Nebius

IThe business

Nebius sells time on graphics chips. A customer that wants to train an AI model, or run one for users, needs thousands of NVIDIA chips wired together in a building with power and cooling. Nebius builds or leases the building, buys and installs the chips, and rents the result by the month, along with storage, networking and software to run it.² The company is headquartered in Amsterdam. It is Dutch, files annual reports on Form 20-F and results on Form 6-K like other foreign companies, and its dollar results are reported under US accounting rules.² The year is the calendar year.

It started in July 2024. That month the company, then called Yandex N.V., finished selling every Russian business, more than 95% of its revenue, assets and employees, and kept a small group of businesses abroad.² Nebius is what was left. In 2023 its continuing business took in $9.8M. In 2025 it took in $529.8M, and in the first half of 2026 $981.3M.²³ There are therefore three audited years for the business I am valuing, which is not enough to build a ten-year record from, and I say so in Valuation.

How it is built

Capacity is counted in megawatts of power. At the end of 2025 about 170 MW was active, meaning it was drawing power in servers and available to earn revenue.² Management guides to 800 MW to 1 GW connected by the end of 2026 and says it has contracted, meaning secured with land and power commitments, more than 2 GW in February and now targets 5 GW by year end.⁵⁸² Those are different things: contracted power is a right to build, connected power has electricity at the building, and active power is earning. The gap between them is where a year of delay hides.

It operates four kinds of site: greenfield, where it owns the land and designs the building, as in Mäntsälä, Finland and in Missouri and Alabama; brownfield; build-to-suit, as at Vineland, New Jersey; and co-location, renting space in someone else’s data centre in places such as Paris, Kansas City, Iceland, London and Israel.² In May it also announced up to 1.2 GW of power and land for an owned site in Pennsylvania.¹² The company says it designs servers and racks itself, that its Finnish site reuses server heat for local homes, and that its own-design data centres carry up to four times lower overhead than the industry average.² I can check none of those efficiency claims from filings. They are the company’s own.

What it sells

Three layers. The bottom is plain capacity: GPU clusters, storage and networking. On top is a layer for machine-learning teams to schedule jobs and manage clusters. At the top is Token Factory, launched in November 2025, which rents inference, running a finished model, by the token instead of by the chip-hour.² Management says Token Factory workloads more than tripled in the second quarter.⁸ The company bought Tavily, an agent search provider, in February, and Eigen AI and the Clarifai core team this spring; the cash paid for acquisitions in the first half was $251.6M net.²³⁸ In October it bought Inferize, an inference software company, on terms it did not disclose.¹⁵

The AI cloud business was 98% of group revenue in the second quarter, $574.9M of $582.3M, with an adjusted EBITDA margin of 49.7%.¹⁶ Two other businesses sit inside the group. Avride, which makes self-driving cars and delivery robots, lost $74.2M at the adjusted EBITDA line in the first half; TripleTen, an education business, lost $19.8M.¹⁶ Nebius also owns a stake in ClickHouse, a database company that raised money in January at a reported valuation of about $15B, and a stake in Toloka; the balance sheet carries its non-marketable equity investments at $1.6B R (reported).¹⁶³

Who pays

Two kinds of customer, and they pay on different terms. The first is the hyperscalers. On 8 September 2025 Nebius signed Microsoft for dedicated capacity at Vineland over five years for about $17.4B, rising to about $19.4B if Microsoft adds to it.⁹ On 1 November it signed Meta for two clusters over five years for about $2.9B.¹⁷ On 13 March 2026 it signed a larger Meta agreement of up to about $27B: $12B of dedicated capacity delivered from early 2027, plus an order of up to $15B under which Meta must buy any capacity of certain clusters that Nebius cannot sell to others.¹⁰ That last order is the unusual one. It is a backstop on future unsold capacity, up to $15B, and it matters for financing as much as for revenue.

The second kind is AI developers and software companies that rent on one-to-three-year contracts or by the hour. The company names Reflection, Cohere and a US quantitative trading firm among Q2 wins.⁸ In September Palantir named Nebius its preferred sovereign AI infrastructure partner.¹⁸ The filings do not name Nebius’s largest customers. They show that three customers each accounted for more than 10% of second-quarter revenue: 24%, 21% and 14%.⁴ One customer was 83% of receivables at the end of 2025.⁴ A business like this is the opposite of diversified, and the 20-F notes that the hyperscalers it serves may also offer competing clouds.²

The backlog, and the money that came with it

Remaining performance obligations, contracted revenue not yet recognized, were $37.5B at 30 June 2026. The company expects 36% to be recognized in the 24 months to June 2028, 40% in the two years after that, and 24% later.⁴ All of it is the AI cloud. The Microsoft and Meta orders I can see announced total about $32.3B at signing, which is a large part of that $37.5B C (calculated).⁹¹⁷¹⁰ The notes do not break it down by customer, so I cannot say how much remains.

Customers also paid in advance. Deferred revenue, cash received for services not yet delivered, rose from $1.58B at the end of 2025 to $5.98B at 30 June 2026, and management expects more than $9B of prepayments in 2026.⁴⁸ In the first half those prepayments, $4.4B, made up almost all of the $4.5B of operating cash flow C (calculated).³ About 70% of Q2 deals carried a prepayment, covering 50% to 60% of the cost of the equipment for those deals, the company says.⁸ This is the heart of how Nebius pays for its growth, and I treat the prepayments as borrowing, since the revenue they cover will arrive without cash.

What it costs to build

Capital spending was $4.07B in 2025 and $8.13B in the first half of 2026 alone; the 2026 guide is $20B to $25B.²³⁵ Property and equipment rose to $13.0B from $5.6B in six months, of which $7.8B was not yet in service.⁴ Leases that have been signed but not started will add $12.1B of payments over lease terms of up to twelve years, and $5.3B of minimum power and energy commitments run to 2030 and beyond.⁴ Those commitments are why I treat the megawatt economics, rather than the next quarter’s revenue, as the thing to get right.

IIThe moat

I ask two questions. Could a competitor take this customer at the same price? And would the customer mind? On the first, today, yes. On the second, for most of Nebius’s customers, not much.

The product is close to a commodity

A rented GPU cluster is largely the same wherever it sits, because the chips come from one supplier and the software that matters to a researcher runs on any of them. The company lists its competitors itself: Amazon, Google, Microsoft and Oracle on the large side, and CoreWeave, Crusoe and Lambda among the AI-specialist clouds, with Fireworks AI and Together AI competing with Token Factory.² Switching a training run between clouds is work, but a customer that has done it once has done it. The best evidence of how little protection the product alone gives is in the company’s own numbers: it repriced older-generation GPUs more than 30% higher in the second quarter because capacity is short, not because the product is better.⁸ A shortage can reverse.

What is hard to copy

Four things, none of them a product feature. First, power: land with electricity secured is scarce, and Nebius says it had contracted more than 2 GW by February, and it has raised its year-end 2026 target from more than 1 GW in August 2025 to 5 GW.²⁸ Second, chips: NVIDIA invested $2B for a pre-funded warrant over 21.1M shares in March, holds 9.3% including those shares, and said in the same announcement that the partnership would enable Nebius to deploy more than 5 GW of NVIDIA systems by the end of 2030.¹⁹²⁰²¹ I take that as evidence of access to supply in a constrained market, though a supplier’s investment is not a promise to allocate chips. Third, credit: contracts with Microsoft and Meta let it borrow against signed revenue, as the $775M July facility at SOFR plus 2.50% shows.²² Fourth, time: it already runs a site and a software platform, and a new entrant needs years.

Where the protection is weakest

The protection has to be paid for every five or six years. Chips age out, and a data centre built for today’s racks must be refitted for tomorrow’s. A software business keeps its protection once built. A GPU landlord must rebuy a large part of it, which is why I model server replacement as a cost, not an afterthought. The second weakness is that most of the best sources of cash are things competitors can also buy. When the capital markets turned in October, Nebius shares fell with CoreWeave’s, which tells you investors saw them as the same kind of exposure.²³²⁴

There is also the question of what customers do at the end of a contract. A five-year deal at fixed prices protects only until it ends. After that, the customer is free to take bids, and the chips in the cluster are five years old. Management’s answer is mix: long contracts for financing, mid-term contracts for the core, short contracts for premium prices.⁸ It is a sensible book. It is also a book that depends on prices staying high.

My judgement: a capital and execution advantage in a market that is short of supply, not yet durable protection. The honest test is whether new capacity keeps earning well above its cost after other builders catch up. I do not assume it does. In my base the income on new equipment fades from 42% to 30% a year on cost by 2033, and in my bear it fades to 27%, where growth stops adding value E (estimate).

IIIThe competitive advantage

The company’s case is vertical integration: it designs the server and the rack, builds the data centre, runs the cloud and offers software on top, so it controls cost and speed at each step.² If true, it shows up in two places. Cost to build a megawatt, and time to bring one online.

Speed

The evidence on speed is mixed but real. Nebius says it has delivered all capacity tranches due to Microsoft under the contract, including two in the month before the August letter.⁸ Microsoft has the right to terminate a tranche if delivery is late after a grace period, and Meta has the same right on its first order, so on-time delivery is paid for by the contract, not just claimed.⁹¹⁷ Active power was about 170 MW at the end of 2025 and I have no later figure.² Connected power guidance for the end of 2026 is 800 MW to 1 GW.⁵ Hitting that will be the first hard evidence that the speed claim holds at scale, and the third-quarter results due on 10 November will show part of it.²⁵

Cost

The company says that new deals closed in the second quarter pay back, counting the equipment and the operating costs that go with it, in one year and ten months, against its earlier two to three years.⁸ That implies the cash it expects to earn each year is a bit more than half the cost of the equipment C (calculated). I use less. For capacity installed in 2027 I assume 42% a year on cost, which is a payback of about 2.4 years before overhead, fading to 30%. The management figure is for deals closed in the best quarter yet, on forecast costs, and includes capacity not yet built.⁸

The cost side is where an advantage would show, and it is where the filings are quiet. Cost of revenues, power, rent for co-location and staff, was 23% of revenue in the second quarter, down from 29% a year earlier, and the company puts the drop down to operating leverage.³ The cost of building a megawatt is not disclosed. The 2026 capex guide of $20B to $25B, set against an increase in connected power of roughly 630 MW to 830 MW over the year, from about 170 MW, works out to $24M to $40M a megawatt. Part of that spending buys capacity for 2027, and active and connected are not the same, so I treat it as a loose cross-check C (calculated).⁵² I use $38M for capacity installed in 2027, with 70% of it servers and networking and 30% buildings and power E (estimate).

Where it wins

Buyers who need capacity soon and cannot get it from a hyperscaler, which is most AI labs and start-ups today. Buyers who want a second source of compute beside Microsoft or Amazon, which is the case for Meta and Microsoft themselves. Buyers in Europe who care where their data sits, which is the pitch to Palantir’s customers.¹⁸ And buyers who want inference, the running of models, priced by the token and bundled with search and tooling.²

Where it loses

Scale and cost of money. Amazon, Google, Microsoft and Oracle can fund data centres out of operating cash flow. Nebius funds them with convertible notes at 0.5% to 4.5%, asset-backed debt at SOFR plus 2.50%, a $3.5B secured bridge at SOFR plus 2.75% rising to 3.50%, share sales and customer prepayments.⁶²²²⁶⁸ Its Chief Executive said, in a video interview reported second-hand, that some capital sources were temporarily closed.²⁷ A landlord that rents equipment on thin margins over its cost of money has less room for error than one that does not. On financing, Nebius competes with CoreWeave, not with Amazon, and one press summary, which I could not check, put CoreWeave’s 2032 notes near 14.8% in mid-October.²³

IVManagement

Arkady Volozh, 62, founded Yandex in 2000 and runs Nebius as chief executive and an executive director.² Ophir Nave, 57, a former Wachtell, Lipton lawyer, is chief operating officer. Andrey Korolenko, who ran infrastructure at Yandex, and Roman Chernin, who led Yandex Search and Maps, are co-founders and run product and infrastructure and the business side. Dado Alonso, who has been at Booking.com and Amazon, became chief financial officer in June 2025; Marc Boroditsky, from Cloudflare and Twilio, became chief revenue officer in May 2025.² It is a team that built data centres and search at scale, and one that arrived in the US cloud market recently. Of the eight on the senior team, the finance chief and the sales chief have been there about a year.

Who controls it

Volozh does, with a minority of the money. The company has two classes of shares. Class A has one vote. Class B has ten. A trust whose beneficiaries include Volozh or his family holds 28,655,509 Class B shares, which is 51.6% of the votes and 11.3% of the shares.² That makes Nebius a controlled company under Nasdaq rules, which lets it skip a fully independent nominating committee, though its board is a majority independent.² At the AGM in August every resolution passed. The share of votes against his re-election was 5%, and the board got a five-year authority to issue up to another 20% of the shares and set aside existing holders’ rights to buy first.²⁸ A holder of Class A shares has a say on paper and little in practice. The board includes Matt Weigand, a partner at Accel, who appears on this page in two roles: as a director, and as the interviewer on the Accel podcast I cite in Section VII.²²⁹

How they are paid, and what that pays them to do

Cash pay is small. The eight senior managers received $5.9M in cash in 2025 as a group, and the six non-executive directors $0.9M.² The rest is stock. In 2025 the senior team was granted 6.35M options and 226,312 restricted share units, vesting over four years.² Many of the options are “premium priced”: Nave, Chernin and Korolenko each hold 625,000 options struck at $100 a share, and Bunina holds 409,500 at $40.² The shares closed at $154.14 on 30 October.¹ Options struck well above the price at the time of the grant pay only if the stock rises a long way, and that pays managers to build as fast as the money allows. It does not tie pay to returns on the equipment, to payback, or to the cost of the debt. Nothing in the pay plan I can see asks for cash.

Volozh’s own incentive comes from ownership. His trust’s Class B stake is worth about $4.4B at $154.14 C (calculated). I see no options for him in the ownership table. That is a real stake and it is large, though the votes he controls are about four and a half times his share of the equity.

What they sold

Directors and officers began filing Forms 3 and 4 in March 2026. Between 31 March and 15 October, ten of them sold 2.01M shares for $423M, at prices from about $100 to $243.¹¹ Most of it was pre-arranged. Volozh sold 117,763 shares for $23.3M, all of it automatic sales to cover tax on vesting shares. Korolenko sold 1,121,294 shares for $224M, of which $207M was under a trading plan adopted in December 2025 and $17.5M to cover taxes. Nave sold 500,000 for $118M on 5 October under a plan adopted on 22 May; his filing says the shares were about 17% of his granted equity and that no further sales will be made under the plan. Ryan, a director, sold 50,000 for $13M and Boynton, the chairman, 37,502 for $8M, with no plan or tax reason stated.¹¹ The one purchaser was a director, Arne Grimme, who bought 2,014 shares in eleven small trades between 15 and 29 October for about $0.3M.¹¹

I draw no conclusion about intent; plans exist to take timing out of the seller’s hands. What the numbers show is a team that took out $423M in seven months, nearly all of it on a schedule, in a year when the company itself raised cash from other investors.

What they did with capital

The cycle is short, so the record is too. Shares outstanding were 361.5M in January 2023 and 235.8M at the end of 2024, after the divestment left 126M shares in treasury. They were 253.0M at the end of 2025, 271.9M at 30 June 2026, and about 288M after the August exchange, plus 21.1M held by NVIDIA as a pre-funded warrant, which is 31% more than at the end of 2024 C (calculated).²³⁶¹⁹ Between June 2025 and August 2026 it sold $14.25B of convertible notes in four deals, took $2.0B from NVIDIA at about $95 a share, and sold 12.7M shares through its at-the-market programme at an average of $223.6 for $2.8B.⁸¹⁹⁶

Some of that timing was very good. The ATM shares were sold close to the June peak. NVIDIA’s $2.0B came when the price was about $112 and was the largest single equity cheque of the year.¹ Some was not. The June 2025 notes had a conversion price of $51.45 and the shares were $223.90 when Nebius exchanged $800M of them for 15.8M shares in August, which at that price is about $3.5B of stock for $0.8B of notes C (calculated).⁴⁷ The holders’ right to convert at $51.45 was written when the stock was cheap. That is not an error, because the notes cost almost nothing to carry, but it is the price of financing a start-up with options on its own stock.

Following the incentives, then: Volozh is paid by the share price, and the share price rewards size. The team has funded itself with new shares or convertible notes in every year I can see. The board gave itself authority to issue 20% more shares for five years, and shareholders approved it with 94% of votes cast in favour C (calculated).²⁸ The accounting record needs saying too. The auditor’s report on internal control was adverse for 2024 and for 2025. The two weaknesses at the end of 2025 were in the controls over property and equipment, including depreciation start dates and an asset-count reconciliation, and over IT controls and revenue for TripleTen, about 10% of revenue.³⁰² The board dismissed the auditor, Reanda, after the 2025 audit and the first-quarter review, and put forward Deloitte; shareholders approved the appointment on 25 August.³¹²⁸ Equipment is almost half of the balance sheet, so a weakness there matters more than one at a software company.

VFinancials

The business I am valuing has three audited years, 2023 to 2025, and the twelve months to June 2026. Before July 2024 these were part of Yandex, whose accounts are in a different business, in a different country, mostly in roubles, and I do not use them. So I cannot show ten years, and I do not think any model can honestly pretend to. Revenue rose from $9.8M to $529.8M in two years and to $1,355M over the latest twelve months, and the company has not had an operating profit in any of them C (calculated).²³

The path is a steep one. The second quarter of 2026 alone was larger than all of 2025 by $52M, at $582.3M against $529.8M.³² Adjusted EBITDA, the company’s own cash-like measure, was negative until the fourth quarter of 2025 and was $236.2M in the second quarter of 2026, a 41% margin.¹⁴³ Cost of revenues fell from 200% of revenue in 2023 to 31% in 2025 and 26% over the latest twelve months C (calculated). Depreciation is the line that matters next: $417.9M in 2025, 79% of revenue, and $765M over the twelve months, 56% of revenue C (calculated). In January 2026 the company lengthened the life of servers from four years to five, which it said would cut 2026 depreciation by about $168M.² That is a genuine choice, and a favourable one. Everything in my model after 2026 uses a longer life too, with a sixth year at a lower price, and I show what a shorter one does.

What an owner could really take out

Nothing, and a long way from nothing. Over the twelve months to June 2026 operating cash flow was $5.3B, but the change in deferred revenue was $6.0B of that, cash paid by customers for service they have not yet received C (calculated).³² Capital spending was $11.1B and stock pay $0.19B. Owner cash flow, after equipment and stock pay, was −$6.1B, and without the prepayments −$12.0B C (calculated). Reported net income from continuing operations was a profit of $42M, helped by $780.6M revaluing the ClickHouse stake in the first half; the operating loss was $684M C (calculated).¹⁶ The gap between the reported figure and the cash an owner could take out is about $6B in one year.

How much of the spending just stands still? In the first half, depreciation was $472M against capital spending of $8,130M, and almost all of the spending was growth. The deeper question is the steady-state number: when the fleet stops growing, servers must be replaced every five or six years. At my base cost of $38M a megawatt with 70% servers, replacing one megawatt costs about $4.4M a year in 2027 dollars, around a quarter of the revenue a new megawatt earns E (estimate). That, not the depreciation charge in the income statement, is what an owner of a mature fleet pays each year before taking anything out.

Quality of earnings

Three things. Adjusted EBITDA leaves out stock pay of $137.8M in the first half, up from $32.1M a year earlier, including $74.9M tied to the Eigen AI deal, and it leaves out depreciation, interest and deal costs of $60.1M.³¹⁶ Interest expense was $182.8M in the half, and in the second quarter $40.2M of interest was capitalized into equipment instead of expensed.³⁴ And revenue per megawatt is not yet visible: the company reports annualized run-rate revenue, the last month of the quarter times twelve, and it was $3.0B at the end of June.⁸ That is a measure of how fast capacity is arriving, not of what the average megawatt earns.

The balance sheet

At 30 June Nebius had $8.04B of cash, $1.06B of restricted cash, $8.50B of convertible notes carried on the balance sheet, which will cost $10.04B at maturity, and $5.98B of deferred revenue.³⁴ Since then it has issued $5.75B of new convertible notes, exchanged $800M of older ones for shares, drawn a facility of about $775M, and signed a $3.5B secured bridge with Citibank that matures in October 2027 and pays a rising margin.⁶²²²⁶ A convertible note is cheap debt that can become stock. Whether the first notes become stock is shown by the price: the June 2025 notes convert at $51.45, the September 2025 notes at $138.75, the March 2026 notes at $180 to $183, and the August 2026 notes at $313 to $325, so at $154.14 the first two groups are in the money and the last two are not.⁴⁷ Property and equipment is $13.0B and the pro forma total of promises made, leases, power contracts and the like, is larger than the balance sheet shows.

VIValuation

First, how this goes badly wrong

Before any numbers I ask what would most likely ruin the business or the investment, and work back. There are three, and they feed each other.

Rental income on new equipment falls to what the equipment costs. Today a megawatt is short, and new deals are priced at $20M to $25M a year.⁸ That is a shortage price. If chip supply catches up, if one big customer builds its own capacity, or if a newer chip makes last year’s cluster less wanted, the price drops and the equipment bought at 2027 prices earns little more than its cost. The company is spending $20B to $25B this year, almost all of it on that bet.⁵ In my bear the income on new equipment falls to 27% a year of cost, and my value for the shares is $7.

The money stops before the buildings are done. The business needs to find about $36B on top of the $13.7B it has, before its cash turns, on my base, and the 20-F says the strategy depends on raising it.² Microsoft can end a tranche if delivery is late, and Meta can end its first order the same way.⁹¹⁷ A missed tranche would be a loss of revenue and a signal to lenders in the same week. The Chief Executive said in October, in an interview I know only second-hand, that some capital sources were temporarily closed.²⁷ The newest debt is shorter and secured: $3.5B due in October 2027 that costs 2.75% over the benchmark rate and steps up by 0.25 points every 90 days.²⁶

A few customers, on terms I cannot see. Three customers were 59% of second-quarter revenue.⁴ A Meta backstop of up to $15B depends on Meta’s credit; the small labs and start-ups on the shorter contracts depend on their own. If AI spending slows, the first sign would be mid-term customers not renewing and short-term prices falling.⁸

So the questions for everything below are: what do new megawatts earn after 2030, can the money be found at a cost the equipment can carry, and how much of the best case is already in the price.

What an owner gets today

Over the past year, operating cash flow was $5.3B, but $6.0B of it was customer prepayments. After equipment and stock pay, an owner would have put in $6.1B, and without the prepayments $12.0B C (calculated).³² The reported profit of $42M is mostly the revaluation of ClickHouse.¹⁶ In my base, the first year in which the business produces more cash than it spends is 2030 E (estimate).

What I cannot judge

Three things, and the answer depends on all of them. One: the long-run income on new equipment. One point of it is about $19 a share in my base, and I have no way to know whether it settles at 25% or 35%. Two: how long a server earns. The company changed its mind about this in January, from four years to five, and my base gives a sixth year at 60% of the price.² Three: whether the bond and bank markets keep lending at these prices. I show ranges on all three instead of pretending to a precise number. If the range from bear to bull looks absurdly wide, $7 to $408, that is the honest answer, and it is how the business is built: a small layer of equity on top of a very large base of equipment, debt and prepaid contracts.

How I value it

I build the business from the megawatt. Each year I choose how many megawatts to add, what they cost, and what they earn. Revenue a megawatt is set from a single idea: the cash income on new equipment, as a share of its cost. A new rental business earns more than its cost of money while capacity is short, and less of an excess as others build. In my base that income starts at 42% a year on cost for 2027 capacity, in line with management’s own $20M to $25M a megawatt, and drifts to 30% by 2033. I take the cash margin on rental revenue at 75%, then charge overheads and stock pay, which together fall from 25% of revenue in 2027 to 10% at scale.³ Servers, 70% of cost, are replaced every six years; buildings and power, 30%, are written off over twenty. Tax is 23% after losses carried forward. The existing fleet is the 850 MW and the $8B of run-rate revenue the company guides to at the end of 2026, which I keep until its servers are replaced in 2032.⁵

The model runs from the second half of 2026, using the company’s guidance for that period: revenue of $3.1B for the year, a cash margin of about 40%, and capital spending of $22.5B, with the first-half actuals taken out.⁵³ I assume the 2026 guide already pays for a quarter of the 2027 builds and that the rest, with half of the 2028 builds, falls in 2027, which puts base capital spending in 2027 at $31.0B; one summary of analysts’ estimates puts it near $29B.²⁴ Management plans to deploy more than 1 GW a year from 2027, against 650 MW in my base and about 1 GW in my bull, and I found no company figure for 2027 capital spending.⁸ The timing is an estimate: if the 2026 guide pays for half of the 2027 builds, base value is $148, and if it pays for none, $119 C (calculated). From 2027 to 2041 I add the cash from the business, then a terminal value on 2042 with growth of 2.0%, 3.0% and 3.5% in the three cases. Capacity grows in the terminal year in step with revenue, and servers are replaced every six years in perpetuity. Values are per share at 30 October 2026, from the 30 June balance sheet adjusted for the August notes and exchange, on 347M shares at the price and more in the bull. I keep each convertible note as debt until the value per share passes its conversion price, and then count its shares. Customer prepayments of $6.0B are treated as debt, since the revenue they pay for arrives with no cash, and the prepayments yet to come are left out because they are borrowing too.

The discount rate. The ten-year Treasury yield was 4.94% on the price date, and Damodaran’s implied equity risk premium 4.20%.³³³⁴ Nebius’s weekly beta against the S&P 500 is 2.83 over two years, which puts the cost of equity at 16.8%, and with the debt at 7% before tax the blended rate at about 14.5% C (calculated).³⁵¹ I use 12%, lower than the formula, for two reasons. A beta measured over a rise from $20 to $287 and a fall to $154 overstates how risky the equipment will be once it is under long contracts. And the Microsoft and Meta contracts announced add to $32.3B at signing, against $37.5B of contracted revenue left, so a large share of the contracted book sits with two companies with strong credit.⁹¹⁷¹⁰⁴ But the rate is a judgment, and the table below shows what 14% does: my base falls to $73.

Bear

Chip supply catches up with demand in 2027 and prices fall. Capacity installed from 2027 earns 32% a year on its cost, falling to 27% by 2030, which is roughly what I need to pay for the money. Capital markets stay open for contracted capacity only, and Nebius has 1,650 MW earning by 2030, and after that only replaces servers. Overheads stay heavy. The existing contracts pay but earn little over their cost, and the shares are worth almost nothing because the debt and prepayments come first.

Base

Nebius delivers 800 MW to 1 GW connected by the end of 2026 and about 650 MW a year for four years, two-thirds of the 1 GW a year management plans. New capacity earns 42% a year on its $38M cost, drifting to 30% by 2033 as other builders arrive. Servers earn full price for five years and 60% in a sixth. Overheads and stock pay fall from 25% of revenue to 10%. Revenue reaches $54.0B in 2030 and $91.1B in 2041. The business needs about $49.9B more than it generates before the cash turns, in 2029.

Bull

Management’s own plan works: more than 1 GW a year from 2027, the Meta backstop is never drawn because the capacity sells, and new capacity keeps earning 36% or more on its cost. Nebius reaches 5,100 MW by 2030 and 9,100 MW by 2041, at a cost of $34M a megawatt, with a cash margin of 72% at scale. Revenue is $129.2B in 2041.

What the price says

Turn it round. At $154.14 the shares, counting the notes that are in the money as shares, are worth $53.5B. Add the other claims and the market puts a value of $54.2B on the business C (calculated). My base is $47.3B E (estimate). The price is 15% above my base, which means it already contains most of what I assume. It also means that I am not looking at a bargain, or an obvious mistake. I am looking at a price that needs the long-run income on new equipment to settle at 31.1% a year, against my 30% C (calculated).

How much the answers move

The point of these tables is that no number here is stable. A change of two points in the long-run income moves the base by about $39. A change from my server life to five flat years moves it by about $47. Moving the discount rate from 12% to 14% moves it by about $61. Those are all inside what I would call reasonable. It follows that anyone who quotes a single value for this company to the dollar has hidden the range. I would say my value is between about $100 and $174 on the base case, at 13% to 11%, and between close to nothing and more than $400 across the three.

The money

The base needs about $49.9B of cash to go out between now and 2029, before it comes back. The company has $13.7B of cash after the August deal, and it receives customer prepayments, which I treat as borrowing.⁶⁸ So the money still to find is about $36B, in the base, over about three years. Since March Nebius has raised about $19B: $10.1B of convertible notes, $2.0B from NVIDIA, $2.8B from share sales since May, a $775M secured loan and a $3.5B bridge C (calculated).^631982226 The pace is possible. The terms are what matter. If a third of the gap were raised as new shares at today’s price, my base would be $138 and my bull $366 E (estimate). My bear would show $36, but only because shares sold at $154 are worth far more than my bear value, and in that case buyers at $154 would not turn up. The risk in the bear is not dilution. It is that the capital is not there at all.

The margin of safety

I set it from the bear case: the price at which my $7 bear value is three quarters of what is paid, so $10. That is almost meaningless as a price, and I say why. In the bear the equity is a thin slice above about $13B of convertible notes and $6B of prepayments. Whatever the price, a bad outcome in the equipment market can take most of it. There is no price I can name at which the downside here is covered.

The price against the range

The price is 15% above my base and 62% below my bull. At $154.14 my base cash flows return 10.5% a year to year 10, my bull 22.7%, and my bear loses nearly all of it E (estimate). Weighting the cases 25%, 50% and 25% gives $171, which is above the price because the range is skewed upward: the bull is a bigger distance from the base than the bear is, in dollars E (estimate). The market may be right and I may be wrong. I may be too low on volume and too high on how fast other builders appear. I would rather show that than move the numbers until they meet.

Against the alternative

A ten-year Treasury pays 4.94%.³³ The nearest comparison on this site is my Oracle page, which is the other large builder of AI data centres funded with debt and customer commitments. On my numbers Oracle returns 11.4% a year from its price to year 10. NVIDIA, which sells the chips Nebius rents and holds 9.3% of it, returns 9.3%.²¹ Nebius’s base return is 10.5%, within a point or two of both, but around it the range is far wider: the bear loses nearly all of the money and the bull is 2.6 times the price. Oracle and NVIDIA have larger, older and profitable businesses with floors under them. Nebius has a higher best case and no floor.

Quality and price, separately

How good is the business? A rented-chip business with strong customers and a shortage in its favour is a good business to be in today, and a hard one to stay good in. It earns well on the way up and must reinvest almost all it earns to keep the capacity. Its reported numbers and its cash numbers are far apart and the controls over its own equipment accounts are reported weak. How good is the price? A little above my base and unsafe on my bear. This looks like a good business in the middle of a boom, at a price that already assumes the boom lasts.

How the odds multiply

The base needs several things to go right together. My guesses at each are in the table. They are not independent, since a strong market helps all five, so the true chance of all five is higher than the product. But each looks more likely than not on its own, and the product is 15%.

Where this could be in 5 and 10 years

Value in year 5 is my model’s value at October 2031 of the cash after that date, minus the debt the build will have added by then, per today’s share count and in that year’s dollars. Debt carries 12%. No new shares are issued. Year 10 is the same at October 2036. Nebius pays no dividend. Owner cash is what the business generates once the debt is repaid. The bull repays it in 2033; the base and bear still owe money in 2036, so they pay out nothing by then. These are estimates of value, not forecasts of the share price.

For the base to hold at five years, Nebius has to have about 3.9 GW earning by 2031 and be carrying about $68.9B of net debt, with new contracts that repay their equipment in under three years. At ten years the base still carries about $20.3B of net debt, and the income on new equipment has to have settled near 30%, with servers earning for five years and a bit. The numbers that drive it are megawatts built, income on cost, the cost of the money, and the discount rate. The most likely way my base is wrong is that new megawatts earn less than 30% after 2030: at 26%, my base today would be $53. Value can sit above or below the price for years.

VIIThe case for the price

What most people miss

Three things cut in Nebius’s favour and rarely come up. First, the Meta agreement of March 2026 includes an order of up to $15B under which Meta must buy any capacity on certain clusters that Nebius cannot sell to others for the five years after it is deployed.¹⁰ That is a floor under some of the demand risk for the clusters it covers, and it is why the company can plan to sell the rest to others. Second, the company finances deals in a way that looks like a bank’s: the July facility was backed by a signed customer contract and covers more than 100% of the cost of the equipment under it, at SOFR plus 2.50%.³⁶²² The company says it has more than $40B of other contracts from investment-grade customers that could support the same structure.³⁶ Third, customers are paying in advance. About 70% of second-quarter deals included prepayments, and they covered 50% to 60% of the equipment cost of those deals.⁸ In a shortage, that makes the customer a lender, at no interest in most cases, to the company that is trying to build.

The bull case

In plain words: demand for AI capacity is larger than supply, the largest buyers have signed multi-year commitments, and Nebius is among the few independent builders with power, chips and balance-sheet access to deliver it. The CEO’s own account of this is in the second-quarter shareholder letter, dated 12 August 2026: new deals priced at $20M to $25M a megawatt, payback of one year and ten months, and, in the letter’s words, “We could sell our entire 2027 capacity on these terms today.”⁸ NVIDIA, which holds 9.3%, said when it invested that Nebius is building “an AI cloud designed for the agentic era”, in a quote from its chief executive, Jensen Huang, dated 11 March 2026.²⁰²¹

On the holder side, Situational Awareness LP, the fund run by Leopold Aschenbrenner, held 4,464,260 Nebius Class A shares at 30 June 2026, according to its 13F filed on 14 August, worth $1.23B at that quarter-end price and 6.1% of the portfolio that filing reports.³⁷ That is a dated position, not a statement by anyone about the company. The same group also filed a Schedule 13G on 27 May 2026 that reported 12,410,060 Class A shares, 5.6% of the class, at 19 May 2026.⁴³ The two filings are six weeks apart and use different tests, and neither says why the numbers differ. I cannot tell from them whether the fall is a sale or a difference in what is counted, so I draw no view from either. A 13F does not show shorts or holdings outside the US, and I draw nothing from it about what the fund holds today or will do. On the sell side, according to a TIKR summary, Bank of America’s Tal Liani reiterated a positive rating with a $310 target on 22 October, citing customer prepayments and sector selling that had no company-specific cause; the same summary says Piper Sandler cut its target to $185 on 21 October, and StockWireX lists that rating as neutral.²⁴²³ Short interest was about 20% of free float in October, per the StockWireX summary of Yahoo Finance and ORTEX data, which I could not check.²³

What the podcasts and interviews say

I looked for what the people whose judgment you might want have said about Nebius and the neocloud model in public: the four hosts of the All-In podcast, Gavin Baker, Brad Gerstner and Bill Gurley, and Volozh himself. For the All-In hosts, Baker and Gurley I found no episode, transcript or show note about Nebius or neoclouds that I could open and date, so I name nobody and say nothing on their behalf. For Gerstner I found nothing about Nebius that I could open and date, so he is left out too. I searched again on 31 October 2026 and the result did not change.

For Volozh I found three interviews, and I could open none of the audio or video. They are all secondhand or thin, and I have marked them so. The Accel “Spotlight On” podcast, season 3, episode 5, has no date on its page and lists a chapter at 14:20 called “The secret to building infrastructure: technology and capital”, but no transcript. It is hosted by Matt Weigand of Accel, who is also a Nebius director, so it is not an outside view.²⁹ A Substack summary dated 9 May 2026 describes a podcast it does not name, and quotes him saying that wholesale bare-metal rental is “the business of a typical NeoCloud”, and that Nebius is “not there”, building layers on top.³⁸ That is the claim to weigh. In the filings the two big hyperscaler contracts are for “dedicated GPU infrastructure capacity”, and in the filings I read the company does not split revenue between rented capacity and software.⁹¹⁷² Nebius may well be building layers, but I cannot see them in the numbers yet, and most of the contracted book is the kind of rental he says it is not. In the same summary he puts unbacked debt for a young company at 10% to 12% and the cost with a hyperscaler behind the contract at about 6%.³⁸ The July facility at SOFR plus 2.50% is consistent with the second figure.²²

The third, an excerpt from a video interview recorded at NVIDIA’s GTC in Berlin and circulated on X on Sunday 25 October, is relayed by 24/7 Wall St. and Yahoo Finance. He is reported to have said that some capital sources, asset-backed and debt markets, were temporarily closed, that “We have a pretty healthy balance sheet”, and that customer advances and other instruments would probably fund at least one gigawatt.²⁷ Checked against the filings: the first is consistent with the October bridge, which is secured on substantially all the assets, costs more and falls due in a year.²⁶ The second is true of the 30 June balance sheet, where cash of $8.0B was nearly the size of the convertible debt, though debt and prepayments have grown since.³ The third is a fair reading of the prepayment data, since one gigawatt at my base cost of $38M a megawatt is $38B and prepayments of more than $9B would cover about a quarter of it by themselves C (calculated).⁸ I could not find the original episode, its host or any timestamp, so treat all three as reported speech, not as confirmed.

Against reality

Today, revenue is up 454% on the year and adjusted EBITDA was $236.2M in the second quarter.³ Contracted revenue is $37.5B and capacity is arriving. If Nebius connects 800 MW to 1 GW by the end of 2026, builds about 650 MW a year after that, prices new capacity near $21M a megawatt a year in 2027, easing to about $16M by 2035, and raises about $36B in debt at 12% or less, my base is possible: $134 a share. The price is 15% above that, and a long-run income on new equipment of 31.1% would justify it, against my 30% C (calculated). So the distance between today’s numbers and my base case is modest. The distance between today’s numbers and the case where the price is right is a build that has not happened, financed with money not yet raised.

What a 10x would take. Ten times $154.14 is $1,541 a share. With every convertible note converted, that is about 393M shares and a market value of about $605B. At 25 times owner cash that needs about $24B a year of owner cash. At the 17% of revenue my base earns after replacing equipment, that is about $141B of revenue, against $530M in 2025, which is 266 times and 1.8 times my base revenue for 2036 C (calculated).² Ten years is also the period in which NVIDIA’s shares rose 115 times and Tesla’s 30 times, from market values of about $38B and $30B C (calculated).³⁹ Nebius starts at about $49B. Ten times from here is 3.8 times my bull. For a business that sells a rented, depreciating asset to a few buyers, I think tenfold is close to impossible, and I would not build a case on it.

Where I come down

What moved me was the financing. A signed contract from Microsoft or Meta that supports secured debt at a margin of 2.5 points over the benchmark is a different business from one that has to sell shares, and I gave it credit in the cost of money. What did not move me is the speed of the revenue, because the model already takes the whole of the company’s 2026 guide. The two things I most want to see are a third-quarter letter that shows new capacity still priced at $20M a megawatt or better, and the bridge refinanced on better terms. If both happen, I would raise my base, mainly by lifting the long-run income on new equipment; each point of it is about $19 a share.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• Delivery. Microsoft may terminate a tranche if Nebius is late after a grace period, and the first Meta order has the same term.⁹¹⁷ Nebius says it has delivered all Microsoft tranches to date.⁸ The Meta order of March 2026 starts in early 2027 on a new chip, NVIDIA’s Vera Rubin.¹⁰⁴⁰
• Chips and power. Three suppliers made up the equipment spending at 30 June, up from two at year end.⁴ The company has signed a partnership with Bloom Energy to deploy fuel cells, one of several ways it is trying to get power quickly.⁸ What a one-year delay would do to the value is in the geopolitical section below.
• Controls and reporting. The auditor’s opinion on internal control was adverse for 2024 and 2025, there are two material weaknesses, the Dutch statutory accounts were late, and the auditor is changing.²⁴¹³¹ A restatement of the equipment accounts would matter more than at most companies.
• Control and dilution. One shareholder holds 51.6% of the votes, and the board can issue 20% more shares for five years without offering them to others first.²²⁸ The older notes were struck at much lower prices than today.⁴
• Tax left over from the divestment. Nebius paid $181.5M of Dutch dividend withholding tax in the first half of 2025, a cost of the Russian divestment, and the 20-F covers the matter.¹⁶²
• Avride and TripleTen. Together they lost $94M at adjusted EBITDA in the first half. Management said in the 20-F it is exploring outside investment into Avride, including a deal in which it might cede control.¹⁶²
• Accounting choices. The longer server life, the capitalized interest of $40M in one quarter, and the revaluation gains in profit all make reported results look better than cash.²⁴

Geopolitical risk, and what it means for the value

Until July 2024 this was Yandex N.V., and more than 95% of its revenue, assets and employees were in Russia. It sold all of them for $2.6B in cash and 162.5M of its own shares, and says it now has no customers, suppliers or partners there.² The past still costs something: $181.5M of Dutch withholding tax on the share part of the price, and, the 20-F warns, slower onboarding and know-your-customer checks with banks and counterparties.² The Dutch domicile matters little to value today. Dividends would bear 15% Dutch withholding tax, but my base pays none inside ten years.²

The live exposures are elsewhere. Nebius owns a site in Finland and rents or builds in France, Iceland, the UK, Israel and the United States, where Microsoft’s capacity is in New Jersey.² The 20-F says it is watching export controls on chips and a proposed US rule that would make cloud providers check foreign customers and let Washington restrict some deals.² It also warns that the conflict involving the United States, Israel and Iran could disrupt its Israeli sites and people, but gives no size for them.² Commerce decides the first, Israel’s security the second, and grid operators and permit bodies the third, power. Contracted power is a right to build, not electricity at the building.⁸

I do not predict events. The odds below are my judgment and rough: the filings name each risk and give no probability, and Nebius has so far delivered every Microsoft tranche on time.⁸ The price allows for none of them by name. At $154.14 it sits 15% above my base, and my 12% discount rate carries ordinary risk, not a specific event.