Nebius has never reported an operating profit — every dollar of bottom-line profit it has shown is a mark on a stake in another company
The AI cloud business is growing at a rate almost nothing in public markets matches, and the group has never once earned money from operations. Both of those are true, and the gap between them is filled by a private-company valuation, a depreciation estimate that lengthened in January, and $4.3 billion of convertible notes with another $4.5 billion proposed eight days after this print.
01
What happened
Nebius rents out AI computing power, and it is growing faster than almost anything on a public market — revenue up 454% in a year. It has also never made a profit from running that business. In both of the last two first-halves it reported net income anyway, and both times the profit came from re-valuing a stake it holds in a different company, the database maker ClickHouse. This quarter there was no such mark, and the result was a $190 million loss. Everything else in this report is about the machinery that sits between those two facts.
02
The numbers that matter
The figures the rest of this rests on, and which way each one cuts.
03
The headline vs. the filing
Every report in this section has found the same shape: the number that leads the coverage is not the number the filing supports.
"Net income from continuing operations was $430.8 million in the six months ended June 30, 2026" — the group has now reported first-half net income two years running.
A loss of about $(349.8)M once the ClickHouse revaluation is removed. In 2025 the same arithmetic gives $(199.2)M.
The gap: The H1 2026 profit sits against a $780.6M non-cash gain from remeasuring the ClickHouse stake; H1 2025’s $398.2M profit sat against a $597.4M gain. In both years the mark is larger than the profit. Q2 2026 is the control case — the gain was nil and the company lost $190.4M. Loss from operations, meanwhile, has been negative in every period disclosed: $(111.2)M, $(175.9)M, $(231.5)M, $(303.9)M. The mark is real, disclosed, and follows ClickHouse’s January Series D. It is also a Level 3 back-solve valuation of a private company, remeasured on a single date, 16 January 2026.
The operating loss widened from $(111.2)M to $(175.9)M — 58% — while revenue grew 454%.
Closer to 97% on the depreciation policy that was in force three months earlier.
The gap: Effective 1 January 2026 Nebius extended the useful life of servers and network equipment from four years to five. The filing quantifies it: the change cut Q2 depreciation by $43.0M and reduced the Q2 net loss by $34.1M ($86.1M and $75.7M for the half). On the prior estimate the Q2 operating loss would have been roughly $(218.9)M. Worth saying clearly in both directions — five years is still shorter than CoreWeave’s six, and shorter than Alphabet or Microsoft. Nebius lengthened toward the industry norm rather than past it. But it lengthened, and it lengthened into the quarter where the loss was going to widen.
"Net cash provided by operating activities" of $4,504.1M in the first half — against $(352.0)M a year earlier.
97.6% of it is customers paying in advance. Deferred revenue contributed $4,395.0M of that $4,504.1M.
The gap: Deferred revenue on the balance sheet went from $1,577.5M to $5,975.2M in six months. That is a genuine vote of confidence from large customers and it funds the build — but it is a liability, not earnings, and it converts to revenue over one to five years. Strip it and underlying operating cash generation is roughly $109M against $8,130.3M of capital spending. Free cash flow for the half was $(3,626.2)M.
04
Q2 2026 — the facts everything hangs on
Show the full print — 19 rows, every figure this report rests on
| Metric | Q2 2026 | vs Q2 2025 | Note |
|---|---|---|---|
| Revenues | $582.3M | +454% | From $105.1M. H1 $981.3M from $156.0M |
| — Nebius AI cloud | $574.9M | +514% | Segment Adjusted EBITDA $285.7M from $9.5M |
| — TripleTen (edtech) | $10.0M | −19% | Adjusted EBITDA $(9.4)M |
| — Avride (autonomy) | $1.0M | from $0.2M | Adjusted EBITDA $(40.1)M |
| Depreciation & amortisation | $259.7M | +245% | ⚠ 44.6% of revenue; exceeds group Adj. EBITDA |
| Total operating costs | $758.2M | from $216.3M | — |
| Loss from operations | $(175.9)M | from $(111.2)M | ⚠ Negative in every period disclosed |
| Useful-life change impact | +$43.0M | Q2 depreciation | ⚠ 4→5 years from 1 Jan; cut net loss by $34.1M |
| Interest expense | $(119.1)M | from $(4.8)M | Roughly 25×; H1 $(182.8)M |
| ClickHouse revaluation gain | nil | vs $597.4M | ⚠ H1 $780.6M; carrying $737.1M → $1,517.7M |
| Other income, net | $81.5M | from $24.6M | Not operating |
| Net income / (loss) | $(190.4)M | from +$502.5M | H1 +$430.8M — all of it downstream of the mark |
| H1 operating cash flow | +$4,504.1M | from $(352.0)M | ⚠ $4,395.0M of it is deferred revenue |
| Deferred revenue | $5,975.2M | from $1,577.5M | Recognised over one to five years |
| H1 capital expenditure | $8,130.3M | from $1,054.5M | ⚠ 8.3× H1 revenue; FCF $(3,626.2)M |
| Cash and equivalents | $8,042.1M | — | Against $8,499.0M of non-current debt |
| Customer concentration | 24 / 21 / 14% | 59% combined | ⚠ Prior year’s 39% and 15% both now below 10% |
| Unsatisfied RPO | $37,490.6M | 36% ≤24 mths | 40% in months 25–48; remainder later |
| Convertible notes | $4,337.5M | gross, H1 | ⚠ A further $4.50B proposed on 19 August |
05
The central tension
The bull and the bear do not disagree on the facts. They disagree on one thing — and it is the whole investment.
- Revenue grew 454%, and the AI cloud business alone grew 514% to $574.9M. Very little in public markets compounds at that rate.
- The cloud segment does make money on its own terms: Adjusted EBITDA of $285.7M against $9.5M a year ago, on a segment basis.
- Customers are pre-paying for capacity. Deferred revenue went $1,577.5M to $5,975.2M in six months — the build is being funded partly by the people who will use it.
- $37.5B of remaining performance obligations, disclosed in the filing rather than as a marketing figure, and explicitly net of the group’s estimate of variable consideration.
- A five-year server life is shorter than the peer set — CoreWeave uses six, as do Alphabet and (at the top of its range) Microsoft. Even after lengthening, Nebius depreciates faster than the companies it competes with.
- The balance sheet is far cleaner than the closest comparable: net debt of $2.2B against CoreWeave’s $46.0B, on $8.0B of cash.
- The ClickHouse and Toloka stakes are real assets, and the ClickHouse mark follows an arm’s-length Series D that Nebius did not participate in.
- No operating profit in any period disclosed, and the loss widened as revenue grew 454%.
- Every dollar of reported first-half profit, in both years, is smaller than the ClickHouse mark that produced it. Q2 had no mark and printed a $190.4M loss.
- A depreciation estimate lengthened in January, cutting Q2 depreciation by $43.0M and the loss by $34.1M, into the quarter where the loss widened anyway.
- 97.6% of operating cash flow is customer prepayment, not earnings — a liability that converts over one to five years.
- The top of the customer base rotated completely in twelve months: last year’s 39% and 15% customers both fell below 10%, replaced by three at 24%, 21% and 14%.
- Capex was 8.3× revenue in the half, and the FY guide of $20–25B against $3.0–3.4B of revenue is roughly 7×.
- $4.3B of converts raised, and $4.50B more proposed eight days after this print — roughly $8.8B of convertible debt inside eight months against $981M of half-year revenue.
- 46.9× trailing EV/sales against CoreWeave’s 12.7×, with 22% of the shares sold short.
The one question
Is the ClickHouse stake a distraction or the point? Read it one way and it is noise: a non-cash mark on a side holding, irrelevant to whether the AI cloud works, and the operating numbers are what matter — in which case Nebius is a hypergrowth infrastructure business that has not yet crossed into profit, valued accordingly. Read it the other way and it is the only thing that has ever made this group money, it arrived on a single day in January from a private round the company did not participate in, and it papers over a business whose losses widen as it scales. The filing supports both readings. What it does not support is the reading where the group has earned anything from operations.
06
What would prove this wrong
The discipline: name in advance what would break each side of the case.
The bull breaks if…
- Q3 shows the operating loss widening again, now that the useful-life change is fully in the base and can no longer flatter the comparison.
- Deferred revenue stops growing — the prepayment engine reverses and operating cash flow falls sharply while revenue still looks strong.
- The $4.50B convertible offering prices on materially worse terms than the H1 raise, or is pulled.
- A top-three customer drops below 10% again, showing the roster rotation is churn rather than growth.
The bear fails if…
- The AI cloud segment reaches group-level operating profit — the first time the business earns money without a mark.
- RPO converts faster than the disclosed 36%-in-24-months schedule, showing the backlog is nearer than the filing implies.
- Customer concentration falls because the denominator grew, with named investment-grade counterparties replacing unnamed ones.
- Capex intensity falls toward peer levels while growth holds, showing the 8.3× was a build phase rather than a run rate.
07
What this means if you don’t trade stocks
There is a habit worth taking from this one, and it costs nothing: when a company reports a profit, check which line it came from.
Nebius reported net income in both of the last two first-halves. Both times, the number came from re-valuing a stake it owns in a different company — not from selling anything. This quarter there was no re-valuation, and the same business posted a $190 million loss. Nothing was hidden; it is all in the filing, in plain language, with the dates and the amounts.
The point is not that Nebius did anything wrong. It is that "the company was profitable" and "the business made money" are different sentences, and only one of them was true. That distinction shows up constantly once you look for it — a one-off gain, an asset sale, a mark on an investment, a tax item.
The practical version: read the line above the bottom line. Operating income tells you whether the business works. Net income tells you what happened to everything else as well. When those two disagree, the disagreement is usually the story.
08
The longer read
Valuation, the risks in order, and the horizon this resolves on.
Valuation — the mirror image of its closest peer
At $223.90 Nebius is worth $61.4B, with an enterprise value of $63.5B. The two neoclouds now covered in this section are almost photographic negatives of each other:
| Nebius | CoreWeave | |
|---|---|---|
| Enterprise value | $63.5B | $96.2B |
| Total debt | $10.2B | $51.6B |
| Net cash | −$2.2B | −$46.0B |
| EV / sales (trailing) | 46.9× | 12.7× |
| 52-week change | +209% | −6% |
| Short interest | 22.0% | 11.1% |
Nebius has the cleaner balance sheet by an order of magnitude and the richer multiple by nearly four times. On 2026 guidance the gap narrows but does not close — roughly 19× EV/revenue against CoreWeave’s ~7×.
22% of the shares are sold short, the highest this section has covered. That cuts both ways and is worth holding lightly: it means a large group of professionals has taken the bear side of exactly the argument above, and it means the price carries mechanical squeeze risk that has nothing to do with the filings.
What a 6-K does not tell you
This is the first foreign private issuer in this section, and the structure matters more than it sounds.
Nebius is Dutch-domiciled and files 6-K and 20-F, not 10-Q and 10-K. The August 12 accession contains three documents: a cover, the MD&A, and the financial statements. There is no press release exhibit and no investor deck.
The consequence is concrete. Every headline operating metric the stock trades on — annualised run-rate revenue, contracted megawatts, revenue guidance, capex guidance — appears in none of them. ARR of $3.0B at end-June, the $7–9B exit-ARR target, FY revenue of $3.0–3.4B, capex of $20–25B and a year-end contracted-power target raised to 5 GW all come from the earnings call and the company’s own website. A US filer would have furnished the release as an 8-K exhibit; CoreWeave’s release points directly at its 10-Q.
None of that is improper — it is what the foreign-private-issuer regime permits. But it means the audited-adjacent record and the numbers in the headlines are two different documents, and only one of them is filed.
What the filing does disclose is better than expected in two places: RPO of $37,490.6M, with the honest detail that only 36% is expected to convert within 24 months and 40% falls in months 25–48; and a full customer-concentration table.
The customer table is the most interesting page in the filing
| Customer | Q2 2025 | Q2 2026 |
|---|---|---|
| A | 39% | under 10% |
| B | 15% | under 10% |
| C | under 10% | 24% |
| D | under 10% | 21% |
| E | under 10% | 14% |
The entire top of the customer base turned over in twelve months. Three unnamed customers are now 59% of revenue, and the two that were 54% between them a year ago are both gone from the table.
At 454% growth this is what you would expect — new customers arriving so much larger than the old ones that the old ones drop below the threshold. It is not evidence of churn on its own. But it does mean the revenue base has no demonstrated persistence at the top, and the filing names none of them.
Risk — each isolated, do not blur
- Earnings quality (dominant). No operating profit in any disclosed period; all reported bottom-line profit downstream of a Level 3 mark on a private company, remeasured on one date.
- The depreciation estimate. A four-to-five-year extension worth $43.0M a quarter, taken effective January, in a business where depreciation already exceeds group Adjusted EBITDA.
- The funding gap. FY capex guidance of $20–25B against $8.1B spent in the half implies $12–17B still to fund, against $8.0B of cash — which is what the $4.50B convertible proposal on 19 August is for.
- Prepayment dependence. 97.6% of operating cash flow is deferred revenue. If bookings slow, the cash flow statement deteriorates before the income statement does.
- Customer concentration. 59% in three unnamed customers, with a fully rotated roster.
- Conglomerate drag. Avride and TripleTen contributed $(49.5)M of Adjusted EBITDA loss in the quarter, roughly 28% of the group operating loss, on $11.0M of combined revenue.
- Dilution. ~$8.8B of convertible notes raised or proposed inside eight months, plus prefunded warrants and treasury share sales in the H1 financing line.
Horizon and sizing (kept separate)
Horizon. The near term is mechanical: whether the $4.50B convertible offering prices and on what terms, and whether Q3 shows operating loss narrowing rather than widening now that the depreciation change is in the base. The thesis resolves over 2–3 years on one question — does the AI cloud reach operating profit before the capital markets tire of funding it?
Sizing considerations (not a recommendation). Nebius is the cleanest balance sheet and the richest multiple in the neocloud pair, which is an unusual combination and probably not a stable one. It also carries something none of the others do: a material part of the equity story is a stake in a private company whose value is set by other people’s funding rounds. With 22% of shares short, the price will move for reasons unrelated to any of this.
How this reads against the other reports
This section has now looked at seven companies in the same AI-capital cycle, and Nebius completes a pair that is worth seeing together.
Nebius and CoreWeave do the same thing — rent out GPUs — and have arranged themselves as near-opposites. CoreWeave carries $51.6B of debt and trades at 12.7× trailing revenue; Nebius carries $10.2B and trades at 46.9×. One has the leverage, the other has the multiple. Both spend multiples of their revenue on capital equipment, and both depend on depreciation estimates that nobody has yet tested through a full replacement cycle.
Ranked by capital spending against revenue, the picture is consistent: Palantir at 0.75%, Microsoft around 32%, Amazon at roughly 105% of operating cash flow, SpaceX at 235%, CoreWeave at about 290%, and Nebius’s first half at 828%. AMD sits outside the scale entirely — it sells the chips the rest are buying.
How this was researched
Every Market Storm report is produced by STORM — a multi-agent research method adapted from Stanford's STORM. Several AI agents each take a different stake (a fundamentals analyst, a short-seller, an industry engineer, a valuation watcher), interview each other while grounded in live web search, and surface the load-bearing claims. A separate skeptic pass then tries to refute each of those claims against primary sources. What survives is written up here — with the caveats it earned. The finance is the payload; the method is the point.
Verification ledger
A separate skeptic pass tried to refute every load-bearing claim against primary sources. Where it bit:
Verified directly against the Q2 2026 6-K exhibits, quoted verbatim where it matters: the four-to-five-year useful-life extension and its $43.0M / $34.1M quarterly impact · the ClickHouse remeasurement of 16 January 2026 taking carrying value $737.1M → $1,517.7M · deferred revenue contributing $4,395.0M of $4,504.1M of operating cash flow · the customer-concentration table · RPO of $37,490.6M with 36% inside 24 months · the full income statement · the 19 August 6-K proposing $4.50B of convertible notes.
Stated plainly because it is the method this section advertises. The four grounded interviews completed and produced six load-bearing claims, but all six skeptic agents failed on a session limit before they could attempt a refutation. Rather than publish the interviews unchecked, every load-bearing claim above was verified by hand against the primary filings — the figures in the confirmed note are quoted from the 6-K exhibits, not from an agent’s summary. That is single-pass verification rather than adversarial verification, and it is weaker: nobody was tasked with trying to break these claims. Read this report with that discount applied. Two claims that could not be verified against a filing are marked below.
ARR of $3.0B at end-June, an FY2026 exit-ARR target of $7–9B, revenue guidance of $3.0–3.4B, capex of $20–25B, and a year-end contracted-power target of 5 GW are all call-sourced, not filed. The August 12 accession contains no press release exhibit and no deck, and the MD&A contains no ARR, megawatt, gigawatt or guidance figure at all. They are reported here as management statements from the earnings call, which is a materially weaker basis than the rest of this report. One source also reported contracted power as "over 4 GW" against the transcript’s 5 GW; both are recorded and neither is averaged.
The 19 August 6-K describes a proposed offering — $2.75B due 2030 and $1.75B due 2034, plus up to $675M of additional notes, "subject to market and other conditions". One research pass asserted it had priced upsized; EDGAR shows no filing after 19 August, so that could not be confirmed and is not stated as fact here. Treat the offering as announced and not yet completed as of this report.
Open questions
Who are Customers C, D and E? Three unnamed counterparties are 59% of revenue, the prior year’s top two have both dropped below the disclosure threshold, and a foreign private issuer is not required to name them.
What is the AI cloud segment’s operating loss on its own? The filing gives segment Adjusted EBITDA of $285.7M and a group operating loss of $(175.9)M, but does not push depreciation down to the segment — so the profitability of the actual business is not directly disclosed.
What happens to operating cash flow when deferred revenue stops compounding? It supplied 97.6% of the half’s operating cash, and the filing says it unwinds over one to five years.
Why extend server lives from four to five years effective January 2026, when Amazon shortened its own estimate the same month a year earlier citing the pace of AI development? The filing gives the assessment but not the underlying usage data.
Sources
- 1Nebius Q2 2026 Form 6-K, Exhibit 99.1 — Operating and Financial Review (2026-08-12)
- 2Nebius Q2 2026 Form 6-K, Exhibit 99.2 — financial statements: useful-life change, ClickHouse remeasurement, customer concentration, RPO
- 3Nebius Form 6-K (2026-08-19) — proposed $4.50B convertible senior notes offering
- 4Nebius newsroom — Q2 2026 results and shareholder letter
- 5Nebius Q2 2026 earnings call transcript — the only source for ARR, power and guidance
- 6stockanalysis.com — NBIS price, enterprise value and short interest
- 7stockanalysis.com — CRWV, for the peer comparison
- 8Nebius EDGAR filing index (CIK 0001513845)
Get the next Market Storm in your inbox
One email when a real market catalyst triggers a new report — the method, the numbers, and what the verification pass caught.