Nebius Q2 2026 Earnings: the ARR Bridge Meets the Vineland Question
Scenario Comparison
Revenue beats (above roughly $600M) AND full-year revenue or ARR guidance is raised above the current range
Revenue beats (roughly $575M to $650M) but full-year guidance is merely reiterated at $3.0-3.4B
Revenue misses (below roughly $520M), or full-year revenue or ARR guidance is cut, or management concedes the year-end connected-power target will not be met
All Scenarios
5Positioning Suggestions
→ Treat this as a sizing decision rather than a rating decision. Realized volatility into the print has been extreme ($148.22 on July 29 to $225.74 on August 4), the target spread across the street is 2.3x, and the expected return on our own framework has thinned to +4.7%. That combination argues for a smaller position through the print rather than a directional bet on it.
→ The accumulate line is $165 and spot is about 16% above it. A miss scenario that takes the stock toward the low $160s would put it at a level where our framework says add, so pre-committing to that level is more useful than pre-committing to a view.
→ Do not add ahead of the print on the strength of the $775M financing. That fact is three weeks old, already in the price, and answers the funding question rather than the delivery question that this print is actually about.
→ If the print beats and the guide is raised, the cohort trade (CRWV, APLD) usually moves later and further than NBIS itself; NBIS carries the idiosyncratic Vineland risk while the cohort only carries the multiple.
→ Hold conviction at HOLD through the print. The re-underwrite is owed immediately after, when the connected-power number and any useful-life disclosure exist.
Predictions
11Q2 2026 revenue comes in at or above the street estimate of about $575M
A beat alone is not a signal to add, because the guide action matters more; wait for the guidance line before sizing up.
Q2 2026 revenue comes in at or above $700M
Would invalidate the retired <$700M bear trigger's premise entirely and argue for re-opening the bull path at the re-underwrite.
Q2 2026 revenue comes in at or above $800M
The old bull threshold. At 39% above the street this is close to a tail outcome; if it happens, the framework needs rebuilding upward, not re-rating.
Full-year 2026 revenue guidance is raised above the current $3.0B to $3.4B range
The cleanest bull confirmation available in this print, because it is the only outcome that beats a street already at $3.36B.
Year-end 2026 ARR guidance of $7B to $9B is reiterated without reduction
A cut here is the single most damaging outcome for the thesis and would justify an immediate bear-probability increase above 25%.
Management discloses a specific connected-power figure in megawatts for the quarter
Non-disclosure should be read as negative regardless of the revenue line, given that Vineland became publicly contested two days before the print.
Connected power at quarter end is disclosed at 400MW or more
This is our own mid-year checkpoint on the path to 800MW-1GW. Clearing it rebuts the Davidson field research on its own terms.
The Vineland or New Jersey facility timeline is addressed explicitly on the earnings call, by management or in Q&A
Management concedes any slippage in the year-end 2026 connected-power target of 800MW to 1GW
A concession converts the bear case from a valuation argument into a delivery fact and should move the bear probability up at the re-underwrite.
A GPU or server useful-life depreciation schedule is disclosed in the filing or discussed on the call
If disclosed and materially longer than the five to six years peers use, that is a rebuild trigger for the thesis, not a re-rate.
Q2 2026 Adj. EBITDA comes in at or above the street estimate of about $173M
Key Questions
- How many megawatts are connected as of quarter end, and does management give the number at all? Refusing to quantify connected power in the quarter that Vineland became contested is itself an answer.
- Does the year-end 2026 ARR guidance of $7B to $9B survive intact? Exit ARR was $1.92B in Q1, so the bottom of the range needs roughly $1.69B of ARR added in each of three consecutive quarters, close to 90% of the entire existing run-rate every quarter.
- Is full-year revenue guidance raised? Visible Alpha consensus is already about $3.36B against a $3.0-3.4B guide, so a reiteration is a de facto miss against positioning.
- What exactly is said about Vineland, and is it scoped to the site or hedged back to contract-level language as the 2026-07-17 filing was?
- Is the GPU or server useful-life depreciation schedule disclosed? This is the specific gap the Burry short exploits, and the thesis currently has no assumption with which to answer it.
- Does the asset-level financing template repeat? The $775M facility of 2026-07-17 was framed as repeatable against more than $40B of contracted revenue; a second facility at a comparable spread converts a proof of concept into a funding channel.
- Any commentary on the asset-light partner model, where infrastructure partners run the Nebius stack in their own data centres? It would reduce capex intensity without reducing revenue, which is the most underpriced structural option in the story.
- Is capex guidance of $20B to $25B held? A further raise without a matching revenue revision is an existing live bear trigger.
Monitoring Checklist
Post-Event Results
Actual Outcomes
- • [R1] Q2 2026 group revenue $582.3M, +454% YoY from $105.1M and +46% QoQ, against a street estimate of about $575M — a 1.3% beat. Note the near-coincidence that Nebius AI cloud segment revenue was $574.9M ([R2]), a hair BELOW the same figure; the beat holds on the group line the street was estimating and on which the pre-event scenario bands were drawn.
- • [R1] Group adjusted EBITDA $236.2M at a 41% margin, against a street estimate of about $173M — a 37% beat, and up from $129.5M / 32% in Q1. The Nebius AI cloud segment ran a 50% adjusted EBITDA margin ([R2]); the group/segment gap is the drag from Avride and TripleTen.
- • [R1] Net loss from continuing operations $(190.4)M against $502.5M of income in Q2 2025; adjusted net loss narrowed 64% YoY to $(33.2)M.
- • [R2] ARR reached $3.0B at 30 June, +598% YoY and +56% from the $1.9B reported at 31 March. ARR is defined as AI cloud revenue in the last month of the quarter multiplied by twelve.
- • [T] FY2026 financial guidance REAFFIRMED unchanged: ARR $7-9B, group revenue $3.0-3.4B, group adjusted EBITDA margin approximately 40%, capex $20-25B. [R2] carries the reaffirmation in words; the metric-by-metric list is management-stated on the call, because the letter defers guidance detail to the webcast. Contracted-power guidance was not simply reaffirmed alongside the financial metrics: it was raised, from above 4 GW to 5 GW.
- • This is the setup risk the pre-event report named as the most underappreciated in the print, and it landed exactly as described: Visible Alpha FY26 consensus was already about $3.36B against a $3.0-3.4B guide, so a reiteration was a de facto miss against positioning. It did not trade that way. In our reading, the 2027 deal economics overwhelmed the 2026 line, though price action alone cannot confirm that cause.
- • [R2] Year-end 2026 CONTRACTED power target RAISED to 5 GW, from the greater-than-4 GW indicated last quarter. Contracted power is defined as secured by contracted land and power commitments.
- • [T] Year-end 2026 CONNECTED power guidance REAFFIRMED at 800MW to 1GW, answering Cantor directly: 'We still expect to meet this guidance from 800 MW to 1 GW of connected power this year.'
- • NO connected-megawatt figure was given for the quarter. The word 'connected' does not appear anywhere in [R1], [R2], [R3] or [R4], and in [T] it appears only inside the year-end-target answer. The nearest approach is an unquantified 'hundreds of megawatts of the generation behind the meter power already as of today', which describes generation access rather than connected capacity. Connected power at quarter end was not disclosed in Q2 2026. Neither the filings nor the call announce a change to disclosure policy, so whether this recurs next quarter is unknown. The mid-year 400-500MW watch trigger written against this metric did not fire this quarter, and it needs redesign because it cannot rely on the metric being published.
- • [T] A reframe that is not a concession, and it is the most consequential thing said on the call for the thesis. Having reaffirmed 800MW-1GW connected by year end, management separated connected power from revenue-generating power: 'connected power represents the data centers, and there are a few steps coming from the data centers to the revenues. You have to commission the data center, build the network, build the clusters, deploy the platform, then onboarding the customers, and then the revenue generation starts.' The monetisation of that capacity was placed 'throughout the first half of 2027'. The headline target survives; the revenue it was standing proxy for moves out.
- • [T] Vineland was the first question on the call, from Morgan Stanley. The site-layout-plan amendment hearing adjourned without a vote; the amendment follows a switch of the project's power source to Bloom on-site fuel cells. Management called the hearing part of the normal process and built into the schedule, said the building itself finished construction earlier this summer with engineering fit-out progressing, and stated 'no significant impact expected on the project timeline'. Andrey Korolenko explicitly kept the site in this year: 'the Vineland is the part of the 2026 capacity and connected power.'
- • [T] On the Microsoft contract, management repeated the contract-level formulation rather than a site-level one: 'we delivered all the tranches that we were required to deliver under the contract up to date. We have all the reasons to believe that we will continue to deliver the remaining tranches as required by the contract.' The pre-event report predicted this hedge and it appeared verbatim in structure.
- • [R4] The useful lives of servers and network equipment were EXTENDED FROM FOUR TO FIVE YEARS, assessed in January 2026 and applied prospectively from 2026-01-01. Q2 2026 effect: depreciation expense lower by $43.0M and net loss lower by $34.1M. Six-month effect: depreciation lower by $86.1M, net income higher by $75.7M. Depreciation of property and equipment was $253.8M in the quarter against $72.6M a year earlier.
- • The depreciation disclosure is the direct answer to the Burry short thesis of 2026-08-06, and it cuts against Nebius. The life extension was a 2026 change that flatters this year's reported profit by $86.1M across the first half, and a five-year server life sits at the low end of the five-to-six-year peer range rather than beyond it. It is disclosed, prospective and quantified, which is the good news; it is also non-cash support from a change in accounting estimate, flowing through operating expenses per the filed MD&A and therefore benefiting both operating income and net income, rather than a genuine improvement in operating performance.
- • [R2] Four landmark AI cloud deals closed in Q2 averaging more than $1B TCV each, including Reflection and Cohere plus an unnamed scaled US neolab and a US quant trading firm. ACV repriced above $20M per megawatt with a $20-25M per megawatt yield, prepayments covering 50-60% of associated capex, and expected payback shortened to 1 year 10 months from a historical two-to-three-year range. Q2 TCV grew nearly 4x QoQ and new-customer TCV more than 9x. Older-generation GPU pricing was more than 30% higher than Q1.
- • [T] The first capacity auction cleared 15% above the highest price ever charged for the Blackwell generation and 20% above pipeline. Short-term capacity deals are being negotiated at $40-50M per megawatt and sometimes above, with the first such deal signed in Q3 and going live in Q4. Both deal types come online later this year and so are explicitly stated to have no material effect on 2026 revenue guidance.
- • The pre-event framework asked whether this print would separate a demand problem from a delivery problem. It answered decisively on demand: pricing, prepayments, deal size and auction clearing all moved the right way at once. It did not answer delivery, because the metric that would have answered it was not disclosed.
- • [R2] Balance sheet and funding: capex approximately $5.7B in the quarter, operating cash flow $2.3B, $8B of cash and equivalents at quarter end, more than $9B of customer prepayments expected across 2026 with roughly 70% of Q2-closed deals carrying one. The ATM was used for the first time: 12.7M Class A shares at a weighted-average $223.6 for approximately $2.8B gross, with 12.3M shares remaining at 30 June. The July asset-backed facility was $775M at SOFR plus 250bp against deployed GPU infrastructure and investment-grade contracted cash flows, with more than $40B of committed backlog available to borrow against.
- • [T] The asset-light partner model was introduced this quarter and already contributed to the margin. Partners finance, build and operate; Nebius supplies the full-stack platform, systems architecture and go-to-market. Management reports dozens of partner inquiries and calls the model early-stage. It is the structural option the pre-event report flagged as most underpriced, and it is now real enough to appear in the margin bridge.
- • [T] No formal 2027 guidance; management committed only to providing it later this year, while stating that 2027 capacity could have been sold out already at current terms and that it is deliberately not doing so. Plans call for deploying more than 1 GW of new capacity per year starting in 2027.
- • [R2] First NVIDIA Vera Rubin NVL72 systems already received and in validation. [T] Management said on the call it expects to start deploying Vera Rubin late this year or early next year, continuing through the following year; the shareholder letter does not itself give a deployment date. [R2] Token Factory inference workloads more than tripled in the quarter and Tavily's developer community grew past 2.5M from 1M in February.
- • [M] Market reaction: NBIS closed 2026-08-12 at $259.20 against a $193.23 close the day before, +34.1%, and ran on to $277.68 by 08-14 before giving most of it back to $223.90 by 08-19. CRWV closed +19.3% on the day, APLD +4.9%, NVDA +3.0%; MSFT closed -2.3% and META -3.4%, neither attributable to this print.
- • The move is the calibration lesson. The pre-event report modelled a reiterated guide as a disappointment against positioning and assigned 0.30 to that arm, and the guide was indeed only reiterated — yet the stock added a third of its value in a session. Our reading, not an established fact from the price alone, is that the market repriced 2027 deal economics and the 5 GW contracted-power raise, both of which the scenario matrix treated as colour rather than as the payload. Sizing into the print on a 2026-revenue frame was the wrong axis.
- • Scoring in this report is complete. As of this writing, downstream reconciliation is outstanding: NBIS's monitoring event calendar still carries this print as upcoming, and the matching thesis catalyst is not yet marked realized. That reconciliation is being handled separately.
Market Reaction
Thesis Updates Needed
Trade Recommendations
→ Prediction #1 (revenue beat) and #11 (adjusted EBITDA beat) both confirmed, so their conditional trades activate. The #1 trade said explicitly to wait for the guidance line before sizing up, and the guidance line was a reiteration, so the beat alone never became an add signal. That instruction was correct in logic and wrong in outcome: the stock closed +34.1% the same session, which in our reading reflects 2027 economics the pre-event frame was not watching, though causation cannot be established from price alone.
→ Prediction #4 (guidance raise) unconfirmed, so the cleanest bull confirmation available in this print did not occur and its trade voids. Do not treat the price move as a substitute for it.
→ Predictions #6 and #7 unconfirmed, and the #6 conditional trade said non-disclosure should be read as negative regardless of the revenue line. That reading stands and should be carried into the re-underwrite as a disclosure-quality mark against the issuer, not reversed because the stock rose.
→ Prediction #9 unconfirmed, so the bear-probability increase it was conditioned on does not fire. The condition it was written to catch has partly happened by another route: connected power was reaffirmed but its revenue was placed in the first half of 2027. Treat that as a monetisation-lag input at the re-underwrite rather than as a fired trigger.
→ Prediction #10 confirmed. Its conditional trade said a materially longer useful life than the five-to-six-year peer range is a rebuild trigger rather than a re-rate. Four to five years is inside the peer range, at its low end, so no rebuild is triggered, but the $86.1M first-half depreciation relief must be carried as explicit non-cash, operating-expense-line support to reported profit, from a change in accounting estimate rather than an improvement in operating performance.
→ The pre-event accumulate line of $165 and the $148-226 realized-volatility band are both obsolete after a $259.20 close, a $277.68 high on 08-14 and a $223.90 close on 08-19. Do not chase into the post-print range; the give-back from the high is larger than the original pre-print range. Re-establish an accumulate level only after the re-underwrite reprices 2027, which is the year that actually moved.
→ The pre-event note that the cohort trade moves later and further than NBIS itself was wrong for this print: CRWV moved the same session at +19.3% and gave back with NBIS. Cohort beta here is contemporaneous, not lagged.
→ Hold conviction at HOLD. The re-underwrite is owed now and it is a rebuild rather than a re-rate, because the metric the framework was scored on (2026 revenue against a $575M street bar) is no longer the metric the market is pricing.