Report contents
Industry landscape
Where the company competes: GPU rental gated by power and chip supply
Nebius sells computing time on Nvidia GPUs. Its annual report on Form 20-F for fiscal 2025 describes the core business as an AI cloud: customers rent accelerated compute by the hour or under longer contracts to train and run models, and the group builds the data centers, servers and software that deliver it. That puts Nebius in AI infrastructure-as-a-service, a submarket narrower than "public cloud" and distinct from the software layers above it. The distinction shapes how the company earns money. This is a capital-asset business that buys chips and electricity and resells compute, not a license business with near-zero marginal cost.
The value chain: the chip supplier captures most of the economics
Trace the chain and the leverage is visible. Nvidia designs the GPUs and decides who receives how many of each generation. Contract manufacturers assemble servers. Utilities and grid operators control the power. A cloud operator pulls those three together into sellable capacity, and AI developers and enterprises buy it.
Bargaining power sits at the ends. Nvidia's pricing claims a large slice of an AI cluster's value before the operator earns anything, and its allocation decisions determine who can offer the newest generation on schedule. On the demand side, large AI labs negotiate hard, spread workloads across providers, and can move them in weeks.
The operator in the middle earns a spread: the rental rate on a GPU-hour minus the chip, the power, the building and the capital tied up in all three. That spread widens when capacity is scarce and compresses when a new generation ships and the prior one becomes abundant. Profit therefore depends on utilization, on the depreciable life assigned to hardware, and on how cheaply capacity can be built and financed. Nebius argues it owns more of that chain than the specialists it competes with, saying it designs its own servers and software and runs data centers it controls, including a large site in Finland inherited from its Yandex predecessor. Owning the real estate and designing its own systems should lower cost per GPU-hour. That is a design claim so far; reported gross margin will be the test of it.
Market structure: a crowded field of credible sellers
The competitive set is crowded and widening. Above Nebius sit the hyperscalers, Amazon Web Services, Microsoft Azure and Google Cloud, plus Oracle's cloud infrastructure, each with capacity, customer relationships and balance sheets an order of magnitude larger. Alongside it sit GPU specialists: CoreWeave, Lambda, Crusoe, Together AI and a long tail of regional and venture-funded providers. Market shares for GPU rental are not published in a comparable series, so rank has to be read from capacity commitments and signed contracts.
Apply the practical test: when the leaders cannot be counted on one hand, entry barriers are usually modest. This segment sits roughly there. Barriers to entry: medium. Capital is the visible barrier and it is not binding, since debt and equity have funded a dozen entrants. The constraints that do bind are slower-moving: grid connections and permitted land, allocation of the newest Nvidia parts, and a reference list of customers willing to sign multi-year contracts. Those take years to assemble, which is why the specialist field is narrower than the available money would suggest. But the hyperscalers hold all three already, and their pricing sets the ceiling on rental rates.
Geography: capacity is local, customers are global
Compute cannot be shipped, so every operator competes where its facilities sit. Power availability and planning permission set the boundary, and data-residency rules mean European workloads are served from European sites. The Finnish campus gives Nebius a European base with existing permits and a long-standing grid connection, a genuine head start in a region where capacity is tight. Its build-out in the United States puts it in the largest market, against rivals with deeper local roots. The group's other operations, autonomous mobility through Avride and the TripleTen education arm, sit outside this market and, for now, outside the question that decides the investment case: whether owned capacity turns into a durable cost advantage per GPU-hour, or gets competed away as the next generation of chips arrives.
Competitive advantage
Does the company have a moat? Not yet: Nebius buys the scarcest input and resells the least differentiated output
Moat rating: Weak. Demand is not the question. Revenue went from $0.12 billion in fiscal 2024 to $0.53 billion in fiscal 2025 and then to $0.58 billion in the June 2026 quarter alone, on the company's reported statements, while gross margin reached 77.1% in that quarter. Profit is the question. The June 2026 quarter lost $0.18 billion at the operating line, and fiscal 2025 lost $0.60 billion there. A GPU-hour reseller that earns a wide gross spread and no operating return is collecting a rent on short supply. That rent is real. It is not a barrier.
The part a reader may miss is that the scarce inputs sit on both sides of Nebius rather than inside it. Nvidia decides who receives each accelerator generation and at what price; grid operators and planning authorities decide who gets power at scale; and a customer can re-point a training or inference job to another provider without moving a person. Nebius owns the middle of that chain, meaning land, shells, servers and orchestration software, and the middle is the part every rival also owns. Its Finnish campus and its US sites may hold down cost per hour. They do not hold the customer.
Concentration and returns have to be read together here. The field is not concentrated. Four hyperscalers sell AI capacity as one line inside much larger cloud franchises, and a funded specialist tier that includes CoreWeave, Lambda, Crusoe and Together AI sits beneath them. Nothing in that structure forces a buyer to stay with one seller, so a 77% gross margin on rental hours reflects the 2025-2026 shortage of accelerators and grid connections, not pricing power that would survive a wave of entry.
Market position and stability: small, unranked and too short a record to judge
Market share and ranking: Weak. GPU rental has no comparable published share series, so rank has to be read off scale and signed contracts. By scale, Nebius is a small seller: $0.53 billion of fiscal 2025 revenue against cloud businesses that resell the same GPU-hours at many times that volume. Among the pure-play GPU specialists it is one of several credible operators, not the reference one.
Stability: Not established. The revenue history cannot be read as an organic share series. Fiscal 2021 shows $4.76 billion and fiscal 2024 shows $0.12 billion, a gap produced by the separation from Yandex rather than by winning or losing AI cloud customers. With fewer than three comparable years, share stability cannot be confirmed, and the rough guide of a five-point average share move over five years implying little protection (under two points implying strong protection) needs a series that does not exist here. What the statements do show is growth from capacity added, which is not the same as share taken from named rivals.
Profitability: the profit arrives from below the operating line
Returns: Weak. Return on equity was 2.59% in fiscal 2025, negative 19.59% in fiscal 2024 and negative 2.17% in the June 2026 quarter. An estimate of after-tax ROIC makes the position clearer. Annualizing the June 2026 quarter's $0.18 billion operating loss gives roughly $0.72 billion of negative operating profit, against invested capital of about $10.9 billion (equity of $10.34 billion plus $8.55 billion of debt, less $8.04 billion of cash). That comes to near negative 7%, an estimate drawn from a single quarter that will move with utilization.
The reported profits are the tell. Fiscal 2025 net income of $0.10 billion and the March 2026 quarter's $0.62 billion both arrived while the operating line was negative; in each case the non-operating line did the work. A moat has to appear in operating profit, and so far it has not.
Margins and scale: Moderate. Gross margin went from 37.5% in fiscal 2024 to 68.6% in fiscal 2025 and 77.1% in the June 2026 quarter, a sequence that reflects scarcity pricing plus new capacity coming into service. The March 2026 quarter's 20.9% does not fit that pattern, which shows how much quarterly margin depends on when capacity starts depreciating. The distance between a 77% gross margin and a negative operating margin is the cost of building ahead of revenue: sales staff, engineering and the fixed cost of data centers. Scale would close that gap if utilization rises, and whether it rises is the operating question behind the rating.
Overall rating: Weak
No measured share, no positive operating return and no captive customer add up to a business whose current profitability rests on a shortage. Profitability carries extra weight in this call, and it is negative. The margins dimension is the only one that reaches Moderate, and it gets there on scarcity rather than on anything rivals cannot copy. The figure that settles the rating is the June 2026 operating margin of negative 31%, which is what a 77% gross margin leaves once the fixed costs of the build show up.
Where the advantage comes from and how long it can last: permitted power at one site, not a lock on customers
The classic sources of protection are thin in GPU rental. The customer arrives with portable work, the seller buys the same accelerators as everyone else, and the price of a rented hour clears in a market with a dozen funded suppliers. The closest thing Nebius has to a barrier is the queue for grid connections and permitted land, which takes years to clear and cannot be bought at short notice. The single test that follows from that is whether the operating margin turns positive on installed capacity before rental rates on that capacity fall. Operating margin is a blunt proxy for the sharper measure, gross profit per GPU-hour after depreciation, but it is the one the statements support.
Supply-side advantage: owning the building helps, and the chip eats the gain
Supply-side advantage: Weak. The cost structure sits between a pure reseller's and a chip designer's. The largest line in any cluster is the accelerator, bought at list price from Nvidia, plus the networking wrapped around it. Designing its own servers and running data centers it controls, including the Finnish site inherited from its predecessor, trims the non-chip part of cost per hour. That part is the smaller half of the bill, so the attainable edge is real and bounded.
What the edge is worth, roughly: the June 2026 quarter turned $0.58 billion of revenue into $0.45 billion of gross profit while absorbing about $0.63 billion of operating cost, the arithmetic being gross profit plus the operating loss. That is an estimate, since the expense lines are aggregated, but it fixes the scale of the problem. A few percentage points of cost saving on the non-chip share of the stack cannot offset $0.63 billion of quarterly fixed cost; utilization and rental price can. The advantage would break if Nvidia allocation loosened and hourly rates fell faster than Nebius's own cost per hour, which is the normal direction for a generation of hardware that is one step behind.
Demand-side advantage: hourly rentals leave the customer free to leave
Demand-side advantage: Weak. There is no repurchase habit in the sense that matters, and switching costs are administrative rather than structural. A multi-year contract helps while it runs, but the underlying workload is portable by design, and the buyers are AI developers with the engineers to re-point jobs between providers. Captivity would show up as pricing power: a rival discount that fails to take volume. The evidence available is consistent with a market that clears on price, since several competitors are adding capacity at the same time and reported gross margins stay wide. A disclosed base of take-or-pay contracts with multi-year terms, or a retention figure for capacity, would settle the question; the public record does not.
Scale and permits: the grid connection is the one asset that takes years
Permits and power: Moderate. This is where the advantage is least easy to copy. Interconnection queues, permitted land and firm power supply are slow, local and cannot be conjured with capital alone. The Finnish campus brings a long-standing connection and permits in a European market where capacity is tight and data-residency rules require local sites. Ordinary licensing and quota regimes do not govern this industry, but planning permission behaves much like one, and that makes it the most protected asset Nebius holds. It is also a depreciating one: a national grid build-out or a policy change that speeds up connections would erode it without warning.
Durability: each chip generation resets the race
The decisive substitute is not another cloud seller but cheaper compute. Each Nvidia generation lowers the work done per dollar, so a fleet bought in one cycle faces declining rates when the next ships. In-house silicon at the hyperscalers and steady software efficiency gains do the same job from other directions, and neither requires anyone to switch providers. That is why capacity, once built, is a wasting asset unless it is contracted forward.
Access to capital is often called a moat here and is not one. Nebius funded its build with roughly $8.55 billion of debt and repeated equity issuance, and a rival with the same lenders can do the same. Total liabilities of $17.62 billion exceed that debt by about $9 billion, and operating cash flow of $2.25 billion in the June 2026 quarter ran far ahead of $0.58 billion of revenue; both point to customers paying in advance, which helps the cash position without protecting the price. The company's Finnish site, its software and its server design are execution advantages that keep it in the race. They do not stop a competitor from undercutting it next quarter.
The questions that would change this assessment are narrow. Does operating margin turn positive on installed capacity before rates on the current generation of accelerators fall, or does free cash flow of negative $3.41 billion in the June 2026 quarter become the pattern for another year? Can a meaningful share of capacity be sold under multi-year, committed contracts that survive a price war? And does that $9 billion of non-debt liability reflect customers locking in years of demand or only the timing of payments? Any evidence of pricing power that shows up as a rival discount failing to win volume would raise the rating; continued operating losses funded by new capital would confirm it as it stands.
Generated by DeepVal AI. Data and models may contain errors or be outdated. For research only; not personalized investment advice.
Continue reading the full report
- Comparison and summaryHow Nebius Group N.V. compares with its closest rivals, and our moat rating
Free to sign up · 20 research points · unlocking this report uses 10
20 points cover 2 full reports. No card needed.
Research reflects its generation date. Preview the first two modules.
More on Nebius Group N.V.
Nebius Group N.V.
Business, moat and key questions
Open company page