An interactive guide · 18 November 2026Independent, not affiliated with Nebius Group · Not investment advice

Nebius, explained

Nebius is a vertically integrated AI cloud. It designs its own servers, builds and runs its data centers, writes the cloud software that sits on top, and sells computing power directly to the companies that use it, from raw chips by the hour to finished AI answers.This page takes the business apart one contract and one site at a time, then puts it back together into an illustrative fair value per share. Every number on the page moves when you change an assumption, and at the end you can download your own version of the model as an Excel file, with formulas.

Questions or corrections? Write to nebiusexplained@duck.com

Assumptions:
In one minute

The short version

WHAT IT IS

A full-stack AI cloud

Not just a site developer or a GPU reseller: Nebius designs the servers, builds and runs the data centers, writes the software and sells to the end customer, from chip-hours to finished AI answers.

THE PLAN

Borrow on blue chips, earn on everyone else

Microsoft and Meta signed about USD 32b of low-priced contracts that banks lend against. With the network built, enterprises now pay more than twice the price per MW and prepay more than the full cost of the capacity, while spot sells every idle hour.

THE SCALE

From USD 0.5b to tens of billions

Revenue was USD 0.5b in 2025 and about USD 2b in the first nine months of 2026. About 4 GW of announced sites come online by 2029; management guides USD 12-13b of revenue for 2027 alone.

THE FUNDING

Customers and lenders pay for most of it

Customers committed about USD 10b of prepayments in 2026 and should commit USD 17-20b in 2027; with loans on investment-grade contracts, that covers 75-85% of 2027 capex.

THE CUSTOMERS

The money comes from paying customers

545+ companies across every industry, from Snowflake, Databricks and Datadog to Revolut and Mount Sinai. Nvidia's money is about 6% of what was raised; Microsoft and Meta own no shares.

THE TEAM

Owners who built a cloud before

The people who ran Yandex's infrastructure for two decades run Nebius. The founder owns about 12% and insiders about 16%, with only small, regular sales, so management gets rich the same way shareholders do.

Behind the funding sits a strong balance sheet: USD 7.4b of cash at the end of September, stakes worth several billion, and over USD 40b of investment-grade contracts still to borrow against.

Section by section

1 · IdeaNebius does every layer of the AI cloud, so it keeps the margin each layer would otherwise pay someone else.
2 · ContractsLow-priced Microsoft and Meta deals financed the network; enterprises now pay more than USD 25m per MW and prepay the full cost.
3 · SitesAbout 4.3 GW across 26 dated sites by 2029, almost three quarters of it in the US; delays are the main operating risk.
4 · Chip lifeA data center lasts 20 years and lives several chip generations; refilling it keeps revenue growing.
5 · Who rentsEach site splits between enterprise (prepaid, steady) and spot (highest price); the base case is half and half, leaning to 70% enterprise from 2028.
6 · CompanyFrom USD 0.5b of revenue in 2025 to tens of billions; see how projections sit against guidance.
7 · FundingPrepayments and contract-backed loans cover most of the spending; the convertible bonds mature after the build-out, when cash is flowing.
8 · Customers545+ companies in every industry; Nvidia's money is about 6% of what was raised, so the demand is not circular.
9 · DealsAbout USD 0.9b of software acquisitions bring millions of developers and a faster inference stack.
10 · SpotA public spot price squeezes GPU brokers and sends their customers straight to the owner of the chips.
11 · Open modelsTokens get cheaper, GPU-hours dearer; as enterprises adopt open models, spending shifts to whoever runs them.
12 · RegulationNo Nebius MW in the highest-risk US states; new rules ask data centers to pay their way, as Nebius already does.
13 · StakesClickHouse, Avride, Toloka and TripleTen add value per share and could be sold or pledged for cash.
14 · ValuationSum of the parts on the 12 months after each snapshot, with bonds converted where it pays holders, discounted back to today.
15 · PeopleA team that ran one of the largest infrastructures outside the US, owning about 16% of the company.

Independent work, not affiliated with Nebius Group. For education only, not investment advice.

1 · The idea

The AI cloud company

Nebius does every layer of the AI cloud, so it keeps the margin each layer would otherwise pay someone else.

Since 2023 a wave of new companies has started selling AI computing power outside the giants that build their own (Amazon, Microsoft, Google, Meta and SpaceX). They tend to be grouped together, yet they do very different jobs, and most handle only one part of the chain. Nebius handles every part.

POWER AND BUILDINGS

The site developers

Often former bitcoin miners. They own land and electricity connections and lease finished data centers to someone else, usually a hyperscaler or another neocloud. They earn rent on a building; they do not run the computers.

CHIPS FOR A FEW BIG CLIENTS

The GPU renters

They rent space in someone else's data center, buy chips, and lease them out, mostly on long contracts to a handful of very large customers. Little software of their own, so the price is what wins the deal.

CAPACITY IN BETWEEN

The brokers

Marketplaces that resell other people's spare chips by the hour. Light on assets, but they own neither the machines nor the customer relationship for long.

EVERY LAYER

Nebius

Designs its own servers and racks, builds and runs data centers (owned or leased), writes its own cloud software (storage, managed clusters, AI inference), and sells to thousands of AI labs, start-ups and enterprises, as well as to Microsoft and Meta.

What each doesSite developersGPU rentersBrokersNebius
Land, power and buildings✓rented–✓ owned and leased
Own server and rack design–––✓
Buys and runs the chips–✓–✓
Own cloud software on top–somesome✓ full platform
Sells finished AI answers (inference)–––✓ Token Factory
Customersa few tenantsa few giantsmany, short-livedgiants and thousands of others

Doing every layer costs more up front, but it means Nebius keeps the margin each layer would otherwise pay someone else, can charge for software and service rather than only for chips, and can rent the same data center to whoever pays best. That is why the rest of this page values Nebius customer by customer, not as a building with tenants.

How one megawatt turns into money

Data centers are measured in megawatts (MW): the electricity they can draw. One MW powers roughly 570 of today's top AI chips. Whatever the customer buys, from raw chips to finished AI answers, every MW Nebius builds runs through the same six steps.

STEP 1 · BUILD
–

to fit out one MW: about for chips and servers, plus for the building if Nebius owns it. Paid before the first customer arrives.

STEP 2 · FINANCE
–

of that comes from somebody else. Enterprise customers now often prepay the full cost, building included. Of what they do not prepay, lenders fund against the signed contract and the machines, at about a year. Only the rest is Nebius's own money.

STEP 3 · RENT
–

a year in rent from an enterprise customer for that MW. Big tech pays less, hourly customers pay more but may leave.

STEP 4 · RUN
–

a year for power, cooling and staff, plus building rent where Nebius leases the space. The rest is cash profit, which repays the loans and their interest first.

STEP 5 · DEPRECIATE
–

a year of depreciation: the chips are written down over , the building over 20. It is not a cash cost, but it is what profit has to cover so the equipment can be replaced.

STEP 6 · WHAT IS LEFT
–

is still owned at the end of the contract: old chips worth about of their cost, and a building with most of its life ahead. Refilled with new chips, the data center earns again with no new building to pay for.

Money in and out of 1 MW, year by year

Cumulative cash for one MW under three kinds of customer. The line starts below zero (the build), then climbs as rent comes in. Where it crosses zero, the money is back.

In one sentence

A deal is good when the rent over the chips' life pays back the building and the chips, with something left over. It is excellent when the customer pays for the chips before they are installed, because then Nebius never has its own money at risk.

That is the whole game. The famous contracts with Microsoft and Meta are the foundation; the money is made with everyone else.

2 · Contracts

The contracts, in the order they came

Low-priced Microsoft and Meta deals financed the network; enterprises now pay more than USD 25m per MW and prepay the full cost.

The contracts follow a plan in three phases. Nebius first proved its cloud with its own chips, rented by the hour. It then signed two giant, low-priced contracts with Microsoft and Meta, whose investment-grade credit let it borrow cheaply and finance the build-out. With the network built, it now puts the capacity to work with customers who pay more: enterprises that prepay, partners who buy the software alone, and spot customers who fill every idle hour. No new hyperscaler contract is planned. Walk through each chapter in order, or jump to any of them. In each one, move the sliders: every slider is one assumption, with a line on why its starting value was chosen. What you change here flows into the whole company further down.

1
2024 - 2025 · PROVE THE CLOUD

Own chips, rented by the hour

  • First fleet in Finland, Paris, Iceland, Kansas City
  • Software and operations tested with paying customers
  • Small, flexible, unfinanced
Price per MW a year: about USD 13m, older chips
→
2
2025 - 2026 · BUILD ON BLUE-CHIP CREDIT

Microsoft and Meta

  • About USD 32b signed: Microsoft USD 17.4b, Meta USD 3b + 12b
  • Low price, about USD 11-16m per MW a year
  • Investment-grade credit that banks lend against
Price per MW a year: about USD 11-16m
→
3
2026 ON · PUT THE CAPACITY TO WORK

Everyone else, at better terms

  • Enterprises: above USD 25m per MW, prepaying the equipment
  • Software-only partners: fees with no chips to buy
  • Spot: every idle hour sold at the market price
Price per MW a year: above USD 25m

–

Money in and out each year: what the customer pays, and what Nebius spends on the building and chips.

RevenueBuilding and chips (below the line)Customer prepayment

What it earns, and when the money is back

Bars: cash profit each year (revenue minus running costs). Line: all money in minus all money out so far. Where the line crosses zero, Nebius has its money back.

Cash profit (EBITDA)Money back so far (cumulative cash)

Scoreboard: every contract on the same 100 MW

Same size, so they can be compared. Click a row to open it in the lab.

3 · Sites

Each site and their build pace

About 4.3 GW across 26 dated sites by 2029, almost three quarters of it in the US; delays are the main operating risk.

Nebius does not open a data center in one go. Each site comes online in steps, as power is connected and data centers are finished. Below is every announced site, with the power planned each quarter. Press play to watch it grow, or delay any site to see what a late building does to revenue and value. The New Jersey site is a reminder that this happens: its third-party builder ran late in 2026, so a little more than half of the Microsoft capacity will be live by year-end and the rest in Q2 2027. Nebius says its own sites are on time or ahead of schedule.

Contracted, connected, in production

1 · CONTRACTED
5 GW

Power Nebius has secured, through land and grid agreements or colocation leases: more than 4 GW today and 5 GW expected by the end of 2026. The sites on this page add up to about 5.2 GW at full build-out. A contracted MW earns nothing yet.

2 · CONNECTED AND BUILT
4.3 GW

Capacity the tracker and announced deals expect to be connected and built by end-2029, with grid power, building and cooling in place. Nebius plans to add more than 1 GW a year, with owned sites making up over half of the next gigawatt and over three quarters of the one after.

3 · IN PRODUCTION
about 0.26 GW

Live and earning at the end of September 2026 on the tracker's timeline, after the biggest quarterly increase in active capacity so far. Only this layer produces revenue. The gap between the three is the growth still to come, and the reason timing matters more than demand.

26:Q3
–
MW live at the date chosen
–
MW under construction (live within a year)
–
MW planned after that, to end-2029
–
Fair value with these delays

Power live by region, quarter by quarter

LiveUnder constructionPlannedOwn / Lease: does Nebius own the building (pays for it once) or rent it (pays every year)?

Site capacities, dates and building type (owned, colocation, build-to-suit) come from the Nebius tracker's data center timeline and deep dive (nebius-tracker.vercel.app, read 18 Nov 2026). Build-to-suit and colocation sites are treated as leased: a third party owns the building and Nebius pays rent. Frankfort, Mäntsälä II and Rotterdam are announced with no dates yet, and the larger 'possible' capacity at some sites is not in the plan. Change any of it.

4 · Chip life

What a chip earns over its life

A data center lasts 20 years and lives several chip generations; refilling it keeps revenue growing.

A data center lasts 20 years; the chips inside it last about . Every data center therefore lives several lives, one per generation of chips. These are the assumptions about chips that the rest of the page uses: how fast rents fall as chips age, what old chips are worth at the end, and what the next generations cost and earn. They are set here, and only here.

Pick a view of chip prices:
1 · THE RENT FALLS

A chip earns most when it is new. Spot prices on older chips fall about a year as newer models arrive; enterprise customers lock in a price for , then renew lower.

2 · THE VALUE WEARS OUT

Accountants spread the cost of the chips over their life: per MW a year of depreciation: not a cash cost today, but the money to replace the chips has to come from somewhere.

3 · THE CHIPS RETIRE

After the chips are sold for scrap ( of cost) and the data center goes dark, unless Nebius buys a new set. That costs about per MW again, and the cycle restarts.

One MW of chips over its life

Rent earned each year by enterprise and spot customers, against the depreciation and what the chips are still worth on the books.

Company revenue by the year the chips were bought

Each color is one generation of chips. Each band shrinks as it ages and vanishes when it retires, and a new band (a refill) takes its place. The dashed line shows what would happen if no data center were ever refilled: the gap between it and the bars is what refilling brings.

Chip life, scrap value and how fast rents fall are the biggest moving parts on this page. The cautious view assumes rents fall fast and old chips are worth little. For years, older chips have held their value better than that, and some are being re-contracted today above what they earned when new, so the base here is steady. New-generation trends apply to chips bought after the 2026-29 plan.

5 · Who rents what

Decide who rents each site

Each site splits between enterprise (prepaid, steady) and spot (highest price); the base case is half and half, leaning to 70% enterprise from 2028.

The signed contracts come first: Microsoft and the two Meta deals each take a fixed amount of power at a given site. Everything else is still to be sold. For each site, choose how its new capacity is split between enterprise customers (prepay, steady), spot customers (rent by the hour, pay most, can leave), or unsold (the building waits, no chips are bought).

Left: all spot (higher price, no prepayment, can leave). Right: all enterprise (prepaid, steadier). Each site can still be set on its own below.
Nebius now makes prepayment the standard structure and chooses the mix to fund the build, so capacity going live from 2028 leans further towards enterprise: 50% becomes 70%.
Or a preset:

Power in use, by kind of customer

Revenue by kind of customer

6 · The company

Add it all up

From USD 0.5b of revenue in 2025 to tens of billions; see how projections sit against guidance.

Every data center at every site, rented to whoever you chose, using the contract terms from section 2. Head-office costs (engineers, sales, admin) come off on top: of revenue. The heavy building years come first; the cash follows.

The strategy behind the mix

Six ideas to keep in mind when reading the numbers below.

01THE UNIT

Every contract is a small business

Money goes out first, for the building and the chips. Rent and service fees come back over the following years.

When the contract ends, Nebius still owns the data center, and often chips with life left to rent again.

02THE FOUNDATION

The famous deals were funding deals

Microsoft and Meta are the least profitable contracts.

They let Nebius borrow cheaply against blue-chip credit, build its network and win everyone else. No more are planned.

03THE PROFIT

The money is made with everyone else

Enterprises that prepay the full cost put none of Nebius’s money at risk.

Spot customers pay the highest price and fill idle capacity; software-only deals show the platform sells on its own.

04THE MIX

Taken from Nebius's own guidance

Nebius guides to roughly half enterprise, half spot on new capacity. That is the base case here.

Change it per site in section 5 to see how much it matters.

05THE OWNERS

Managers with large stakes of their own

Insiders own about 16%, the founder about 12%. Their sales are small: tens of thousands of shares a quarter against a stake of about 30 million.

Expect management to slow down rather than sign bad contracts to hit a quarterly number.

06THE CAVEAT

Treat the years after 2030 with caution

The site plan stops at 2029, and today’s figures are weighed down by the early hyperscaler deals.

Growth after 2029 comes only from the theoretical capacity set below: a what-if, not a forecast.

The site plan stops at the end of 2029, so without this the company stops growing. This adds new owned data centers every year from 2030 to 2035, rented to the same enterprise / spot mix and built at the same cost per MW. It is a what-if, not a plan: for scale, the sites already announced could hold close to 1,000 MW more than is scheduled. It does not change the fair value, which is set on the business at the end of 2027.

Revenue and profit

Cash: spending, prepayments and what is left

2025 and January to September 2026 use the revenue and adjusted cash profit (adjusted EBITDA) Nebius reported, quarter by quarter: 2025 revenue USD 529.8m and adjusted EBITDA of -USD 64.9m; 2026 to September revenue USD 1,962.1m and adjusted EBITDA USD 831.5m. Columns marked A are reported: 2024 and 2025 come from the FY2025 results (continuing operations; profit after depreciation is adjusted EBITDA less depreciation, net profit is as reported, and capex, prepayments and cash are from the cash flow statement). Growth for 2026 9m A compares with the first nine months of 2025 (USD 302.1m). Shaded columns: lavender = reported; grey = 2030 onwards, which includes theoretical capacity not yet announced; yellow = 2026 estimate, which adds the page's Q4 to the reported first nine months (2026 capex: USD 15.4b reported to September). Prepayments are not reported as a separate line: for 2026 9m A the table uses the rise in customer prepayments not yet earned (deferred revenue), from about USD 1.6b at end-2025 to USD 9.4b at end-September, and for 2026E management's expectation of about USD 10b of prepayments in 2026. Depreciation is not reported for nine months, so that cell is blank. The site plan stops at the end of 2029. After that, each data center is refilled with new chips when the old ones wear out, and data centers freed by Microsoft and Meta are rented to the same mix as the rest of their site: section 4 explains why that matters.

Management's 2027 guidance, and how this page compares

With the Q3 2026 results, Nebius gave its first guidance for 2027. Management has guided very conservatively, and the assumptions needed to match its numbers show how much. Initial guidance from a company this capacity-constrained is best read as a floor: it covers only capacity Nebius is confident it can deliver on time, after a year in which a third-party site ran late and some 2026 revenue slipped into 2027.

Guidance case or the page's own base case

The page opens with settings adjusted so it matches all of management's guidance at once; the table below checks each item. Switch to the page's own base case to see what the sites and contracts add up to without that caution.

7 · Funding

Who pays for the building

Prepayments and contract-backed loans cover most of the spending; the convertible bonds mature after the build-out, when cash is flowing.

Nebius spends years ahead of the rent. The gap is filled, in order, by customers who prepay, by lenders who lend against signed contracts and equipment, by cash in the bank, and only then by selling new shares. How much lenders will advance decides how many new shares are needed.

How the money was raised, in three steps

The order in which money was raised matters as much as how much was raised. It shows a finance team keeping its cheapest, most reliable source for last.

MOVE 1 · 2025-26 · SECURE

Sign investment-grade contracts

Microsoft and Meta sign about USD 32b of multi-year contracts. Both are investment-grade companies, so their promise to pay is something banks will lend against, at rates far below what Nebius could get on its own. Nebius gave them a low price; in return it got their credit.

MOVE 2 · 2025-26 · DON'T TOUCH IT YET

Raise elsewhere while the window is open

Instead of borrowing against those contracts straight away, Nebius sold about USD 14b of convertible bonds and some new shares while investors were eager to buy them, on cheap terms. Only USD 0.8b was borrowed against the contracts. The best collateral stayed in the drawer.

MOVE 3 · 2027 ON · DRAW WHEN OTHERS CLOSE

Use the reserve when markets turn cautious

Markets are now warier of lending to Nebius on its own name. But lenders still want Microsoft and Meta credit, and that capacity is almost untouched: about USD 21b could be borrowed against those contracts. It is the money that carries the 2027 peak, without depending on the mood of the stock market.

The point: a company that had borrowed against its best contracts early would now have to raise money when it is expensive. Nebius did the opposite, raising when it was cheap and keeping the safest source in reserve for exactly this moment. Enterprise customers who now prepay more than the cost of the equipment add a third source that needs no lender at all.

Cash needed each year, and where it came from

The least available source today is loans secured only on chips with no customer contract behind them: lenders now price GPU debt on the contract, so the page assumes they advance just 25% on spot capacity. 2024 to September 2026 are reported: cash used after capex plus the cash added to the bank, against the money raised (convertible bonds at face value, USD 3.75b in 2025 and USD 9b in 2026; Nvidia's USD 2b and the USD 700m placement; the USD 775m secured loan). "Other and balance" covers the 2024 sale of the Russian business and items not broken out, such as other share sales. From Q4 2026 the bars are this page's projection.

Net debt compared with a year's cash profit

Debt minus cash at each year-end, divided by that year's cash profit (EBITDA). The line takes the convertible bonds out of debt, as if they all turned into shares.

Reading the funding

So far Nebius has raised almost exactly what its projects needed, the small gap coming from cash in the bank. From 2027, nearly all of the gap is expected to be prepaid by customers or borrowed against what is already signed or built. Lenders funding 70% of contracted equipment not prepaid is a normal ask for secured financing; on spot capacity, with no contract behind it, the page assumes only 25%.

If new shares are sold, it is because spending runs ahead of revenue during the build-out, not because the business loses money. The cautious guidance case, with data centers opening later and filling more slowly, needs more of them than the page's own base case, where the heaviest year is 2029. The first loan repayments also fall due by then (the one-year Citi facility is assumed not to be rolled over), and secured lenders like the borrower to put in a deposit of its own. How much depends on that deposit and on how much debt lenders accept relative to cash profit; under five times is assumed, well below peers.

The balance sheet at each year-end

Dec-24, Dec-25 and Sep-26 are the reported balance sheets (2024 restated for the Toloka deconsolidation, debt at carrying value); "other" lines are the reported totals less the lines shown. From 2026 the figures are this page's projection.

Every bond and loan raised so far

Convertible bonds pay a low interest rate because the buyer can swap them for shares if the price rises above the conversion price. Each was sold at a premium of 40-58% to the share price of the day. Two of the newer series also grow their principal until maturity (accrete), so the price at which they turn into shares rises over time.

RaisedInstrumentOutstanding (USD m)Interest rateInterest a year (USD m)MaturesTurns into shares at (USD)Note
Jun 2025Convertible bonds, two series2002.00% / 3.00%52029 / 2031about 51USD 1.0b issued; USD 800m swapped for about 15.8m shares in Aug 2026
Sep 2025Convertible bonds1,3751.00%142030about 13950% premium to USD 92.50; principal accretes, about USD 160 effective at maturity
Sep 2025Convertible bonds1,3752.75%382032about 139
Mar 2026Convertible bonds2,2501.25%282031183.22Premium of about 57.5% to USD 116.33
Mar 2026Convertible bonds1,7502.625%462033180.31
Aug 2026Convertible bonds3,4500.50%17Feb 2030313.4640% premium to USD 223.90; principal accretes to 110%, about USD 345 effective at maturity
Aug 2026Convertible bonds2,3004.50%104Feb 2034324.6545% premium to USD 223.90; principal accretes to 125%, about USD 406 effective at maturity
All convertibles12,7002.0% cash
4.0% with accretion
252
+ about 250 accretion
Accretion is extra principal repaid at maturity (or converted), not cash paid each year
Oct 2026Citi senior secured facility3,500SOFR + 2.75%, up to 3.50% (about 6.7% today)233364 days–A bridge: Citi is lead arranger, agent and security agent; Nebius points to its over USD 40b of long-term contracts
Jul 2026Asset-backed loan on an investment-grade contract775SOFR + 2.50% (about 6.5% today)50––The first facility of its kind; the Microsoft and Meta contracts, and more than USD 40b of similar committed revenue, can be financed the same way
All interest-bearing debt above16,9753.2% cash
4.6% with accretion
535
about 785 with accretion
Weighted by amount outstanding

Two things stand out. The whole debt costs about 3% a year in cash (USD 0.47b), or about 4.5% counting the principal that some convertibles add each year, far below what a young, fast-growing company would pay on ordinary bonds. And each series turns into shares only above its conversion price, set 40-58% above the share price on the day it was sold: if the shares never get there, the bond is repaid as debt and creates no new shares. Option sizes on the August 2026 notes (up to USD 750m more) are not included. The June 2025 bonds are assumed split evenly between the two series. SOFR is taken at about 3.9%, its 30-day average in mid-November 2026. Sources: Nebius pricing releases of 10 Sep 2025, 17 Mar 2026 and 19 Aug 2026, and of the Citi facility on 22 Oct 2026; June 2025 terms from Nebius's 6-K.

What if the shares stay below the conversion prices?

With the shares trading below most conversion prices, a common worry is that the bonds become a debt wall that forces Nebius to sell shares cheaply. The timing says otherwise.

The bonds fall due once the data centers are earning. Almost nothing matures before 2030, and each year's repayment is a fraction of the cash profit the business makes by then.

A low share price actually helps shareholders here. If the bonds do not convert, Nebius repays about USD 12b of debt instead of handing over about 61m shares, which would be worth far more than USD 12b at the fair value on this page.

And repayment is the worst case, not the likely one. The page assumes each bond is either converted or repaid from cash. In practice a maturing bond is usually refinanced with a new bond or a loan, as Nebius has done repeatedly. The risk to watch is a late build-out, not the bonds: delay every site by two quarters with the shares still low, and the years of negative cash stretch further into the largest repayments. That is a question about delivery, covered in section 3, not about the bonds.

8 · Customers

Who the customers are, and where the money comes from

545+ companies in every industry; Nvidia's money is about 6% of what was raised, so the demand is not circular.

A common charge against AI clouds is circular financing: a chip maker invests in a cloud, the cloud uses that money to buy the chip maker's chips, and the same money goes round in a loop with no paying customer at the end. The test is simple: follow the money. At Nebius, the cash comes overwhelmingly from paying customers, bond buyers and banks that have nothing to do with Nvidia.

545+
companies found using Nebius by an independent tracker
57%
of them (313) never announced by Nebius itself
202 / 112
on the AI Cloud / on Token Factory (inference)
USD 12b+
signed since mid-2026, incl. multiyear deals with Snowflake, Databricks and Reflection

What the customers do

Companies in the Nebius Customer Atlas, by category. Most are not AI labs at all, but businesses putting AI to work.

Source: Nebius Customer Atlas (nebiusatlas.com), an independent directory not affiliated with Nebius, read 18 Nov 2026. A listing can be a large contract, a small account or a benchmark: it counts customers, not dollars.

A few of the names

Big tech and large companies

MetaMicrosoftDatabricksSnowflakeCloudflareDatadogPalantirShopifyAccenturePorscheGEICOProsus

AI labs and model builders

MistralCohereReflectionBlack Forest LabsLumaWorld LabsRecraftPhotoroomPrime IntellectHiggsfield

Software, finance and health

CursorCognitionJetBrainsBraveRevolutbunqMount SinaiSword Health

Robotics and research

1XRoboForceMultiply LabsCaltechPrincetonStanfordCambridgeUniv. of Washington

Source: Nebius Customer Atlas and Nebius customer stories.

The big enterprise deals of 2026, and their terms

After the hyperscaler deals, the contracts got bigger and better. Enterprise customers now sign billion-dollar deals at more than twice the hyperscaler price per MW, and prepay all of the equipment.

WhenWhoSizeYearly price per MWPaid up frontNote
Sep 2025Microsoft (hyperscaler)USD 17.4b, 5 yearsabout USD 11.6mabout USD 7.0b (40% of the contract)300 MW at Vineland; the benchmark the enterprise deals are compared with
Q2 2026Four AI cloud customersover USD 1b each, on averageUSD 20-25m50-60% of the capexThe first billion-dollar enterprise contracts
2026Reflection (model developer)over USD 1b––Compute agreement
Q3 2026Snowflake, Databricks, Datadog and other large tech firmsover USD 12b in allabove USD 25mover 100% of the capex, building and networking includedPayback on one-to-three-year deals down to 1 year 7 months; prices per chip-hour up over 20% in the quarter
Q4 2026Not yet namedmulti-billion––Management says multi-billion AI cloud agreements were already signed in Q4

Contract sizes and terms as reported in Nebius results coverage (Q2 and Q3 2026) and the Microsoft 6-K and annual report. Individual sizes for Snowflake and Databricks are not disclosed. The enterprise price used on this page (USD 25m per MW a year, prepaying the equipment) comes from these terms.

What they run on it

From Nebius's own customer stories: production workloads, for companies that sell to their own customers.

Finance

Revolut

The digital bank runs fraud prevention, customer support and model training on 200+ H100 chips.

Read the story on nebius.com ↗

Voice AI

Krisp

Trains its real-time accent-conversion models for call centers on Nebius H100s, which finished training 50–80% faster than the A100s it used before.

Read the story on nebius.com ↗

Model builder

Recraft

Trains its design AI models on Nvidia B200 systems.

Read the story on nebius.com ↗

Health

Sword Health

Runs Dawn, its AI wellbeing service, in production on the AI Cloud.

Read the story on nebius.com ↗

Robotics

RoboForce

Builds industrial robots for hazardous work, training on Blackwell chips and storage.

Read the story on nebius.com ↗

Science

Prima Mente

Pre-trains foundation models of brain activity on a reserved 32-chip cluster.

Read the story on nebius.com ↗

Biopharma

Quantori

Runs a molecule-generation pipeline for drug discovery.

Read the story on nebius.com ↗

Open source

vLLM

The open-source inference engine ran large experiments, including DeepSeek R1, on Nebius.

Read the story on nebius.com ↗

Follow the money

Roughly where Nebius's cash came from since the 2024 restart, and where it sits today, in USD b. The Nvidia flows are in lime. Nvidia has put in about USD 2b, about 6% of the total; Nebius has spent far more than that on Nvidia chips, paid for by everyone else.

WHERE THE MONEY CAME FROM WHERE IT WENT Nebius about USD 31.1b Customer prepayments still held · USD 9.4b mostly Microsoft and Meta, for capacity not yet delivered · 30% Revenue from customers, Jan-Sep 2026 · USD 2.0b 545+ companies, from Meta to Revolut · 6% Convertible bond buyers · USD 12.8b seven series, Jun 2025 to Aug 2026 · 41% Other equity investors, Dec 2024 · USD 0.7b Accel, Orbis and others · 2% Banks · USD 4.3b Citi facility and loans backed by signed contracts · 14% Nvidia · USD 2.0b Dec 2024 shares and Mar 2026 warrant · 6% Nvidia chips and systems · USD 13.5b estimate: about 70% of equipment Buildings, power, network · USD 5.8b builders, utilities, other suppliers Cash still in the bank · USD 7.4b 30 Sep 2026 Running costs, interest, other · USD 4.4b the balance
THE BIGGEST PAYERS OWN NO SHARES

Microsoft and Meta are customers, not investors

About USD 32b of contracts, with billions prepaid in cash, from two of the strongest balance sheets in the world. Neither holds a stake in Nebius: they pay because they need the computing power.

NVIDIA IS A SUPPLIER AND A SHAREHOLDER

Nvidia owns 9.3%, but funded little of the build

Nvidia bought about 1.19m shares in Dec 2024 and paid about USD 2b for a warrant over 21.07m shares in Mar 2026, 9.3% in all. That is a vote of confidence, but a small part of the money raised, and Nvidia is not among the contracts the build-out is financed on.

THE REST IS ARM'S LENGTH

Bond buyers, banks, hundreds of customers

About USD 12.75b of convertible bonds sold to ordinary investors, a USD 3.5b Citi facility, and revenue and prepayments from 545+ companies in every industry. None of it loops back to Nvidia except as payment for chips.

How the picture is built: customer prepayments are the deferred revenue on the 30 Sep 2026 balance sheet; revenue is as reported for January to September 2026; bonds are the face value of the seven convertible series at issue; banks are the Citi facility plus the first USD 775m of contract-backed loans; equity is the Dec 2024 placement (USD 700m, of which Nvidia about USD 25m at USD 21 a share) and the Mar 2026 Nvidia warrant (about USD 2b). On the right, property and equipment and cash are from the 30 Sep 2026 balance sheet; the split between Nvidia chips and the rest assumes chips are about 70% of equipment spend, in line with the cost per MW in section 2. The last bar is the balance. The cash left from the 2024 sale of the Russian business is not shown. Figures are rounded and approximate.

9 · Acquisitions

The 2026 acquisitions

About USD 0.9b of software acquisitions bring millions of developers and a faster inference stack.

In 2026 Nebius made four software deals, for about USD 0.9b in disclosed upfront prices. None of them adds much revenue on its own. What they add is the software layer on top of the chips, and a community of millions of developers who start on Nebius instead of on someone else's cloud.

AnnouncedCompanyWhat it doesPriceWhat it brings to Nebius
Feb 2026Tavily (Israel)Web search built for AI agents: an agent asks a question, Tavily returns fresh, sourced facts from the webabout USD 275m, up to USD 400m with milestonesMore than 1m developers at signing, over 2.5m later in 2026, according to management; 3m+ SDK downloads a month; about 300m requests a month by Aug 2026; Fortune 500 customers including IBM. Priced at roughly 98x its 2025 revenue
May 2026Eigen AI (San Francisco)Inference optimization: makes open models run faster and cheaper on the same chipsabout USD 643m, cash and sharesA team from MIT's HAN Lab, Nebius's first Bay Area engineering office, and Eigen's existing customers, who move onto Nebius capacity. Closed about 40 days after signing
May 2026Clarifai team and IP (US)Compute orchestration: runs many models across many chips efficientlynot disclosedThe core engineering and research team, the patent portfolio and a perpetual license to the technology; founder Matthew Zeiler becomes head of research
Oct 2026Inferize (Israel)Fixes "cold starts", the delay before a model answers its first requestnot disclosedTechnology and engineers folded into Token Factory
THE FUNNEL

Developers first, contracts later

A developer who builds an agent on Tavily's search, or tests a model on Token Factory, is one click from running it on Nebius chips. Small accounts today are tomorrow's enterprise contracts.

THE MARGIN

More answers from the same chip

Eigen, Clarifai and Inferize all make inference cheaper to serve. Selling tokens rather than chip-hours, a faster stack means more revenue per MW, or a lower price that wins more customers.

THE PRICE

Small next to the spending plan

About USD 0.9b upfront, against about USD 27b of capex planned for 2026. The deals are a rounding error in the capex plan, and part was paid in shares.

An illustrative sales impact

A what-if, not a forecast, and not included in the rest of the page: how much yearly revenue the developer community could turn into, if a small share of it becomes paying Nebius accounts. Move the sliders.

Yearly revenue from developers who convert

Prices and figures from the companies' announcements and press coverage (Bloomberg via Globes and PYMNTS for Tavily, Nebius and SiliconANGLE for Eigen AI and Clarifai, Autobound and Privsource for Inferize). Developer counts are Tavily's and Nebius's own. The sales impact is illustrative: the conversion rate and spend per account are not disclosed by Nebius.

10 · The open spot market

A public price for every idle GPU

A public spot price squeezes GPU brokers and sends their customers straight to the owner of the chips.

Since October 2026, Nebius sells its spare GPU hours on a live spot market and publishes the price of every chip in every region. The price moves with demand, between a floor and a ceiling one cent below the on-demand rate, and changes up to every 15 minutes. Customers either set the most they will pay, or follow the market (see the live prices). Publishing the price looks like a small product feature. It is aimed at the middlemen of the GPU market.

How a GPU broker makes money, and what an open price does to it

A large part of the hourly GPU market goes through brokers: companies that own few chips, buy spare capacity wholesale from whoever has it, wrap it in easy software and resell it per second at one flat price, well above what they paid. That works as long as nobody can see the wholesale price.

BEFORE: AN OPAQUE CHAIN Idle GPU ownerssell spare hours on private,unpublished terms GPU brokerbuys wholesale, hides the source,adds its software and a markup Developerpays one flat price perGPU-second, never sees the cost Price: wholesale+ broker margin NOW: AN OPEN MARKET Nebius spot marketlive price per GPU and region,published, updated every 15 min Free, open-source toolsdstack and SkyPilot find the cheapestcapacity, run the job, shut it down Developersees the price, sets a maximum,buys straight from the owner Price: the market,nothing in between
PRESSURE 1 · THE AUCTION

Out-bid in the pool

A broker must buy below its resale price to make a margin, so it can rarely be the highest bidder. In an open market, capacity goes to whoever pays most: often the end user directly.

PRESSURE 2 · THE INTERRUPTION

Capacity can be taken back

A preemptible machine stops when the market price passes the buyer's maximum, or when Nebius needs it for a regular customer. A broker holding spot capacity can lose it at short notice and see it resold to a higher bidder, unless it pays up.

PRESSURE 3 · THE FLAT PRICE

Brokers lose their flat price

Brokers sell predictability: one headline price per GPU. When their cost moves every 15 minutes in public, they either absorb the swings in their margin or pass them on and lose the simplicity that is their product.

PRESSURE 4 · THE BENCHMARK

Other chip owners see what idle hours sell for

Data center owners who now dump spare hours on opaque markets can read the market price off a public page. That weakens the discount brokers buy at.

PRESSURE 5 · THE SOFTWARE

The broker's software, for free

Nebius integrated the open-source dstack in April 2025, and in October 2025 launched a managed server for SkyPilot, an open-source tool that finds the cheapest capacity across many clouds, runs the job and shuts it down. The convenience layer brokers charge for is now free.

THE LAST GAP · COLD STARTS

Why Nebius bought Inferize

Brokers keep one edge: "scale to zero". Use a GPU for 10 seconds, pay for 10 seconds, because their stack loads a model almost instantly. Inferize, bought in October 2026, fixes exactly that delay. Once it is built in, a developer no longer has to choose between a broker's convenience and the owner's price.

Broker or direct, job by job

Illustrative. A broker bills only the seconds used, at a flat premium price. Going direct means paying the spot price, plus a few minutes to start the machine and load the model, plus some work redone after interruptions. Short jobs favor the broker; anything longer than a few hours favors the owner, typically by half or more.

Cost of one job, by how long it runs

Spot pricing rules: Nebius preemptible VM page and docs, and coverage of its 8 Oct 2026 launch. Tools: Nebius blog posts on dstack (10 Apr 2025) and the Managed SkyPilot API server (23 Oct 2025). The broker price and the start-up and interruption overheads are illustrative assumptions, not any one company's figures. Brokers also sell plain computing (CPUs), which is less exposed to GPU prices; this comparison covers GPU work only.

11 · Enterprise AI

Open models, closed models and the price of a token

Tokens get cheaper, GPU-hours dearer; as enterprises adopt open models, spending shifts to whoever runs them.

Companies putting AI to work choose between two kinds of model. Closed-weight models (the frontier labs' best) can only be rented through the lab's own service. Open-weight models can be downloaded, fine-tuned on a company's own data and run wherever it likes. Enterprises are moving fast towards open weights. Fine-tuned on a company's own data, an open model can beat a general-purpose one on that company's tasks, at a fraction of the cost, and the company owns the result. The model developer is still paid, through a license or royalty, but most of the bill is for running the model.

More tokens, cheaper tokens, dearer GPU-hours

Bars: tokens processed per week on OpenRouter, in trillions (log scale). Line: Silicon Data's H100 rental index, USD per GPU-hour.

Sources: OpenRouter usage data and newsletter (to Sep 2026), Menlo Ventures chart of OpenRouter data (week of 26 Oct 2026), Silicon Data H100 neo-cloud rental index. OpenRouter is one marketplace, not the whole market, but it is the largest public record of token volume.

Why the GPU-hour is worth more

Volume. Weekly tokens on OpenRouter went from 0.5 trillion in January 2025 to 128 trillion in September 2026, about 250 times more, and reached about 194 trillion in the last week of October. AI agents drive much of it: agent API keys now use about seven times as many tokens as human ones.

Price per token. The same quality of answer gets about 10 times cheaper each year. Yet the price actually paid per token has nearly stopped falling: down only 6% in 2026 to May, after 39% in late 2025, because buyers move up to better models as the old ones get cheap.

Price per GPU-hour. Renting an H100 cost USD 2.00 an hour on Silicon Data's index in December 2025 and USD 2.88 in October 2026, up 44%, on a chip now two generations behind the newest. Nebius itself raised its on-demand H100 price from USD 3.85 to USD 4.50 on 1 October 2026.

The arithmetic behind it: what a GPU-hour earns is the tokens it turns out times the price of each token. Better software makes the same chip turn out two to three times more tokens, so even as each token gets cheaper, the hour that makes them is worth more. Speculative decoding alone took one H100 from 145 to 424 tokens a second on the same model.

Nebius sells both the hour and the token: Token Factory serves 20 open models on OpenRouter, which carries only a fraction of Token Factory's overall volume: most of it comes from customers buying directly from Nebius. See Nebius on OpenRouter ↗

Why Nebius is well placed

01 · NEUTRAL

It does not compete with its customers

Amazon, Microsoft and Google each build or back their own frontier models. A lab or an enterprise training on their clouds is feeding a rival. Nebius has no model of its own to push: it only sells the compute and the tools.

02 · THE STACK

It owns the open-model stack

Token Factory serves open models by the token; Eigen AI, Clarifai and Inferize make them run faster and start instantly; Tavily gives agents live web search. Because Nebius also owns the chips, it serves open weights at a cost a reseller renting chips cannot match, and keeps the margin.

03 · CONTROL

Sovereignty and data control

An enterprise that fine-tunes an open model on its own data owns the result, instead of renting a closed service. Nebius runs that model on infrastructure in Europe, the US and the Middle East, under rules the customer chooses, and now inside Palantir's secure perimeter.

04 · BOTH CAMPS

It wins whichever side grows

Closed-model builders rent its chips by the thousand (Microsoft, Meta, the AI labs on section 8). Open-model users buy its tokens and fine-tuning. If frontier labs keep growing, Nebius builds for them; if open weights take over, Nebius serves them.

SEP 2026 · PARTNERSHIP

Palantir names Nebius its preferred sovereign AI infrastructure partner

After an integration period, Palantir will bring Nebius compute and inference endpoints inside its enterprise perimeter for commercial customers. The two firms say open models and customer data looped back into them create the smartest domain AI, with better security.

Why it matters. Palantir sells to large companies and governments that will not send their data to a public AI service. It already had the software and the models; the missing piece was the compute underneath. Palantir becomes a sales channel that brings Nebius into enterprises it would take years to reach alone.

Financial terms were not disclosed, and no Palantir revenue is assumed anywhere on this page.

Token volumes from OpenRouter; GPU rental prices from Silicon Data's H100 index; token prices from YipitData and a16z. Survey figures as reported by Techstrong.ai (ETR), Mozilla, Menlo Ventures and the a16z CIO survey. Partnership details from the Nebius and Palantir announcement of 8 Sep 2026.

12 · Regulation

US pushback on data centers, site by site

No Nebius MW in the highest-risk US states; new rules ask data centers to pay their way, as Nebius already does.

Opposition to data centers became a national political issue in 2026. The rules now being written are mostly about who pays for power and water, not about stopping data centers, and Nebius's US sites sit outside the states where the fight is fiercest.

What changed in 2026

STATES

Two statewide pauses

New York paused state environmental permits for data centers of 50 MW or more on 14 July 2026, for up to a year. Texas froze new grid connections on 3 August and later extended the pause to environmental permits. Maine passed a ban on sites over 20 MW, but the governor vetoed it.

TOWNS

A wave of local moratoriums

Trackers count about 1,100 local moratorium records in 47 states, most of them active, and over 800 local opposition groups. On 3 November voters backed limits in 40 of 41 local ballot measures. Ohio, Michigan, Georgia, North Carolina and Indiana have the most.

WASHINGTON

The fight is over who pays

The White House's Ratepayer Protection Pledge (March 2026) asks AI companies to "build, bring or buy" their own power. The House passed a Ratepayer Protection Act 417-3 in September. Some House Democrats called for a national moratorium on 17 November; nothing has been enacted.

Where the risk is, and where Nebius builds

Each square is a state. Outlined squares with a lime tag are Nebius sites, with the MW scheduled by end-2029.

AKAKMEMEWIWIVTVTNHNHWAWAIDIDMTMTNDNDMN · Nebius: Woodbury, 10 MW by 2029MN10 MWILILMIMINYNYMAMAORORNVNVWYWYSDSDIAIAIN · Nebius: Frankfort (no date), 0 MW by 2029INno dateOHOHPA · Nebius: Butler, 1200 MW by 2029PA1,200 MWNJ · Nebius: Vineland, 300 MW by 2029NJ300 MWCTCTRIRICACAUTUTCOCONENEMO · Nebius: Independence, Kansas City, 1225 MW by 2029MO1,225 MWKYKYWVWVVAVAMDMDDEDEAZAZNMNMKSKSARARTNTNNCNCSC · Nebius: Spartanburg, 50 MW by 2029SC50 MWDCDCOK · Nebius: Oklahoma City, 20 MW by 2029OK20 MWLALAMSMSAL · Nebius: Birmingham, 300 MW by 2029AL300 MWGAGAHIHITXTXFLFL
Statewide pause (NY, TX)Most local moratoriums, or a statewide ban passed (OH, MI, GA, NC, IN, ME)Nebius site

How much of the plan is exposed

About three quarters of the power Nebius plans to run by 2029 is in the US. None of it is scheduled in a state with a statewide pause or among the states with the most local moratoriums. The one overlap is Frankfort, Indiana, an announced site with no date yet, which the plan on this page does not count.

That does not mean no local debate: Independence, Birmingham and Butler all drew public opposition. Each has its approvals, zoning or power agreement in place, as the cards below show.

A global footprint, built up over four years

Power scheduled by end-2029, by country (MW)

Power live at each year-end (MW)

From the site timeline in section 3. Announced sites without dates (Frankfort, Indiana; Rotterdam; Mäntsälä II), about 800 MW in all, are not included.

What Nebius does locally

Every large US site comes with a published community plan, public meetings and a commitment that local households do not pay for its power.

INDEPENDENCE, MISSOURI · 1.2 GW

USD 650m+ for the community

Payments in lieu of taxes projected above USD 650m over 20 years to the city, schools and other local bodies, plus about USD 30m a year to the city-owned utilities. About 1,200 construction jobs and 130 permanent ones. STEM and AI-literacy programs, first-responder support, a resident engagement panel, and a donation that cleared school meal debt. Closed-loop cooling uses about as much water as a petrol station.

BIRMINGHAM, ALABAMA · 300 MW

A disused office site, reused

Built on a former bank back-office campus. About USD 80m for city and county schools, 1,000+ construction jobs and 100+ permanent ones. Nebius pays for the new substation and the full cost of its power, so other Alabama Power customers are not affected. GPU grants, K-12 AI literacy and training for career switchers.

BUTLER, PENNSYLVANIA · 1.2 GW

Already zoned for data centers

Bought for USD 187.5m in an industrial park that already has a data center zoning overlay, so no rezoning is needed. PPL supplies the power, Nebius covers its full cost, and local ratepayers' bills do not rise. Plans were presented to residents at the local high school.

MÄNTSÄLÄ, FINLAND · 75 MW

Heating the town

Server heat goes into the district heating network: 19.5 GWh in 2025, about 65% of the town's heating needs, cutting connected households' heating bills by about 10% and avoiding about 4,000 tonnes of CO2. Cooled by outside air, with no water or refrigerants.

The projects were debated: dozens of residents spoke against the tax incentives in Independence, and in Birmingham residents protested and a local humane society has sued the city over the approval. In each case the project went through the public process and is under construction.

Permit and power rules are moving Nebius's way

LARGE-LOAD TARIFFS

Paying your own way is the norm

By May 2026, 23 states had approved at least one large-load tariff, which makes data centers pay for their own grid upgrades. Pennsylvania's regulator adopted a model tariff 5-0 in April. Nebius already pays the full cost of its power at its own sites.

BRING YOUR OWN POWER

Faster grid connections for sites with their own power

PJM, the grid operator for Pennsylvania and New Jersey, created an expedited connection track in January 2026 for large loads that bring their own new generation. New Jersey passed a law encouraging it. Behind-the-meter supply and on-site batteries, like the system planned at Butler, are encouraged rather than challenged.

PERMITS

Pennsylvania chose rules over a ban

Pennsylvania chose rules over a moratorium: data centers leave the permit fast track and state agencies may no longer sign secrecy agreements with them. Nebius already publishes its plans, holds public meetings and does not depend on fast-track permits.

How much water and electricity a site uses

Electricity

US data centers used about 176 TWh in 2023, about 4.4% of US electricity, and could reach up to 12% by 2028. Nebius's US plan of about 3.1 GW would use at most about 27 TWh a year if every MW ran flat out, roughly 0.6% of US electricity. Its owned sites are paid for under large-load terms, with the power infrastructure funded by Nebius.

Water

US data centers consumed about 17 billion gallons of water directly in 2023, mostly through evaporative cooling. Nebius uses closed-loop liquid cooling in the US and outside air in Finland. In Independence it would not rank among the city's top 50 water users; the city's plant produces about 28 million gallons a day. The water fear is about older, evaporative designs, not this one.

Why the host community usually comes out ahead

TAXES

The biggest taxpayer in town

In Loudoun County, Virginia, data centers provide about 38% of the general fund and about USD 26 of tax for every USD 1 of services they use, which let the county cut other taxes. Prince William County collected USD 280m in 2024.

JOBS

Construction jobs first, permanent jobs after

Each large campus brings 1,000 or more construction jobs for several years and 100-130 well-paid permanent jobs. Research finds a clear lift in local employment in cities and a smaller one in rural areas.

POWER BILLS

Little effect on power bills when the data center pays its share

The evidence is mixed: one study finds about a 5% rise in retail power prices where it can be measured, another finds no effect for investor-owned utilities. Large-load tariffs and self-funded power, as at Nebius's sites, are exactly what removes that cost from households.

State and local actions as tracked by LLM Pulse (Moratorium Nation dataset), Ballotpedia via the same tracker, and press reports, as of mid-November 2026; the risk shading is this page's own reading of those trackers. Community figures from the cities, Nebius and local press. Electricity and water from the Lawrence Berkeley National Laboratory report via the Congressional Research Service and Pew; the Nebius electricity figure is this page's estimate.

13 · The stakes

The other companies Nebius owns a piece of

ClickHouse, Avride, Toloka and TripleTen add value per share and could be sold or pledged for cash.

When Nebius was carved out of Yandex in 2024, it kept a handful of businesses beyond the AI cloud. Some it has since partly sold to outside investors, keeping a stake. They do not appear in the cloud's revenue, but they are worth something, and that value belongs to Nebius shareholders. Each is valued at a recent or estimated company value, then cut by a haircut for being private, hard to sell and taxed when sold. The stakes are also a source of funding: they can be sold or pledged as collateral to raise cash without issuing a single new Nebius share.

What the stakes add to each share

Stakes and company values are estimates based on the latest funding rounds and stake sizes reported in the press, not disclosed prices. The per-share value uses today's 298m shares, with no discounting: these are values today, not in 2027.

14 · Valuation

What the company is worth per share

Sum of the parts on the 12 months after each snapshot, with bonds converted where it pays holders, discounted back to today.

Take each part of the business at the end of 2027, price each dollar of revenue it bills over the following 12 months the way a buyer would (more for steady, high-margin business), subtract the debt, divide by the shares, and bring it back to today. Then add the stakes in other companies from section 13.

Where does "inference" come from?

It is not a seventh contract or an extra site. Part of the AI cloud revenue modeled in sections 2 to 6 comes through Token Factory, Nebius's service where customers pay per AI answer (per token) rather than renting the chips. It runs on the same data centers and chips, so its revenue is already in the totals above. For the valuation it is simply carved out of the AI cloud line, about USD 1.5b a year at end-2027, and priced like software at a higher multiple, because customers buy a finished service, not hardware. The cloud line is reduced by the same amount, so nothing is counted twice. Set the inference share to zero to value everything as plain cloud.

The value uses the revenue billed in the 12 months after the snapshot date, not a single quarter, so a data center that runs a quarter late loses one quarter of revenue, not its whole year. Delays still cost more than the time alone suggests: while revenue more than doubles each year, a business that is a quarter behind bills noticeably less in any given year. Pick 2028 or 2029 to see the business once the data centers are built.

From the business to a fair value per share

Building delays against the price of cloud revenue

Illustrative fair value, USD per share, keeping everything else as set above. Green is above your current setting, red below.

Words

Glossary

Megawatt (MW)
How much electricity a data center can draw. The unit Nebius sells in. 1 GW = 1,000 MW.
Hyperscaler
The giants that run their own clouds: Microsoft, Meta, Google, Amazon. They sign big, long, low-price contracts.
Enterprise customer
AI labs and software companies renting guaranteed capacity, often paying for the chips in advance.
Spot / on-demand
Renting by the hour or month, with no long contract. The highest price, and the first to go in a slowdown.
Service-only (asset-light)
Nebius runs its cloud software on chips someone else paid for, and earns a fee.
Capex (equipment spend)
Money spent on chips, servers and buildings.
Prepayment
Rent paid up front. It funds the chips, carries no interest, and is "repaid" in computing time.
Cash profit (EBITDA)
Rent minus the bills to run the sites and the head office. Before depreciation, interest and tax.
Profit (EBIT)
Cash profit minus depreciation: the slice of chips and buildings used up each year.
Yearly return (IRR)
The interest rate a project pays on the money put into it.
15 · The people

The founders and the shareholders

A team that ran one of the largest infrastructures outside the US, owning about 16% of the company.

Nebius looks like a two-year-old start-up. It is in fact one of the most experienced engineering teams in Europe, restarting with a clean sheet, a listed company and a few billion dollars, and owned in large part by the people who run it.

1990s - 2021 · THE PEDIGREE

Yandex

Arkady Volozh co-founds Yandex, which grows from Russia's answer to Google into a sprawling tech group: search, maps, ride-hailing, payments and one of the earliest self-driving programs. Its holding company, Yandex N.V., is Dutch and listed on Nasdaq.

2022 - 2024 · THE BREAK

Leaving Russia

After the invasion of Ukraine, trading in the shares is frozen and Volozh is sanctioned by the EU; he publicly condemns the war and is taken off the list in March 2024. In July 2024 the Russian business is sold for about USD 5b, roughly half its fair value: the price of a clean exit.

2024 · THE RESTART

Nebius Group

What is kept: the Finnish data center, offices in Amsterdam and Tel Aviv, the Nasdaq listing, the international businesses (ClickHouse, Avride, Toloka, TripleTen) and the engineers. Trading resumes in October 2024 after almost three years.

DEC 2024 → · THE BACKERS

Nvidia, Accel, Orbis

Six weeks after trading resumes, Nvidia, Accel and Orbis lead a USD 700m private placement at USD 21 a share. Microsoft and Meta follow in 2025-26, and Nebius is chosen to build Israel's national supercomputer.

Why their track record matters

Building and running a cloud is mostly a software and operations problem. The people who now run Nebius spent two decades running one of the largest internet infrastructures outside the US: their own data centers, their own servers and their own cloud. That is why Nebius designs its own hardware and writes its own platform, rather than reselling someone else's.

It also explains the start. The team began with about 400 engineers in AI infrastructure, a data center already running in Finland, and cash from the Yandex sale, which let it skip the years most start-ups spend getting there.

Ownership and alignment

The founder owns about 12% and insiders about 16% in total, and long-term investors such as Accel and Orbis sit on the register. Nebius also has about 33m Class B shares, which carry more votes than the Class A shares traded on Nasdaq. Insider sales have been regular but small: the founder's quarterly sales in 2026 were a few tens of thousands of shares, out of a stake of about 30 million. Their wealth rises and falls with the share price, like everyone else's.

Sources

Where the numbers come from

Regulation and communities

Tokens, GPU prices, open models and Palantir

The spot market

Acquisitions

Customers and the Nvidia stake

Company results and balance sheet

Sites and capacity

Bonds and loans

History, people and ownership

Which site serves which customer, other than Microsoft at Vineland, is an assumption. Stake values are estimates, not disclosed prices. All figures in USD; m = million, b = billion. The page shows how the parts fit together, not what will happen.