adopting Nebius’s own software platform.

The table compares CoreWeave, Inc.; Microsoft Corporation; Alphabet Inc.; Amazon.com, Inc.; Oracle Corporation; IREN Limited; Applied Digital Corporation; Core Scientific, Inc.; Lambda, Inc.; and Crusoe Energy Systems LLC.

The rivals

The important distinction is not simply who owns GPUs, but who can turn them into a useful service for the customer.

2.4 — What protects it

Nebius’s strongest protection is the engineering and operating capability that makes many separate components behave like one dependable computer.

A large AI job may spread across hundreds or thousands of machines. Fast networking lets them exchange information, storage keeps them supplied with data, and orchestration software assigns work and replaces failed machines. If one connection is slow or one machine repeatedly fails, the customer may pay for GPU-hours without completing the job. Nebius designs servers and racks in-house and integrates them with its cloud platform, so reproducing the service requires more than buying the same GPU.

Independent testing provides evidence of competence. In September 2026, an assessment placed Nebius in the highest ClusterMAX tier, Platinum, alongside CoreWeave, after examining live clusters, networking, storage, software and recovery from faults. The same assessment found CoreWeave ahead in several technical areas. Nebius has demonstrated a capable service, not an unassailable lead.

Facility efficiency may improve its cost position. Power Usage Effectiveness, or PUE, measures total data-center energy against the energy used by computing equipment. Nebius reported average PUE of 1.25 in 2025 against a stated global industry average of 1.54. For the same IT workload, those ratios imply about 19% less total facility energy.

That does not mean Nebius’s cloud is 19% cheaper or more profitable. GPUs, financing, labor, networking, maintenance and electricity prices also matter. The comparison has measurement limits, including excluded commissioning months and differences between company sites and the wider industry benchmark.

Nebius has built credible operating capability; the next question is whether customers value it enough to stay.

2.5 — Can Nebius keep an edge as AI cloud capacity grows?

Nebius has credible technical differentiation, but repeat business and paid utilization will decide whether that advantage survives plentiful supply.

A cluster is a group of connected computers working as one larger system. Orchestration software coordinates their jobs, while monitoring detects unhealthy equipment. This matters because the cheapest GPU-hour is not necessarily the cheapest way to finish a training run. Failures and slow connections can make a low listed price expensive in practice.

Nebius reported leading results for selected single-node B300 configurations in a June 2026 standardized training test. Those results show capability within defined workloads, not the lowest cost for every customer. Customer evidence is also encouraging: Higgsfield, a Nebius customer, increased usage more than twentyfold since its first contract. One successful customer cannot establish the retention rate of the whole business, but it shows how expansion might work.

Software may make the relationship harder to replace. Nebius offers managed Kubernetes, Soperator and managed inference alongside storage and computing. Kubernetes and Soperator are listed as free, so software adoption may improve retention and support a better computing price without becoming a separately billed software business. The key measure is gross margin: the share of revenue left after direct service costs.

Nebius is also testing ways to keep equipment productive. Utilization means the share of available equipment in use, but paid utilization asks a harder question: does that use earn enough to cover the cost? On October 21, 2026, Nebius announced a pilot spot pool in which customers can release idle GPUs for short-term rental and receive a rebate. That resembles selling an empty airline seat cheaply: some revenue is better than none, unless it displaces a customer willing to pay more.

Hyperscalers can combine AI capacity with existing relationships, while other specialists can improve their systems. Nebius’s durable edge will show itself when comparable customers choose it again because it completes work reliably at a competitive total cost.

Nebius Group N.V. sold the businesses that made its predecessor Yandex large and used the recovered capital to build an international AI-computing company. The result is a striking growth story whose returns remain unfinished: money has become GPUs, data centers, acquisitions and customer-funded expansion, but not yet proven lifetime profit.

3.1 — How it got here

Nebius is the international remnant of Yandex, not a continuation of Yandex’s old profit engine. Yandex began as a search engine in 1997, and its Dutch parent listed on the Nasdaq stock exchange in May 2011. Russia’s invasion of Ukraine halted trading on February 28, 2022, while the company negotiated a separation.

The decisive change came in 2024, when Yandex sold the businesses responsible for more than 95% of group revenue in the first nine months of 2023. It retained an international AI cloud, Toloka, Avride, TripleTen, a Finnish data center and minority investments. The parent adopted the Nebius name in August, and Nasdaq trading resumed on October 21.

This was a carve-out: like removing a department from a supermarket, the parent remained but its old accounts still contained businesses that no longer belonged to it. Nebius therefore prepared pro forma accounts, an accountant’s reconstruction of what the retained businesses might have looked like earlier. Those accounts showed a pro forma 2023 loss attributable to the parent of $325.8 million, not the audited result of a separately operating Nebius.

The old accounts also cannot tell us the recurring cost of independence. One-time transaction costs and temporary transition services do not establish what it costs to run the new group every year. That makes the next question less about old Yandex profits than about what management did with the recovered capital.

3.2 — The big decisions

Nebius’s first major capital decision was a constrained recovery from Russia. The announced transaction valuation was approximately $5.4 billion, but approval required at least a 50% discount to fair value. The closing announcement reported $2.8 billion of cash and 162.5 million parent shares returned to treasury. Shares were ownership claims, not cash available to build facilities.

The audited cash bridge is therefore more useful than the headline valuation. It recorded $2,642.3 million of cash consideration but removed $1,174.9 million of cash that belonged to the businesses sold. The cash-flow contribution from the disposal was consequently smaller than the apparent sale price.

Nebius then put the capital behind AI infrastructure. It sells access to powerful computers, mainly GPU clusters, so customers can train and run models without buying the machines themselves. Like a rental-car company, Nebius must buy and prepare the fleet before it earns revenue. A five-year agreement with Microsoft, a large software company, committed dedicated GPU infrastructure at Vineland, New Jersey, with an announced potential value of $17.4 billion.

The company also bought software and engineering capability. Tavily closed for $189.4 million of accounting purchase consideration, and ClarifAI followed for $97.4 million. Eigen AI closed for $331.4 million, while service-linked shares for its employees carried a $508.1 million fair value and vest over four years. These acquisitions can make the hardware more productive, but Q2 2026 revenue reached $582.3 million, up 454.0% from a year earlier, including acquired operations. The spending has produced growth; it has not yet proved the acquisitions’ returns.

3.3 — Where the cash went

Nebius has used its money to build productive capacity, so free cash flow was negative throughout the expansion period. The shortfall went into GPUs, servers, networking equipment, land, power connections and data-center construction. Those assets must be installed, accepted and kept busy before they earn an adequate return.

In December 2024, Nebius chose new equity over a potential buyback. It sold 33,333,334 shares at $21.00 each, raising about $700 million. That caused dilution: existing shares represented a smaller percentage of the enlarged company, but the company gained capital without an immediate repayment obligation. The bet works only if the new assets create more value than the ownership surrendered.

Debt then became a larger part of the funding structure. Net cash of $2,428.6 million in 2024 became net debt of $374.6 million in 2025. The company kept construction moving, but future operating returns now had to satisfy lenders as well as shareholders. December 2025 annualized revenue reached $1.25 billion, above management’s earlier target, but annualized revenue is not profit.

Customer advances supplied another source of cash in 2026. First-half operating cash inflow was $4,504.1 million, while investing outflow was $8,288.8 million. Much of the operating inflow came from deferred revenue: customers paid before Nebius delivered the promised computing service. Like a gym membership paid in advance, the cash arrives today but creates an obligation for later service. It helps fund construction, yet it is not free profit.

The table traces free cash flow through dividends, buybacks, acquisitions, disposals, investments, new shares and debt.

Follow the money: where the free cash flow went (US$ millions)

Nebius has therefore built a larger operating base by accepting dilution, debt and delivery obligations. To decide whether that financing worked, the results must next be separated by business and by share ownership.

3.4 — What the history and portfolio mean

Nebius’s growth can be judged only after separating the AI cloud from accounting changes and minority investments. Before the disposal, Yandex’s consolidated accounts included businesses later sold. Today’s retained group includes the cloud, Avride, TripleTen and other holdings, while later software acquisitions have joined the cloud operation.

Toloka shows why control matters. In 2025, outside investment reduced Nebius’s voting interest in Toloka to 49%. Nebius stopped consolidating Toloka, meaning it no longer reported Toloka’s sales and costs line by line. Instead, it used the equity method, recording its share of Toloka’s profit or loss. A change in reported revenue around that point could therefore reflect an accounting boundary rather than weaker customer demand.

The portfolio offers both promise and complication. Avride, the autonomous-driving business, received up to $375 million of combined investment and commercial commitments from Uber and Nebius. ClickHouse later received an outside financing at an approximately $15 billion valuation, and Nebius’s recorded stake rose to $1,517.7 million. These events support the existence of value, but they do not mean Nebius has sold the assets or stopped funding them.

The denominator changed too. The disposal returned shares to treasury, the placement issued new shares, and later financings and employee awards added further claims. Pro forma 2023 diluted shares were about 189.7 million, while the Q2 2026 basic weighted-average count was about 280.4 million. Growth in the company therefore does not automatically mean equal growth in value per share.

Nebius has built a much larger, better-funded platform, but the evidence so far proves expansion and outside validation more clearly than returns to each shareholder.

Nebius Group N.V. rents computing infrastructure for artificial-intelligence workloads, so its future depends on both technical execution and disciplined financing. Its leaders have delivered rapid growth, but founder voting control and the cost of expansion leave outside shareholders with limited influence and an important question: are management’s interests truly aligned with theirs?

4.1 — Who runs it

Nebius combines a technology founder with a finance chief whose job is to make expansion affordable.

Arkady Volozh has led Nebius since 2024, but his experience is much older than the company. He ran Yandex, the search and technology group from which Nebius emerged, between 2000 and 2022. His background in applied mathematics and computing matters because an AI-computing service needs its chips, buildings, networks and software to work together reliably.

The operating team divides that challenge. Roman Chernin leads the business, Andrey Korolenko oversees infrastructure and product, and Danila Shtan leads technology. Ophir Nave combines computer engineering with corporate-law experience, which helps when sites, equipment, contracts and financing must fit together.

Dado Alonso became chief financial officer in 2025 after more than 25 years in international finance. Her central problem is timing: Nebius pays for GPUs, buildings, electricity and engineers before customers generate the full return. In July 2026, it borrowed $775 million against deployed equipment and contracted customer cash flows. That asset-backed loan reduced the immediate need to issue shares, but it increased repayment obligations.

The business therefore has two linked risks. Scarce GPUs and power can delay capacity, while unused capacity earns little. Fully used capacity can be attractive, but only if customer payments cover infrastructure costs, interest and the risk of equipment losing value.

The officers table shows the named executives and directors by role and start date.

Who runs it

That balance between technical ambition and financing is also the board’s test, which makes oversight the next question.

4.2 — The board

Nebius has relevant outside expertise, but founder control limits how independent oversight can be in practice.

Chairman John Boynton provides continuity. Charles Ryan brings banking and infrastructure-financing experience, Kira Radinsky brings applied-AI expertise, Arne Grimme brings Dutch corporate-law and transaction experience, and Matthew Weigand brings venture-capital experience. Those skills suit a company that must assess both demanding technology plans and the financing behind them.

Formal independence is not the same as freedom from every relationship. Boynton’s long association with the company and Weigand’s connection to an investor in Nebius’s financing do not necessarily disqualify them. They do mean shareholders must ask whether a director will challenge an expensive proposal, not merely whether the director satisfies a rule.

Nebius reports a majority-independent board and independent audit and compensation committees. The nominating and governance committee is not entirely independent, however. Because one holder controls a majority of election votes, Nebius qualifies as a controlled company and can use exemptions from some independence requirements. Directors may therefore be formally independent while still depending on a founder-controlled voting block for their positions.

At the August 25 annual general meeting, shareholders reappointed the directors and renewed authorities to issue shares and repurchase them. Those permissions preserve financing flexibility; they do not show that shares were issued or that cash was returned.

The board table shows each director’s start date and independence status.

The board

But formal oversight cannot be separated from who can choose directors, so control begins with ownership.

4.3 — Who owns it

Nebius separates economic ownership from voting control, giving Volozh’s family trust decisive influence over elections.

Class A shares carry one vote each. Class B shares carry ten. LASTAR, Volozh’s family trust, controls about 52% of the votes while representing about 11% of the company’s economic interest. Outside investors may therefore own most of the financial value yet remain unable to replace the board against the trust’s wishes.

Volozh and other executives also hold meaningful financial stakes. That gives management a reason to care about the long-term share price, but it does not remove the control gap. The founder-linked block can determine elections while bearing a much smaller share of the company’s economic exposure.

Institutional ownership requires careful reading. NVIDIA, an AI-chip supplier, owns a disclosed Class A position, while Goldman Sachs’s reporting entities describe a shared position rather than two stakes to add together. Situational Awareness and its affiliated entities and principals likewise report one shared position. Beneficial ownership can mean shared authority to vote or sell; it does not mean every named person owns a separate pile of shares.

On March 11, 2026, NVIDIA invested approximately $2 billion through a pre-funded warrant for Class A shares. A warrant is a right to obtain shares, and “pre-funded” means most of the purchase price has already been paid. The investment can fund infrastructure and align NVIDIA with Nebius’s growth, but issuing the shares can dilute existing owners. A supplier’s success and a minority shareholder’s return are related, not identical.

The owners table identifies insiders and principal holders, their share classes, stated value, recent transactions and reporting date.

Who owns it: insiders and principal holders

That control structure also changes how to read insider transactions, which come next.

4.4 — Insider buying and selling

Insiders sold heavily, but the reason for each sale matters more than the headline total.

In the stated twelve-month period, officers and directors sold 1,603,247 shares for US$329,741,619.56 in open-market transactions. Selling can signal reduced confidence, but it can also reflect taxes or a planned reduction in a concentrated investment.

Volozh’s October 1, 2026 filing shows the clearest example of a sale that says little about his outlook. When restricted share units vested, some shares were sold automatically to cover estimated withholding taxes. He received the award, incurred the tax obligation and sold only enough shares to meet it. The filing said the transaction was not discretionary.

Ophir Nave’s October 5 sale followed a Rule 10b5-1 plan adopted on May 22, 2026. Such a plan sets trading instructions in advance, so the execution date is less informative as a fresh judgment. It still reduced his exposure, but the plan explains its timing.

Arne Grimme’s Form 4 records an October 19, 2026 purchase under a previously adopted plan. That prevents a simple conclusion that insiders only sold. The register must be read transaction by transaction, including its footnotes and retained awards.

These transactions are easier to judge once the pay system and operating promises are clear.

4.5 — Pay, and promises kept

Nebius’s incentives encourage retention and a rising share price, while its execution record still must be judged per share.

Senior management receives cash and equity compensation. Options give executives the right to buy shares at a fixed price; restricted share units deliver shares when their conditions are met. Both can align management with shareholders, but time-based vesting mainly rewards staying and does not require a particular return on infrastructure spending.

The strongest hurdle was a premium-priced option award approved in April 2025. The $100 exercise price meant executives needed substantial share appreciation before the options became attractive. Volozh declined his award. The arrangement rewards a higher share price, but it does not directly measure cash generation or returns on invested capital.

Stock compensation cost increased in 2025, even as its share of revenue fell because the business grew. Year-end shares outstanding also increased between 2024 and 2025, and Nebius spent nothing on buybacks in 2025. The average share count used for earnings and the year-end ownership denominator answer different questions, so a lower average does not prove dilution disappeared.

Nebius has nevertheless delivered important operating promises. In February 2025, Volozh reaffirmed a December annualized run-rate target of $750 million to $1 billion. Nebius finished 2025 at $1.25 billion. By August 2026, it had also delivered all Microsoft capacity tranches due by then. Annualized run-rate revenue shows the pace of a period, not guaranteed future revenue, profit or cash.

The stock-pay table compares compensation, cash-flow effects and share counts across the listed years.

Stock pay and the share count (US$ millions; per-share figures in US$; shares in millions)

Nebius’s leaders are partly on a shareholder’s side: they own meaningful stakes and have executed strongly, but founder voting control, debt, equity awards and possible dilution mean outside owners still carry risks they cannot fully control.

Nebius Group N.V. rents powerful computing infrastructure, mainly GPUs, to customers building and running artificial-intelligence systems. Its accounts show a company growing at exceptional speed, while still asking whether demand can become durable profits and cash.

5.1 — Growth

Nebius is growing by putting scarce AI computing capacity into service. It rents access to GPUs much as a workshop rents equipment: a GPU-hour is one GPU available for one hour, but only customer-paid use produces billable work. Capacity is scarce because Nebius must secure chips, electricity, buildings, cooling, networking and financing before it can sell the service.

Continuing revenue multiplied between 2023 and 2025, led mainly by the cloud business. But the comparison changed when Nebius’s voting interest in Toloka fell to 49% on May 2, 2025, so Toloka stopped being consolidated. Earlier periods were recast as well. The reported growth rates therefore describe a changed reporting boundary, not a perfectly continuous business.

ARR, or annualized run-rate revenue, answers a different question. It multiplies the latest month’s AI-cloud revenue by twelve: ARR rose from $90 million in 2024 to $1,250.0 million in 2025, then reached $3.0 billion by June 2026. It is a speedometer, not a promise. A new cluster or customer ramp can lift it before the same pace becomes recognized revenue and collected cash.

Active power and contracted sites support the expansion story, but contracted capacity, installed equipment and revenue-producing service are separate stages. The specialist therefore watches whether new machines become paid, profitable workloads.

The results table shows the company’s reported revenue, profits, margins, per-share figures and dividends across the five years.

Results (US$ millions; per-share figures in US$)

Net income — With reported tax unchanged, removing $598.9 million of investment gains changes the $9.8 million profit into a $589.1 million loss.

The company’s own operating measures show how quickly ARR, active power and planned capacity have changed.

The company's own measures (US$ millions)

5.2 — Profit

Nebius’s operating economics are improving, but reported profit is not yet proof that its infrastructure earns an adequate return.

Gross margin improved sharply, yet depreciation and amortization sit below cost of revenue. That means gross profit excludes much of the economic cost of supplying computing capacity. A $100 server creates a recurring expense as it wears out, even though the cash was spent when Nebius bought it. EBITDA can show whether operations are improving before depreciation, interest and tax, but it cannot show whether the servers repay their cost.

Reported net income was lifted by investment revaluations. ClickHouse is a private software company in which Nebius owns an investment; when outside investors bought shares at a higher price, Nebius increased the recorded value of its holding. The 2025 gain was $597.4 million. No customer paid Nebius for that increase, and Nebius did not sell the investment.

The same issue became larger in 2026. The first-half ClickHouse gain was $780.6 million, so reported profit could coexist with an infrastructure business still losing money. These are noncash gains, not funds available to build data centres.

Restructuring, equipment lost in transit, acquisition costs and tax adjustments also move earnings. The 2026 change from four-year to five-year equipment lives reduced depreciation, but that was an estimate change rather than a cash saving. The useful question is therefore whether operating profit survives full equipment costs and recurring investment.

The table separates Nebius’s adjusted operating measures from its reported tax rate.

The company's own adjusted measures, and tax (US$ millions; per-share figures in US$)

The next test is cash: whether improving operations fund the infrastructure they require.

5.3 — Cash

Nebius produced operating cash in 2025, but expansion consumed far more cash than operations generated.

Working capital supplied much of the improvement. Working capital means short-term operating balances such as customer advances, receivables and supplier bills; its contribution was $363.4 million in 2025. Strategic customer agreements brought in $982.5 million during that year. A customer may pay ahead for dedicated capacity, giving Nebius cash before the service is delivered. The receipt is helpful, but it also creates an obligation.

Capital spending then absorbed the operating inflow. Nebius bought GPUs, buildings and networking on a scale far beyond current operating cash, leaving free cash flow deeply negative. Adding stock compensation to the analysis makes the owner’s economic shortfall larger, because employees receive part of the business even though no cash leaves at that moment.

Positive operating cash therefore does not yet demonstrate a self-funding cloud business. It may partly reflect the next wave of customer prepayments financing the next wave of construction. The harder evidence will be cash earned by a stable fleet after expansion slows, including the cost of repairs and replacement.

The cash-flow table shows operating cash, working-capital effects, capital spending, free cash flow and the stock-pay adjustment.

Follow the money: real free cash flow (US$ millions)

5.4 — Balance sheet and debt

Nebius has substantial liquidity, but its fixed commitments have expanded with its equipment base.

Cash and short-term investments provide near-term protection, while the current ratio measures current assets against current liabilities. That ratio remains strong but has fallen sharply, which means obligations are growing faster than the old cushion. Total assets also mix operating equipment, construction assets, leased-facility rights and private-company investments; these do not have equal ability to produce or release cash.

Convertible notes add both borrowing and possible dilution. Their stated coupons range from 1.00% to 3.00%, but the 2029 notes carry a 7.06% effective rate after accretion and issuance costs. Accretion means the recorded obligation grows toward the amount payable. Conversion may avoid repayment in cash, but it can issue more shares to lenders.

Leases create another future claim. Recognized lease liabilities were $845.4 million at year-end 2025, while signed leases not yet started represented about $9.76 billion of undiscounted payments. Those payments were expected to begin in 2026–27, not immediately. They nevertheless matter because rent and financing continue even if capacity sits idle.

The balance sheet table shows the company’s liquidity, liabilities, debt, equity and calculated coverage measures.

Balance sheet (US$ millions)

The specialist matches debt maturities and lease commencements with customer receipts, not just with cash on hand today.

5.5 — The latest quarter

The June 2026 quarter showed a larger and more efficient revenue engine, but not yet self-funding growth.

Revenue continued to accelerate, supported by capacity added earlier, better utilization, newer services and acquisitions. The eight-quarter record shows repeated expansion rather than one isolated surge. Gross margin and operating margin improved because revenue grew faster than operating expenses.

The absolute operating loss still widened. Acquisition and corporate transaction costs were $49.9 million in the June 2026 quarter. The longer equipment lives also reduced depreciation by $43.0 million. Removing acquisition and corporate transaction costs improves operating income, while removing the depreciation benefit worsens it; neither adjustment proves that the deployed fleet earns its full cost.

Cash from operations was positive while capital spending remained much larger, so quarterly free cash flow stayed deeply negative. Customer funding and infrastructure deployment therefore continue to move on a greater scale than recognized sales.

Reporting quality also matters. The December 2025 controls assessment identified material weaknesses in fixed-asset accounting and TripleTen revenue recognition. Published 2025 quarterly operating losses totalled $596.2 million, while the audited annual operating loss was $611.7 million. A specialist uses the audited figure and treats the difference as unresolved rather than inventing a fourth-quarter result.

The table gives the revenue, margins, earnings, cash flow and capital spending for the last eight quarters.

The last eight quarters (US$ millions; per-share figures in US$)

Operating income — Acquisition and corporate transaction costs were $49.9 million; without them, the operating loss would be $126.0 million.

The latest quarter confirms operating progress, but the next question is what each deployed GPU actually earns.

5.6 — What each deployed GPU earns

A GPU creates value only when paid work covers its operating costs and eventually repays the capital installed around it.

A GPU-hour measures availability, but billable GPU-hours measure customer-paid use. A machine may be installed, reserved, technically busy or performing paid work; those states are not interchangeable. Reserved capacity can earn a fee even when the customer leaves it idle, while busy equipment may still earn poor returns if prices are discounted.

Megawatts measure electrical capacity, not computing sold. Nebius distinguishes contracted, connected and active power, with active power showing electricity used by IT equipment. A proper return calculation matches revenue and gross profit with average active power or billable hours over the same period. It also separates GPU generations, because newer processors may deliver more work per watt while making older equipment less attractive.

The operating bill includes electricity, cooling, networking, maintenance and staff. Colocation adds payments for space and facility services in another data centre. Managed inference—running trained models to answer requests—may improve returns by selling a service rather than bare machine time.

Nebius’s Q2 payback estimate was one year and ten months, but it remains a forecast based partly on future capacity. The specialist tests it against deployment dates, realized prices, operating costs and cash recovered after service begins.

5.7 — How durable and diversified customer commitments are

Customer commitments make growth more visible, but their value depends on delivery, collection and concentration.

Backlog estimates future business under agreements. Remaining performance obligations, or RPO, are contracted services not yet delivered. Deferred revenue is cash or consideration received before the related service becomes revenue. A prepaid restaurant meal plan illustrates the sequence: the signed plan is a commitment, the cash may be deferred revenue, and each meal becomes revenue only when served.

RPO was $37,490.6 million at June 2026. Nebius expected 36% to become revenue during the following twenty-four months. It expected another 40% during months twenty-five to forty-eight. The figure excludes contracts lasting one year or less and includes estimates of variable consideration, so it is not a near-term sales forecast.

Prepayments provide stronger evidence than an unfunded expression of interest. For some large Q2 deals, they were expected to cover 50–60% of associated capital spending. That reduces near-term financing pressure, but the customer has purchased future service, so Nebius still owes delivery and may face remedies for delay or failure.

Customer dependence remains material. Three customers supplied 59% of Q2 revenue. At June-end, two customers represented 16% and 29% of net receivables. A larger customer roster helps only if revenue, receivables, prepayments and renewals become less concentrated.

5.8 — How long each GPU generation can earn

The crucial equipment-life question is how long each GPU keeps earning an adequate return, not how many years the accounting policy assigns it.

Depreciation spreads an equipment purchase across its expected useful life. A server bought for $100 creates less annual depreciation when management extends that life, but the original cash outlay does not disappear. In January 2026, Nebius extended server and network equipment lives from four years to five.

During the first half of 2026, that change reduced depreciation by $86.1 million and increased net income by $75.7 million after tax. The benefit is prospective and noncash. It should therefore be separated from both operating improvement and investment revaluation gains.

There is a commercial case for longer lives: older GPUs may still serve inference or less demanding workloads after newer chips become preferred for frontier-model training. But falling prices, better performance per watt and customer migration can make an older fleet unattractive before its accounting life ends.

The decisive evidence is generation by generation: paid hours, realized prices, power use, repair costs, redeployment, retirement age, replacement spending and resale value. Longer accounting lives improve reported earnings; longer commercial lives improve returns. Only the second can justify continued expansion.

Nebius Group N.V. sells access to powerful computing infrastructure used for artificial intelligence. Its stock price asks investors to believe that this young business can turn rapid expansion into durable cash generation.

6.1 — What an investor pays today

At US$171.57 a share on October 22, 2026, Nebius is priced for a much larger cash-producing business than the one operating today.

The stock market values shareholders’ ownership at US$46,642.2 million, while enterprise value is US$47,145.8 million. Enterprise value adds debt and subtracts cash, so it estimates the cost of buying the operating business itself. The distinction matters because Nebius must earn returns for lenders as well as shareholders, while its cash may already support construction. Investors are paying 34.8 times last-twelve-month revenue.

Profit measures are harder to read. EBITDA means earnings before interest, taxes, depreciation and amortization. Nebius reported US$236.2 million of adjusted EBITDA alongside a US$175.9 million operating loss in Q2 2026. A US$100 GPU system that lasts five years might produce US$40 of revenue and cost US$10 to operate, leaving US$30 of EBITDA; after US$20 of annual depreciation, only US$10 remains as operating profit. EBITDA therefore does not remove the eventual cash cost of replacing equipment.

Customer prepayments reduce the need for outside financing, but they are not profit. In Q2 2026, roughly 70% of deals included prepayments covering 50–60% of associated capital expenditure. Like a builder receiving money before completing an extension, Nebius can use the cash to build capacity but must still deliver the service. Its −7.9% free cash flow yield shows that current cash generation does not support the valuation by itself.

The price also implies value beyond the accounting assets. Nebius trades at 4.5 times book value, while it extended server and networking lives from four years to five years beginning in 2026. That lowers reported depreciation without improving the machines’ cash earnings. Concentrated control adds another consideration: Arkady Volozh’s family trust held about 52% of voting power through an approximately 11% economic interest as of March 31, 2026.

Look at the current valuation measures together, especially the revenue multiple, book value and cash-flow yields.

What an investor pays today (US$ millions; per-share figures in US$)

The price is therefore a claim on future execution, which becomes clearer when set against Nebius’s own history.

6.2 — Against its own past

Nebius became more valuable because investors began pricing future execution, not because established earnings suddenly justified the price.

The year-end share price rose from US$27.70 in 2024 to US$83.71 in 2025, while price-to-book rose from 2.0 times to 4.6 times. Investors paid more for each dollar of recorded equity as confidence in AI infrastructure grew. Book equity can also rise when a company raises money and buys equipment, so a higher share price does not by itself prove that the valuation multiple expanded.

The September 2025 Microsoft agreement gave that confidence a concrete foundation. Microsoft contracted for dedicated GPU infrastructure at Vineland, New Jersey, with about US$17.4 billion of value through 2031 and potential growth to about US$19.4 billion. But a contract becomes revenue only after Nebius secures equipment, power and space, installs the network and clusters, and delivers usable capacity.

That capacity is scarce because GPUs, electricity, suitable data centers and financing are limited. Nebius said urgent customers sometimes pay US$40–50 million per megawatt, compared with US$20–25 million for mid-term deals. Its AI cloud revenue grew 514% year over year in Q2 2026 to US$574.9 million, but a small starting base makes rapid percentage growth easier.

Expansion also changes each shareholder’s claim. Nebius issued 12.7 million Class A shares in Q2 2026 for approximately US$2.8 billion. The money can finance construction, but each existing share then represents a smaller fraction of the company. The relevant question is whether future cash generation grows faster than the share count.

The history table shows how price, earnings and book value changed across the earlier years.

Five years of history (US$ millions; per-share figures in US$; shares in millions)

Its past explains the optimism; rivals show how demanding that optimism has become.

6.3 — Against rivals

Nebius commands a demanding valuation beside both GPU-cloud rivals and profitable technology companies.

CoreWeave is the closest listed comparison because it also sells AI computing infrastructure. Its US$34.20 billion market value divided by US$5,131 million of 2025 revenue gives approximately 6.7 times sales. Nebius has a larger equity value despite a much smaller recent revenue base, while its enterprise-value-to-sales ratio is 34.8 times. The comparison is not perfectly matched, but it shows how much expansion Nebius investors are buying.

Microsoft, Alphabet, Amazon and Oracle provide a different benchmark. Their whole-company P/E ratios range from Alphabet’s 16.71 times to Microsoft’s 29.62 times, while their net margins range from Amazon’s 10.8% to Microsoft’s 40.3%. These companies combine cloud with other businesses, so their multiples cannot simply be applied to Nebius. They do show what investors pay for demonstrated profits.

Nebius’s 514% AI-cloud revenue growth in Q2 2026 is powerful evidence of demand, but growth alone does not settle the comparison. Investors must ask whether signed demand becomes working capacity, paying customers and collected revenue at acceptable cost.

Margins require the same discipline. Nebius reports depreciation separately from cost of revenues, so its gross margin may look better than a rival that includes equipment costs directly. Operating margin adds wider business costs, while net margin also reflects financing, taxes and investment effects. The final comparison is cash left after power, operations, equipment replacement and expansion.

The table puts the public and private rivals beside Nebius across revenue, margins, market value and P/E.

The rivals

Those comparisons explain why the stock can rise on good news and still fall when the market changes the price it will pay for growth.

6.4 — How the price moved

Nebius’s stock has delivered a strong annual gain while repeatedly changing its mind about the value of future success.

At US$171.57 on October 22, 2026, the share price was approximately 42.8% below its 52-week high of US$299.86. It had gained 74.0% over one year but fallen 27.3% over one month. These returns are not contradictory: different starting dates produce different stories.

The Microsoft agreement was a commercial catalyst because it replaced imagined demand with a named customer and a large commitment. It also supported the idea that financing could follow demand. But investors still had to estimate the cost and timing of building the promised infrastructure.

Nasdaq-100 inclusion created a separate trading catalyst. Nebius entered the index effective June 22, 2026, and funds tracking the index had to adjust their holdings. Such purchases can increase demand regardless of whether a fund manager believes the shares are cheap. They explain trading flows, not underlying value.

The August results brought another burst of attention, with contemporary coverage reporting a 34.1% rise on August 12. Later weakness reflected concerns about AI valuations and data-center expansion costs. The precise cause of each trade is unknowable, but the larger lesson is clear: the business can improve while investors reduce the multiple they will pay for distant cash flows.

The price and return tables separate the stock’s current level from its performance over different periods.

The share price today and over 52 weeks (US$ per share)

Share price performance

The central valuation question is therefore not whether Nebius is growing, but what that growth must eventually earn.

6.5 — What the price implies

Today’s valuation requires Nebius to turn its construction program into several billion dollars of durable annual cash generation.

Suppose a mature Nebius were valued at 20 times annual cash flow available to all capital providers after operating costs, taxes and necessary investment. Its US$47,145.8 million enterprise value would then require approximately US$2.36 billion of annual cash flow. At an assumed 20% cash margin, that would require approximately US$11.8 billion of annual revenue. These are illustrations, but they reveal the scale embedded in the price.

Timing raises the hurdle. If that value arrived three years from now and an investor required a 10% annual return, enterprise value would need to reach approximately US$62.8 billion. At the same assumptions, annual revenue would need to approach US$15.7 billion. That is far beyond management’s guidance for US$3.0–3.4 billion of recognized group revenue in 2026.

Annualized run-rate revenue is only a pace, not completed annual sales. Nebius reported US$582 million of group revenue in Q2 2026 and a US$3 billion annualized run rate at June’s end. Multiplying a current pace by twelve assumes it continues; it does not mean the company already earned that amount.

Management’s estimated one-year-and-ten-month payback on Q2 deals supports the possibility of attractive economics, but it depends on forecast costs and future capacity, including sites not yet built. The final test is return after reinvestment. Nebius must replace GPUs, serve lenders, absorb dilution and still produce rising cash flow per diluted share. Its 0.0% dividend yield means investors receive no current cash income while waiting. The stock price assumes that future is not merely possible, but successfully built and paid for.

Nebius Group N.V. is trying to turn scarce electricity and expensive AI hardware into a fast-growing cloud business. The central question is whether its promised capacity can become billable service quickly enough, and whether customer cash and financing can carry the buildout until it does.

7.1 — What management has guided

Management’s plan depends on turning infrastructure commitments into customer-ready computing.

Revenue records services delivered during a period. ARR, or annualized run-rate revenue, takes the latest month’s revenue and multiplies it by twelve. A bakery opening a huge counter near year-end might show little annual sales but a very large annual pace in its final month. Nebius faces the same effect: a cluster that begins serving customers late can lift ARR while contributing little revenue during that year.

Power commitments show a similar gap. Contracted power secures a future place to build and an electricity commitment; connected power means electricity physically reaches the site. Neither proves that GPUs, cooling, networking and software are ready for customers. The guidance table therefore matters most when read as a sequence from future power to usable capacity.

Nebius’s AI cloud is rented GPU computing for training and running AI models. Management pairs its growth plan with a 40.0% adjusted EBITDA margin target and customer prepayments. Adjusted EBITDA describes operating performance but excludes interest, taxes, depreciation, amortization and specified adjustments. Prepayments bring cash forward, but they also create an obligation to deliver computing later.

Site counts offer less evidence than commissioning dates, capacity and customer start dates. The table sets out management’s guided periods, ranges and targets for revenue, margins, power, prepayments, sites and deployment.

Guidance in force (US$ billions; non-currency units stated in rows)

That is why the next question is not whether the targets sound large, but when analysts think the capacity can earn them.

7.2 — What analysts expect

Analysts broadly accept the near-term revenue story but disagree about the speed of the following ramp.

The 2026 consensus was $3.36 billion, based on six estimates. The 2027 consensus was $12.02 billion, based on seven estimates. Those estimates ranged from $8.50 billion to $17.99 billion. This is a small sample, but its spread is economically important.

The disagreement is mainly about timing. Capacity serving customers late in one year can contribute a full year of revenue afterward. Capacity completed late in the next year contributes much less. Analysts may agree that customers want more computing while reaching very different conclusions about construction dates, occupancy and prices.

Recent results offer mixed evidence. Nebius recorded three earnings beats and one miss across the last four reported comparisons. Fourth-quarter 2025 revenue was $227.7 million against approximately $246.05 million expected. First-quarter 2026 revenue was $399.0 million against approximately $375.13 million expected. The pattern is not yet a dependable habit of beating expectations.

The useful question for analysts is therefore simple: how many megawatts serve customers each quarter, at what revenue per megawatt and with what utilization? That bridge leads directly to the spending required to make the forecasts possible.

7.3 — The spending ahead

Nebius must spend heavily before the resulting computing can earn revenue.

Capex means spending on long-lived assets. Here it includes GPUs, servers, networking, cooling and data-center construction. Management maintained a 2026 spending plan of $20–$25 billion, while suppliers and builders must often be paid before customers begin using the machines.

First-half capex reached $8,130.3 million. Operating cash flow was $4,504.1 million. Deferred revenue—customer cash received before the related service is delivered—contributed $4,395.0 million to that cash flow. Even with that help, the first-half shortfall after capex was $3,626.2 million.

Customer advances therefore reduce the funding gap, but they do not remove it. A customer paying $50 toward a $100 cluster gives Nebius useful cash, while Nebius still must buy equipment and deliver the remaining service. The undelivered amount remains a promise, not profit.

Management estimated a one-year-and-ten-month payback for certain large deals and said prepayments could cover 50–60% of associated capex. That is a forecast for selected future capacity, not proof that the whole estate produces cash after replacing equipment. The decisive test is whether each spending wave produces active capacity and collections on schedule.

7.4 — Can they pay for it

Nebius has substantial funding access, but its construction pace makes liquidity a timing problem.

Using June 2026 balances, Nebius had net debt rather than net cash. Cash rose through March 2026 and then fell by June, while first-half free cash flow remained negative. The table shows the underlying cash, investments and debt movements.

The first-quarter shortfall after capex was about $214.9 million. The second-quarter shortfall widened to $3,411.3 million. Reaching the annual capex plan would require another $11,869.7–$16,869.7 million in the second half, so June cash alone cannot fund the entire buildout.

Nebius issued $4,337.5 million of convertible principal on March 20, 2026. It later raised $5.75 billion through new convertibles on August 24, 2026. Convertibles provide cash now but may become shares later, diluting existing owners. A separate $800 million exchange for approximately 15.8 million shares reduced debt but supplied no new cash. The August notes mature in 2030 and 2034, giving the company time but leaving future repayment or dilution.

A new secured bank facility adds nearer-term pressure. It has a 364-day maturity. Interest starts at Term SOFR plus 2.75%, with the margin able to rise to 3.50%; SOFR was 3.88% on October 21, 2026. Access to a facility is not the same as drawn cash, and a short-term loan used for long-lived assets may need refinancing before those assets repay their cost.

The first table shows cash, short-term investments, debt and the resulting net cash or net debt across eight quarters.

Cash and debt, last eight quarters (US$ millions)

The second table shows the amounts of debt coming due across the stated maturity periods.

Debt coming due after the latest fiscal year end (US$ millions)

The funding question is therefore whether cash arrives before construction bills and debt maturities.

7.5 — What to watch

The next useful evidence will be the alignment of commissioning, customer receipts and financing.

Through December 31, 2026, investors should compare actual service starts and collections with the guidance. A strong exit ARR without a clear account of active capacity would leave questions about durability. Large advances without matching delivery would instead show a growing service obligation.

Early 2027 should show whether capacity supporting the second agreement with a major customer has begun service and billing. Nebius must also address the short-term facility well before its maturity. Longer-term funding against performing customer contracts would be stronger evidence than repeated short extensions.

Spot pricing offers a nearer commercial test. Nebius now sells interruptible GPU capacity at live prices, alongside short reserve auctions and pay-as-you-go access. A listed H100 spot price was $0.67 per GPU-hour on October 22, 2026. Spot sales can monetize idle machines, but falling prices or weak occupancy would signal weaker scarcity.

Exceptional items should remain separate from recurring cash generation. Disposal-related Dutch withholding tax was $181.5 million after the February 2025 declarations. At December 2025, Nebius recognized $3.3 million for litigation and disclosed $43.6 million of equipment lost in transit. None should become an assumed future receipt.

The deeper issue is whether power and construction can become dependable billable capacity.

7.6 — Can power and construction become billable capacity?

Nebius earns a return only when electricity, buildings, hardware and customers are ready together.

A typical customer-ready cluster may take 2–4 months after power becomes available at around 1,000 GPUs. The sequence starts with enforceable access to the site and electricity. Then engineers install distribution, cooling and networking; GPUs arrive and pass integrated testing; finally, customers run workloads under the agreed service conditions.

Vineland, New Jersey, illustrates the dependencies. Nebius moved the project to on-site fuel cells, while the building had already been completed and engineering fit-out was progressing. An amended site layout still required approval. Solving one power bottleneck can therefore create new engineering and permitting work.

Nebius should keep three scorecards: contracted power, connected power and active capacity serving customers. Securing power is valuable because it protects future expansion, but only active capacity produces bills. The best evidence would reconcile each site’s megawatts, service dates, budget, final spending and elapsed time from electricity availability to customer use.

That site-by-site record will determine whether the funding plan is supporting production or merely postponing the test.

7.7 — Can the funding plan support the buildout?

The funding plan works only if customer cash, construction and debt maturities stay in step.

Before a new cluster earns its first bill, Nebius must fund GPUs, servers, cooling, networks, construction, deposits and working capital. Customer advances reduce that burden only when they are actually available. A valuable contract can still leave a large interval that Nebius must finance itself.

Each funding source makes a different claim on the eventual return. Customer advances create future service obligations. Secured lenders may claim equipment and contract receipts first. Share issuance brings permanent capital but spreads future earnings across more shares. Convertibles combine debt with possible dilution.

Nebius’s asset-light model changes that equation. Partners finance and own facilities and hardware, while Nebius provides systems architecture, cloud software and sales. It resembles a hotel brand operating from an investor-owned building: less capital leaves Nebius, but more execution depends on the partner, and the economics must be shared.

The final test is a delayed-build and slower-collections case shown quarter by quarter. Nebius must demonstrate enough cash, financing headroom and operating flexibility to survive postponed billing while interest, rent and committed power costs continue. If it can do that, its secured power and customer demand may support durable growth; if not, the plan may produce capacity faster than it produces returns.

The forecast gives NBIS substantial upside, but execution risk makes the range unusually wide.

The weighted price is $277.86 in October 2027, a 62% return including dividends. It rises to $427.22 in October 2028, a 149% return. By October 2029, the weighted price reaches $536.43, a 213% return.

The October 2029 range runs from $91.15 in the bear case to $1,405.50 in the bull case. Today's reference price is $171.57.

What the shares could be worth (USD per share at each date; in brackets the return since today's 171.57, dividends included)

8.2 — The model

The forecast prices NBIS on future adjusted EBITDA, then adjusts for net debt and dilution.

At each date, enterprise value equals the selected EV/adjusted EBITDA multiple applied to the preceding twelve months' EBITDA. Net debt is then subtracted, and the result is divided by the shares outstanding. This fits Nebius because data-center and GPU investment vastly exceeds current revenue, while customer advances and external financing fund the buildout.

Adjusted EBITDA captures the operating payoff from commissioned capacity. Net debt reflects whether capex and financing have outrun cash generation. The share count matters because Nebius has used equity and convertible financing, while the forecast assumes no dividends or buybacks.

8.3 — The drivers

EBITDA is the main operating driver, and the base case assumes a steep but staged improvement.

The base path is $3.1B in October 2027 and $5.6B in October 2028. It reaches $8.5B in October 2029. The assumptions combine revenue of $10.0B at a 31% margin in 2027, then $16.0B at 35% in 2028 and $22.7B at 37.4% in 2029. The evidence is Q2 2026 adjusted EBITDA of $236.2M and management's 40% AI-cloud margin target.

The bear case gives $0.9B, $2.0B and $3.5B across the three dates. It assumes delayed commissioning, weaker utilization and margins below plan. The bull case gives $4.8B, $8.5B and $12.5B, requiring timely sites, strong pricing and margins above the initial target.

Shares begin at an estimated 287.655M, based on 271.855M reported shares plus 15.8M issued in the note exchange. The base count rises to 303M in 2027 and 319M in 2028. It reaches 336M in 2029 because the case assumes awards and equity funding, including $3.0B and $3.5B of proceeds in the first two years, then $4.0B in 2029.

The bear count reaches 310.2M and 335.6M before 363.1M, reflecting larger equity raises at lower prices. The bull count reaches 289.7M and 298.7M before 319.1M, because customer funding limits new equity but convertible notes still create dilution.

Dividends remain $0.00 in every case and every year. That follows the five-year record of zero payouts, negative free cash flow and a $20–$25B 2026 capex plan.

The base multiple falls from 35x in 2027 to 30x in 2028. It reaches 25x in 2029 as growth matures and cash generation remains unproven. The bear case uses 22x, 18x and 15x; the bull case uses 45x, 40x and 35x. The starting valuation is about 49.9x on annualized Q2 EBITDA, versus approximately 12.1x for CoreWeave, so the premium requires faster growth and reliable execution.

Net debt starts at negative $296.4M, or estimated net cash after the note exchange. The base path reaches $4,103.6M in 2027 and $3,103.6M in 2028. It becomes negative $3,396.4M in 2029 as EBITDA, customer advances and lower capex overtake spending. The bear path reaches $10,403.6M, $16,603.6M and $19,403.6M because spending remains high while collections lag. The bull path reaches $2,203.6M before turning to negative $1,496.4M and negative $10,996.4M, supported by timely commissioning and large prepayments.

The forecast's table (base case, then bear / bull in brackets; USD millions, per share in USD, unless shown)

8.4 — Bear and bull

The bear case assumes demand remains real but capacity arrives too slowly to cover leases, debt and construction spending.

It starts at $30.29 in 2027 and ends at $91.15 in 2029. Its 40% probability reflects power, construction, GPU installation and customer-acceptance risk, alongside three customers supplying 59% of Q2 revenue.

The bull case assumes sites commission on schedule, utilization and pricing remain strong, and prepayments fund most deployment.

It starts at $737.99 in 2027 and ends at $1,405.50 in 2029. Its 15% probability is lower because $12.5B of EBITDA, premium valuation and approximately $11.0B of net cash must arrive together.

8.5 — What would change the view

The view moves toward the bear case if site delays in 2027–2028 leave power, leases and GPUs ahead of billings.

Customer concentration or slower backlog conversion through 2027–2029 would weaken collections and force more debt or dilution. By October 2029, persistently negative free cash flow would compress the multiple.

The view moves toward the bull case if connected power becomes customer-ready capacity during 2027–2028. Strong paid utilization, pricing, repeat workloads and prepayments through 2029 would improve margins and cash conversion. Demonstrated execution by October 2029 would justify a premium multiple and reduce dilution pressure.