Nebius Group
A landlord for AI computing power: it fills data centres with NVIDIA's chips and rents them to AI labs by the hour or on contracts lasting years.
The scores
Growth3/4 ↑
Revenue up 454% and ARR more than doubled in six months, but three customers are 59% of revenue and the record is two years.
Business quality3/4 ↑
Top-tier rating and rising prices in a shortage, but untested by a glut and attacked by CoreWeave, Nscale, Google TPUs and its own customers.
Financial position2/4 →
Cash roughly matches debt, but the 2026 plan and 1 GW a year from 2027 need money not yet raised.
Execution record3/4 →
Beat every 2025 target and delivered Microsoft's tranches, but the record is short, connected power slips to 2027 revenue, and controls are weak.
Per share capture2/4 ↓
Shares up about a sixth in a year, an ATM in use, a further 20% authorised and in-the-money convertibles worth roughly 50M shares.
Potential3/4 ↑
A large upside on a short chain: $37.49B already signed and power secured, needing only build, financing and price to hold.
Risk profile2/4 →
Three customers are 59% of revenue, NVIDIA is the only supplier and one holder controls the votes; each is disclosed and none single-step.
Valuation2/4 →
At about 9 times December's expected yearly billing, the price needs the plan and 2028's sevenfold revenue at Q2 prices.
Evidence quality3/4 →
Revenue, backlog and contracts are filed, but deal prices and capacity are management's word and no counterparty names Nebius.
If it works 5.5× · if it fails you keep up to 25%.
The Verdict
We rate Nebius neutral. It fills data centres with NVIDIA's chips and rents them to AI labs, and the business is mixed: demand is real and customers have signed billions of future contracts, but the building is paid for partly with new shares. The price already assumes 2028 revenue about seven times this year's, at contract prices set during a chip shortage, so it is full, not cheap. As a bet it is speculative: if those prices survive 2027's flood of new chips, the shares could be worth several times today's; if they do not, shareholders keep under a fifth.
The long caseDemand for AI computing is growing faster than data centres can be built, so a GPU cloud can stay busy, if it can fund the building.
The evidence under this - 10 claims
- The rise applies to pay-as-you-go prices; the long contracts that make up most of the backlog are priced at signing and are untouched by it.
- Rothschild & Co Redburn's Sell case of September 21 rests on falling GPU rental prices over its horizon.
- SemiAnalysis sells consulting to GPU clouds and their buyers; it is independent of Nebius but has a stake in the market it rates.
- The rating is of service quality, not of whether the quality survives a glut: the same report notes CoreWeave still sets the technical bar.
- The vendor's $58.5B market value counts about 252M shares and so understates the company by roughly $13B.
- Third-quarter capex net of prepayments is not yet disclosed and will have reduced the cash used here, raising total value.
- At the 2026 base of $12M per megawatt the same revenue needs about 1.9 GW, which the company has not said it will have billing by 2028.
- Management's 5 GW of contracted power would carry far more than this if built and priced at Q2 terms, so the price may assume less than the plan.
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
The same index fell roughly three-quarters from early 2024 to late 2025, so the series has already shown how fast a chip generation loses its price once supply catches up.
Four analysts is a thin base, and the same panel still expects a loss per share in 2028.
The company
Training an AI model takes thousands of NVIDIA chips working as one machine for weeks. Few labs can build that themselves. Nebius builds the data centres, fills them with NVIDIA's newest systems and rents time on them by the hour or on contracts of months to years, with its own software for running the jobs.
Buyers would otherwise rent from Amazon, Microsoft, Google or Oracle, or from rival CoreWeave. An outside rater puts Nebius's clusters in its top tier with CoreWeave alone, above all four big clouds.
The evidence under this - 3 claims
- SemiAnalysis sells consulting to GPU clouds and their buyers; it is independent of Nebius but has a stake in the market it rates.
- The rating is of service quality, not of whether the quality survives a glut: the same report notes CoreWeave still sets the technical bar.
How it makes money
Customers pay for chip time. Big buyers sign contracts of several years and often pay part of the bill upfront, which helps fund the building. Microsoft and Meta must pay whether or not they use the capacity.
The catch is cost: every new contract needs billions of dollars of chips and buildings before the first dollar of rent arrives.
The evidence under this - 4 claims
- Microsoft's own filings never name Nebius; the figure rests on Nebius's disclosure alone.
- Microsoft's obligations began only once Nebius confirmed it had financed the build, which ties the contract to the capital markets.
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
Where the money comes from, stream by stream
| Stream | Share of sales (Q2 2026) | How it earns it |
|---|---|---|
| Nebius AI cloud | 98.1% | Rents NVIDIA GPU clusters with its own cloud software, on demand or under multi-year reserved contracts such as Microsoft's and Meta's. |
| TripleTen | 1.7% | Online courses retraining adults for technology jobs, paid per student, mainly in the US and Latin America. |
| Avride | 0.2% | Autonomous-driving and delivery-robot technology, pre-commercial and funded partly by outside investors including Uber. |
Who runs it
Arkady Volozh, who founded and ran Yandex for over twenty years, has run Nebius since 2024 and owns about a ninth of it. His special shares give him just over half the votes, so he controls the company. The finance chief joined in 2025. Top managers are paid mostly in shares that vest with time and in options, so they gain when the stock rises; no bonus target on sales or profit is disclosed.
The evidence under this - 4 claims
Several executives hold options struck at $100, so part of the equity pays only above a price, and the 2024 PSUs were forfeited rather than replaced
Why it wins, and who could beat it
Why customers come back
In a shortage, whoever has chips switched on when a lab needs them sets the price. Nebius secured power early, more than 75% of it on sites it owns.
The evidence is pricing. It raised its pay-as-you-go prices by 17% to 21% from October 1, the second rise in three months, and an independent price index for older NVIDIA chips is up 38% since December. That advantage lasts only as long as the shortage does.
The evidence under this - 4 claims
No filing tabulates megawatts by site, and the 20-F still lists co-location in France, Iceland, the UK, Israel and Kansas City, so the owned share is of power contracted for the future, not of capacity running today.
- The rise applies to pay-as-you-go prices; the long contracts that make up most of the backlog are priced at signing and are untouched by it.
- Rothschild & Co Redburn's Sell case of September 21 rests on falling GPU rental prices over its horizon.
The same index fell roughly three-quarters from early 2024 to late 2025, so the series has already shown how fast a chip generation loses its price once supply catches up.
- SemiAnalysis sells consulting to GPU clouds and their buyers; it is independent of Nebius but has a stake in the market it rates.
- The rating is of service quality, not of whether the quality survives a glut: the same report notes CoreWeave still sets the technical bar.
The threat
CoreWeave sells the same thing with $103.7B of signed contracts against Nebius's $37.49B, and the newcomer Nscale has won tens of billions from Microsoft and Anthropic. Google now sells its own AI chips to outsiders. And Microsoft and Meta, behind most of Nebius's contracts, build their own data centres and never mention Nebius in their filings.
The evidence under this - 5 claims
CoreWeave's margin fell partly because it brings capacity online ahead of the revenue it earns, which Nebius's own 2027 ramp will repeat.
Anthropic also signed up to $44.6B with Nscale, a neocloud renting NVIDIA chips, so the move to TPUs adds a supplier rather than replacing one.
Meta says most of its $349.31B of commitments are third-party cloud capacity, so its own building has not stopped it renting.
The rivals: CRWV, GOOGL, META, MSFT, AMZN, ORCL, IREN, and 1 more that is private, in-house or listed abroad.
Who else does this, rival by rival
| Ticker | Rival | How far along | Priced against it |
|---|---|---|---|
| CRWV | CoreWeave, Inc.Priced against it: Yes | On sale | Yes |
| Private | NscalePriced against it: Private | On sale | Private |
| GOOGL | Google TPU systems sold to outside customersPriced against it: Yes | On sale | Yes |
| META · MSFT | Meta's and Microsoft's own data-centre buildsPriced against it: Yes | On sale | Yes |
| AMZN | Amazon.comPriced against it: Yes | On sale | Yes |
| ORCL | OraclePriced against it: Yes | On sale | Yes |
| IREN | IRENPriced against it: Yes | On sale | Yes |
How big it could get
How big this could get
Specialist AI clouds like Nebius sold more than $25B of computing time in 2025, by an outside count, and $9B in the last three months alone, more than triple a year earlier. The electricity used by AI servers is expected to grow 30% a year to 2030. Nobody counts how much capacity is rented in total, only the dollars.
The evidence under this - 2 claims
Synergy's forecast of close to $400B by 2031 is a projection, not a measurement, and is not used here.
The IEA's own range runs from a lower efficiency case to a 1,260 TWh lift-off case; the base case is one scenario of several.
Who has signed
Microsoft has committed up to $17.39B through 2031. Meta has signed $12B of dedicated capacity starting in early 2027, plus up to $15B it buys only if Nebius cannot sell that capacity elsewhere. Four other deals in Q2 2026 averaged more than $1B each.
The evidence under this - 3 claims
- Microsoft's own filings never name Nebius; the figure rests on Nebius's disclosure alone.
- Microsoft's obligations began only once Nebius confirmed it had financed the build, which ties the contract to the capital markets.
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
The results
The numbers
Sales grew 479% in the year to Dec 2025, slower than the 834% in the year before.
Operating margin is −115%, up from −12× in the year to Dec 2022.
Running costs take 184¢ of each sales dollar, up from 55¢ in the year to Dec 2021.
Shareholders' money earned 2.1% a year, up from −5.0% in the year to Dec 2021.
It used $3.68B of cash in the latest year, so the $3.68B it holds lasts about 1.0 years at that pace.
It owes $450M more than the cash it holds, about 1.2 years of the cash its operations bring in.
The share count fell 32% in 4.0 years, and pay in shares costs 16% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-04. Fiscal years end in December.
| Dec 2021 | Dec 2022 | Dec 2023 | Dec 2024 | Dec 2025 | |
|---|---|---|---|---|---|
| Size and growth | |||||
| Sales | $4.79B | $13.5M | $9.80M | $91.5M | $530M |
| Sales growth | +62% | −100% | −27% | +834% | +479% |
| Profit | |||||
| Gross margin | 51% | −110% | −100% | 52% | 69% |
| Operating margin | −3.7% | −12× | −29× | −437% | −115% |
| Net margin | −4.1% | 55× | 25× | −701% | 16% |
| What growth costs, per sales dollar | |||||
| Research and development | 14% | 432% | 889% | 125% | 33% |
| Selling, general and admin | 35% | 424% | 16× | 279% | 72% |
| Pay in shares | 5.8% | 27× | 38× | 131% | 16% |
| Capital spending | 13% | 108% | 846% | 883% | 767% |
| Cash | |||||
| Cash from operations | $125M | $697M | $830M | $246M | $385M |
| Free cash flow | −$476M | $682M | $747M | −$562M | −$3.68B |
| Balance sheet | |||||
| Cash and short-term investments | $1.18B | $293M | $116M | $2.43B | $3.68B |
| Debt | - | $728M | $558M | $6.10M | $4.13B |
| Net cash | - | −$435M | −$442M | $2.43B | −$450M |
| Returns and shares | |||||
| Return on capital | - | - | - | - | −7.3% |
| Shares (diluted) | 362M | 368M | 371M | 281M | 248M |
What growth costs
Growth is making the running costs lighter: staff and office spending now take a far smaller slice of each sales dollar, so losses have shrunk. The heavy cost is building: each year the company spends several times its sales on chips and buildings, and that spending turns into depreciation as the new sites switch on.
The evidence under this - 2 claims
The last report (12 Aug)
Quarter ended June 2026, filed figures.
Q2 2026 revenue was $582.3M, up 454% on a year earlier, and the yearly pace of billing reached $3.0B. Before depreciation, interest and tax the business earned 41 cents on each dollar of sales.
Under the official accounts it still lost $190.4M. The chips wear out on paper over five years, interest is rising, and a one-off acquisition cost landed in the quarter. Spending on equipment was $5.66B, against $2.25B of cash coming in, most of it customer prepayments.
What it did to the thesisIt proved the demand: revenue and the billing pace are soaring. It also showed the cost: heavy losses on the official accounts, spending far above the cash coming in, and depreciation stretched to flatter profit. The price assumes the plan works, so the view stays Neutral.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $-0.12 against $-0.72 expected.
The evidence under this - 4 claims
Management called four years conservative in May 2025 and still calls the policy conservative; CoreWeave has used six years for its computing equipment since 2023, so five is not out of line with the industry.
What management did with the money
Management has raised money every way it can: $5.75B of bonds that can turn into shares in August, $2.8B of new shares sold on the market in Q2, and $2B from NVIDIA. Shareholders have approved up to 20% more shares. No insider at the top has chosen to sell.
Insiders in 2026Q3: 11 purchases and 48 sales, from their Form 4 filings.
The evidence under this - 4 claims
The price, and where we disagree
In short. Full. Buyers are paying for the plan working and the years beyond it: about $22.7B of 2028 revenue against $3.0B to $3.4B this year, at prices set during a shortage.
The shares have roughly doubled in a year and swing hard on each piece of news: up 34.1% the day after the August results, down 9.9% a week later on the bond sale.
The next check is the Q3 2026 results, due around November 10. Billing needs to keep climbing fast for the year-end target of $7B to $9B to hold, because the capacity switched on this year only starts earning through the first half of 2027.
The evidence under this - 3 claims
- First-half revenue was $981.3M, so the second half must roughly double it.
- The Q1 guide for a lower Q2 margin was wrong in the favourable direction, which says the quarterly outlook is loosely held both ways.
Contracted power means land and power secured, not buildings; the company has given no current active or connected megawatt figure since about 170 MW at the end of 2025.
How we read the multiple
Trailing sales lag a business growing this fast, so the better question is what the price needs.
At 2026's earlier contract prices, that 2028 revenue would take nearly twice the chips. The price is a bet that shortage prices last.
- The company is worth about $71.7B.
- That is about 9 times the yearly billing targeted for December.
- It is 3.1 times the $22.7B of 2028 revenue analysts expect.
- At Q2 2026 deal prices, that needs about 1 gigawatt billing.
Against its own history. Far cheaper on this measure than a year ago, because sales grew much faster than the price; the price has been chasing the business, not running ahead of it.
The evidence under this - 4 claims
- The vendor's $58.5B market value counts about 252M shares and so understates the company by roughly $13B.
- Third-quarter capex net of prepayments is not yet disclosed and will have reduced the cash used here, raising total value.
Four analysts is a thin base, and the same panel still expects a loss per share in 2028.
- At the 2026 base of $12M per megawatt the same revenue needs about 1.9 GW, which the company has not said it will have billing by 2028.
- Management's 5 GW of contracted power would carry far more than this if built and priced at Q2 terms, so the price may assume less than the plan.
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
Priced against its rivals
The competitors its own annual report names, each confirmed against the rival's filing. Measured on company value for each $1 of sales, cheapest first:
| Ticker | Company | company value for each $1 of sales |
|---|---|---|
| AMZN | Amazon.com | 3.7× |
| ORCL | Oracle | 7.4× |
| GOOGL | Alphabet | 9.4× |
| MSFT | Microsoft | 11.9× |
| CRWV | CoreWeave | 12.4× |
| IREN | IREN Limited | 24.5× |
| NBIS | Nebius Group | 44.7× |
The middle of the other 6 is 10.7. Nebius Group N.V. is dearest of 7.
Rivals' figures come from a data provider, so read the spread of the group rather than any single pair.
Buyers pay more for each dollar of its past sales than for any rival's. That reflects its own speed of growth, not something the whole group shares.
Private or foreign rivals that cannot be priced: Crusoe, Lambda.
The bet: what it pays if it works
If it works, the shares could be worth several times today's price. The best case we can defend is the 2028 revenue analysts expect: about one gigawatt of chips rented out at this year's contract prices. Valued like the other listed clouds, that is about four times today's company value, before the new shares the build will take.
If it fails, with rents back down and money harder to raise, shareholders keep a cloud renting today's chips, valued like the cheapest cloud: at most a fifth of today's value.
| Revenue in two to three years | Megawatts billing needed | Valued like its peers | Like the cheapest |
|---|---|---|---|
| $22.70B (the big win) | about 1.0K at 100% kept | 5.5× | 1.9× |
| $45.00B | about 2.0K at 100% kept | 11.0× | 3.8× |
To pay 2× (a Fair bet) it would need about $8.21B of revenue (about 365 megawatts billing); to pay 5× (a Cheap bet), about $20.53B of revenue (about 913 megawatts billing), at the same multiple.
The steps it has to take
- Done 2025 targets beaten: $1.25B year-end run-rate against $0.9B to $1.1B guided
- Done Microsoft and Meta signed: up to $17.39B and $12B plus up to $15B
- Done Q2 2026 deals priced at $20M to $25M a year per megawatt
- In progress Annualised revenue of $4.5B or more at September 30, 2026 due 2026-11-10
- In progress Year-end 2026 run-rate inside the $7B to $9B guide due 2027-02
- In progress Microsoft's and Meta's capacity delivered without a termination or delay due 2027-06
- Not started Contracts for 2027 capacity averaging $20M a year per megawatt or more due 2027-08
- Not started 2028 revenue near $22.7B, about 1 GW billing due 2029-02
The detail behind the label
The evidence under this - 6 claims
Four analysts is a thin base, and the same panel still expects a loss per share in 2028.
- At the 2026 base of $12M per megawatt the same revenue needs about 1.9 GW, which the company has not said it will have billing by 2028.
- Management's 5 GW of contracted power would carry far more than this if built and priced at Q2 terms, so the price may assume less than the plan.
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
Where we disagree with the market
A top-tier AI cloud whose demand will outrun supply for years; a vocal minority calls it an over-borrowed chip landlord.
Demand is real and prices are rising, but the price already assumes those prices survive 2027's new supply, which only management has vouched for.
The record contract prices from Q2 2026 are management's figures, on capacity not yet built, and no buyer has confirmed them. The same chips fell to about a quarter of their 2024 rent once supply caught up.
The case against us is strong: rents have risen all year by every outside measure, and the company says it could sell all of 2027 on these terms today.
What analysts say, and the evidence
The evidence under this - 5 claims
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
The same index fell roughly three-quarters from early 2024 to late 2025, so the series has already shown how fast a chip generation loses its price once supply catches up.
- At the 2026 base of $12M per megawatt the same revenue needs about 1.9 GW, which the company has not said it will have billing by 2028.
- Management's 5 GW of contracted power would carry far more than this if built and priced at Q2 terms, so the price may assume less than the plan.
What could go wrong, and how we would know
The risks
- Prices fall when supply arrives. New NVIDIA systems flood in during 2027; if rents fall back to 2026's earlier level, the same chips earn about half as much.
- Three customers, most of the money. Three customers were 59% of Q2 2026 revenue, and Microsoft can cancel any capacity Nebius delivers late.
- Building needs constant new money. Equipment spending runs far ahead of cash coming in, so new shares and convertible bonds are likely every year.
- Buyers become rivals. Microsoft, Meta and Google build their own data centres, and Google now sells its own AI chips to outsiders.
- Weak accounting controls. The 2025 annual report found two material weaknesses, one over the fixed assets that make up most of the balance sheet.
- One man controls the votes. The founder holds about 52% of the votes, so outside shareholders cannot stop new share issues.
The evidence under this - 8 claims
- No counterparty has disclosed any of the four deals' terms; Reflection announced only 'more than $1 billion'.
- The payback of 1 year 10 months is stated on forecast costs and on capacity not yet built, in the letter's own footnote.
- Management said the deals will not materially change 2026 revenue, so none of the yield is in reported numbers yet.
The share is falling as Microsoft's tranches reach full revenue and the Q2 deals are with four other buyers, though none of those will bill materially until 2027.
- Microsoft's own filings never name Nebius; the figure rests on Nebius's disclosure alone.
- Microsoft's obligations began only once Nebius confirmed it had financed the build, which ties the contract to the capital markets.
More than $9B of 2026 prepayments, if received, plus GPU-backed debt like July's facility could fund most of the second half without new shares.
Meta says most of its $349.31B of commitments are third-party cloud capacity, so its own building has not stopped it renting.
Anthropic also signed up to $44.6B with Nscale, a neocloud renting NVIDIA chips, so the move to TPUs adds a supplier rather than replacing one.
What we still do not know
- What price per chip-hour the Microsoft, Meta and Q2 contracts actually carry; no one discloses it.
- How many megawatts are actually switched on and billing today; the last figure was about 170 at the end of 2025.
- Whether older chips still earn their keep in years four and five, now that they are depreciated over five years instead of four.
The evidence under this - 3 claims
Contracted power means land and power secured, not buildings; the company has given no current active or connected megawatt figure since about 170 MW at the end of 2025.
Management called four years conservative in May 2025 and still calls the policy conservative; CoreWeave has used six years for its computing equipment since 2023, so five is not out of line with the industry.
- Microsoft's own filings never name Nebius; the figure rests on Nebius's disclosure alone.
- Microsoft's obligations began only once Nebius confirmed it had financed the build, which ties the contract to the capital markets.
What would change our view
- Neutral→AttractivePricing holds and the multiple comes in
It needs all of:
- ○The company's value falls to 25 times its past year's sales, with the year-end billing target still intact.checked daily
- ○New contracts for 2027 capacity average at least $20M a year per megawatt.checked at the next results, due by 10 Nov
- Neutral→UnattractiveThe run-rate misses or prices crack
It needs any of:
- ○The yearly pace of billing at the end of 2026 comes in below $7B.checked at the next results, due by 10 Nov
- ○Nebius cuts its pay-as-you-go chip prices across generations.watched as it happens
We would also be wrong if
- Less than 800 megawatts connected at the end of 2026.
- Microsoft or Meta ends a block of capacity because it arrived late.
What is coming
- 10 November 2026 Q3 2026 results: Whether ARR reaches $4.5B toward the $7B to $9B year-end guide and an active-power figure is given.
Every dated event ahead
- 10 November 2026 Q3 2026 results: Whether ARR reaches $4.5B toward the $7B to $9B year-end guide and an active-power figure is given.
- February 2027 Q4 and full-year 2026 results: Year-end ARR against the guide, the 2027 capex and funding plan, and whether Q2's contract prices held for new deals.
- March 2027 Meta's dedicated capacity begins: Whether the $12B Meta capacity starts on time and earns revenue, or slips like 2026's connected power.
- April 2027 2026 annual report (20-F): Whether the two material weaknesses in internal control are remediated, and the updated ownership and pay disclosures.
Since we wrote this
- 22 OctNewsA $3.5B one-year bank loan adds debt, not sharesIt funds the build for now, at a higher rate than July's loan and secured on nearly all assets; the case and the view are unchanged.
- 19 OctBet moveBet re-struck: Fair bet to Cheap betRe-priced at the 16 October close.
- 5 OctThesis reviewThe bet now names what it ridesThe owner asked that every bet name the macro trend behind it, after BWXT's bet was missing its chief executive's case for nuclear over gas.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the method silent on who runs the company, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the 20-F and interim filings.
- 2 OctThesis writtenNeutral, medium convictionDemand and pricing are real and filed, but the price already assumes 2028 revenue that needs Q2's record contract prices to survive 2027's new supply.
Also checked 1 time with nothing new to act on.
Sources 38
Sources - 38
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-04-30Item 4 business description, site models and server design; the competition risk factor naming CoreWeave, Crusoe, Lambda and the four general clouds; the Microsoft and Meta contract notes; customer concentration; the two material weaknesses; the four-year depreciation policy and the planned change; Class B voting controlbacksC01C02C03C09C10C13C18C25C30C31C48C49C50C51C52C53
- S02primarycompany release2026-08-12revenue, adjusted EBITDA, net loss, capex, the balance sheet and the share count at June 30, 2026backsC14C15C16C17C20C28C29
- S03primarycompany release2026-08-12ARR, the four Q2 deals and their yield per megawatt, prepayment cover, payback, the 5 GW contracted-power target, the five-year depreciation life, older-generation GPU pricing against Q1backsC01C06C08C14C23
- S04primarycompany filing2026-08-12remaining performance obligations and their timing, customer concentration by quarter, property and equipment including assets not yet in use, the convertible notes table, the ATM sales, the depreciation change's effect, uncommenced lease commitmentsbacksC09C13C16C18C19C28C47C48C49C53
- S05primarycompany filing2026-08-12the drivers of the Q2 loss (Eigen AI compensation, deal fees), operating cash flow driven by deferred revenue, pricing modelsbacksC16C20
- S06primarytranscript2026-08-12Nebius Q2 2026 earnings call transcript (FMP, via scripts/transcript_fetch.py)guidance reaffirmed across all metrics, the $9B prepayment figure, connected power active only through the first half of 2027, the claim that all of 2027 could be sold on Q2 terms. Questions were read by investor relations from a portal, so there were no follow-ups and the company chose which were read
that's true about the connected, but I would think that about being active in -- throughout the first half of 2027.
backsC08C21C23 - S07primarycompany release2026-05-132026 revenue and margin guidance, owned share of contracted power, the cost-of-revenue basis, contracted power raised to more than 4 GWbacksC04C17C21
- S08primarytranscript2026-05-13Nebius Q1 2026 earnings call transcriptcapex guidance raised to $20B to $25B, the Meta option explained, the demand anecdote
we are typically seeing 4 or more customers competing for every GPU we bring online.
backsC04 - S09primarycompany release2026-02-12full-year 2025 revenue, year-end ARR of $1.25B against $0.9B to $1.1B guided, about 170 MW active against a 100 MW target, the 2026 targetsbacksC15C24
- S10primarycompany filing2025-09-08contract value through 2031, the financing condition precedent, Microsoft's termination right for late deliverybacksC10
- S11primarycompany filing2026-03-16$12B of dedicated capacity from early 2027 and up to $15B of optional capacity Nebius intends to sell to others firstbacksC11
- S12primarycompany filing2026-08-24$3.45B of 0.50% notes due 2030 convertible near $313.46 and $2.3B of 4.50% notes due 2034 near $324.65; $800M of older notes exchanged for about 15.8M sharesbacksC26
- S13primarycompany filing2026-07-17SOFR plus 2.50% to October 31, 2030, secured on deployed GPUs and an investment-grade customer's contracted cash flowsbacksC29
- S14primarycompany filing2026-07-2022,256,412 Class A shares or 9.3%, of which 21,065,936 sit under a pre-funded warrant bought on March 11, 2026backsC31
- S15primarycompany filing2026-08-26authority to issue up to a further 20% of share capital for five years, and to buy back up to 20%backsC28
- S16primarycompany filing2026-08-12revenue, $103.7B of remaining performance obligations, customer concentration, $35.6B of debt principal; its 10-K names only the general clouds as competitors and never NebiusbacksC12C32
- S17primarycompany release2026-08-111.5 GW active and about 3.7 GW contracted power; adjusted operating margin of 5% against 16% a year earlierbacksC32
- S18primaryindustry data2026-09-23SemiAnalysis, ClusterMAX 3.0 rating of GPU clouds SemiAnalysisNebius and CoreWeave the only Platinum-rated clouds; Google and Oracle Gold, Azure Silver, Crusoe Bronze; Nebius asking prepayments of up to 100% on one-year commitments. SemiAnalysis sells consulting to clouds and buyers, so it is independent of Nebius but not of the market
Nebius joins CoreWeave in the Platinum tier. While CoreWeave still sets the technical bar for others to follow, Nebius is now established as a provider that consistently commands a premium pricing over others.
backsC05 - S19primaryindustry data2026-10-01$2.77 per GPU-hour on October 1, 2026 (index page: 'As of Oct 1, 2026 2.77 USD / GPU-hour'), against $2.00 in December 2025 (blog at silicondata.com/blog/h100-price-spike)
H100 cloud rates plummeted from around $8/hr in early 2024 to the $1.50-$3.00/hr range by late 2025 across specialist providers.
backsC07 - S20primaryindustry data2026-04-02SemiAnalysis, The Great GPU Shortage SemiAnalysisthe one-year contract price series, a different measure from the on-demand index
H100 1-year GPU rental contract pricing has shot up almost 40% from a low of $1.70/hr/GPU in October 2025 to $2.35/hr/GPU by March 2026.
backsC07 - S21primaryindustry data2026-04-02Synergy Research Group, neocloud market forecast Synergy Research Groupthe measured size of the specialised AI cloud market in 2025, and Synergy's forecast of close to $400B by 2031
neocloud revenues are scaling at an unprecedented pace, reaching $9 billion in Q4-up 223% year-over-year-and exceeding $25 billion for full-year 2025.
backsC33 - S22primaryindustry data2026-07-30Synergy Research Group, Q2 2026 cloud market Synergy Research Grouptotal cloud infrastructure of $143.4B in Q2 2026 and $500B trailing twelve months; GenAI-specific cloud services growing 165% a year
Among the tier two cloud providers, those with the highest growth rates include CoreWeave, OpenAI, Oracle, Crusoe, Nebius, Anthropic and Nscale.
backsC33 - S23primarygovernment data2025-04-10data-centre electricity more than doubling to about 945 TWh by 2030; accelerated-server capacity growing almost five times. The iea.org page refused (403); the PDF was read from the IEA's own file host
Electricity consumption in accelerated servers, which is mainly driven by AI technology adoption (Box 2.1), is projected to grow by 30% annually in the Base Case
backsC34 - S24primarycompany filing2026-07-29$329.1B of datacenter leases not yet commenced; reliance on cloud infrastructure providers; no mention of NebiusbacksC12
- S25primarycompany filing2026-07-30$349.31B of commitments mostly third-party cloud capacity; 2026 capex guided at $130B to $145B; no mention of NebiusbacksC12C36
- S26primarycompany filing2026-07-24sale of TPU systems to outside customers; $24.1B of backstops for third-party data-centre build-outbacksC35C36
- S27primarycompany release2026-04-06a large AI buyer moving part of its compute to non-NVIDIA chips sold by a general cloud
We have signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity that we expect to come online starting in 2027.
backsC35 - S28primarycompany filing2026-09-18Nscale Form S-1 Nscalestatements of work with Microsoft worth up to about $43.8B and Anthropic agreements up to about $44.6B; names Nebius as a primary competitorbacksC37
- S29primarycompany filing2026-08-27a $9.7B five-year Microsoft agreement; names Nebius as a competitorbacksC37
- S30secondarynews2026-09-17Reuters, 'Nebius hikes AI cloud prices again as demand for computing power soars' (Euronext copy) Reuterson-demand prices for selected GPUs up 17% to 21% from October 1; reuters.com refuses this client, so the syndicated copy was read
Cloud provider Nebius said on Thursday it will raise pay-as-you-go prices for leasing selected Nvidia chips from October 1, marking the company's second price hike in three months.
backsC06 - S31secondarynews2026-08-07the short at $211.77 and the stated objection; his own post is paywalled and was not read
The fish have gorged themselves on off-balance sheet liabilities. Backstops. Uncommenced leases. Purchase commitments
backsC38 - S32secondarynews2026-09-21the bear case as an analyst stated it; the note itself was not read
The firm flagged three converging pressures: falling GPU rental prices, hyperscalers scaling captive fleets, and rising financing costs on multi-billion-dollar buildouts.
backsC38 - S33primaryvendor data2026-10-02Research pack, frozen at content/NBIS/research-pack.jsonprice, returns, the price-move decomposition, the analyst split and its drift, short interest, consensus revenue estimates, the filed competitor cohortbacksC39C40C42C45
- S34primarycompany filing2026-10-02Form 4 filings, read for the Rule 10b5-1 flag and footnotes via scripts/insider_fetch.py84 open-market sales and no purchases since March 31, 2026; every chief-executive and finance-chief sale was an automatic sale to cover tax on vesting sharesbacksC44
- S35primaryregulator2026-07-30up to seven publicly anchored AI sites in the EU, bids closing November 12, 2026; no bidder named
The joint public funding provided by the Union and Member States will act as an anchor customer, significantly de-risking the significant capital required
read, but no claim rests on it - S36primarycompany filing2026-03-11NVIDIA's $2B purchase and its support for more than 5 GW of Nebius capacity by the end of 2030backsC31
- S37primarycompany filing2026-03-02a six-year useful life for computing equipment since January 2023; Microsoft at about 67% of 2025 revenue; a competitor list naming only the general cloudsbacksC18
- S38primarycompany filing2026-10-22the $3.5B Citi bridge facility: size, rate, 364-day maturity, security and covenant, and management's account of whybacksC54