| Aug 12, 2026 | +34.14% | Q2 FY2026 | Read transcript briefingQuarter in one view- Beat the guided margin dip: group revenue $582M, +454% YoY, +46% QoQ; Nebius AI $575M, +514% YoY, 98% of group. Group adj. EBITDA $236M / 41% margin vs the explicitly guided Q2 dip below Q1's 32% — the "nonlinear path" resolved upward a quarter early. Nebius AI margin 50% (from 45%).
- ARR $3.0B at end-June, +58% from $1.9B — ~$1.1B of Q2 adds, a step up from the ~$650M quarterly pace of Q4/Q1. Full-year guidance reaffirmed across all metrics (ARR $7–9B, revenue $3.0–3.4B, ~40% margin, CapEx $20–25B).
- Deal economics disclosed for the first time: four ~$1B+ midterm AI cloud deals closed in Q2 (Reflection, Cohere, a scaled U.S. neolab, a large U.S. quant firm) at $20–25M/MW with prepayments covering 50–60% of associated CapEx; short-duration premium deals being negotiated at $40–50M+/MW; first capacity auction cleared 15% above the highest prior Blackwell price (20% above pipeline pricing).
- Contracted backlog now quantified: >$40B — first backlog disclosure after multiple quarters of refusal, though still without composition/concentration detail.
- Funding mix shifted to equity: ATM used for the first time — 12.7M shares at $224 weighted average, ~$2.8B gross; first asset-backed facility $775M at SOFR+250 in July; cash $8.0B (down from $9.3B despite $2.3B operating cash flow — CapEx of ~$5.7B in Q2 is the draw).
- Contracted power target raised again: 5GW by end-2026 (from ≥4GW), with >1GW/yr of new capacity planned from 2027.
What management is focused on- A three-tier deal architecture (new framing this quarter): midterm 1–3yr AI cloud contracts ($20–25M/MW), short-duration ≤6-month premium capacity ($40–50M/MW+), and long-term investment-grade contracts used as financing collateral. The strategic claim: "we could sell today our entire 2027 capacity... but we are not doing this" — deliberately holding capacity for higher-priced short-term sales.
- Price discovery as a product: the capacity auction is framed as a real-time market signal tool, not just a sales channel; management plans repeat rounds.
- Asset-light model: partners finance/build/operate facilities; Nebius supplies platform and demand for "high-margin revenue" with minimal balance-sheet capital. Already contributing to Q2 revenue and margin per the CFO; "dozens of inquiries" from potential partners; targeted at 2027+ capacity.
- Selling closer to deployment: tactically shortened how far ahead capacity is sold, explicitly to capture rising prices — a shift from the prior "pre-sell 2027" posture toward optionality.
- Financing hierarchy restated: operating cash flow → customer prepayments (>$9B expected in 2026) → asset-backed debt against the $40B backlog → corporate debt (currently "almost no corporate-level debt") → equity/convertibles. GPU-as-asset-class financing flagged as an emerging additional source.
- Open-weight / inference stack: Token Factory day-0 support for frontier open models (GLM 5.2 cited with 100% quality score via Artificial Analysis), Eigen/Clarifai "fully integrated" into Token Factory, ARM/CPU deployments added for agentic workloads.
Key numbers and quarter mechanics- Revenue $582M (+454% YoY, +46% QoQ); Nebius AI $575M (+514%); non-core implied ~$7M.
- ARR $3.0B at June 30 vs $1.9B at March 31 → ~$1.1B adds, ~70% above the prior two quarters' pace. Drivers cited: Q1 capacity additions, higher utilization (infrastructure efficiency), and early asset-light/Token Factory/M&A revenue.
- Group adj. EBITDA $236M / 41% (vs $129.5M / 32% in Q1); Nebius AI 50% margin (from 45%). Note the transcript contains a CFO misstatement ("$286 million" corrected to $236M in the transcript itself) — the core EBITDA dollar figure should be verified against the press release.
- Q2 CapEx ~$5.7B — first quarterly CapEx disclosure after repeated requests; annualizes within the $20–25B guide with H2 weighting.
- Cash $8.0B (from $9.3B); operating cash flow $2.3B again; prepayments at all-time high, ~70% of Q2 deals included upfront payment; >$9B of 2026 prepayments expected.
- ATM: 12.7M shares at $224 avg = ~$2.8B gross; 12.3M shares remain under the 25M-share program.
- Asset-backed facility: $775M, SOFR+250, mid-single-digit all-in, secured by deployed GPUs and investment-grade contracted cash flows; >$40B committed backlog available for similar structures.
- Guidance reaffirmed: ARR $7–9B, revenue $3.0–3.4B, ~40% group margin, CapEx $20–25B. Connected power 800MW–1GW by year-end reaffirmed, but Korolenko added that revenue activation from that capacity runs "throughout the first half of 2027" — a meaningful lag caveat.
- 2027 framing: Q2's $20M+/MW, <2-year-payback deals come online "from late Q4 and onwards" and serve as the 2027 pricing baseline; formal 2027 guidance promised "later this year."
- Vineland: building construction finished earlier this summer; engineering fit-out progressing; power source switched to Bloom fuel cells (on-site, "no significant impact expected on the project timeline"); site-plan amendment hearing adjourned without a vote; Vineland confirmed as part of 2026 connected capacity.
- Vera Rubin: in labs with expected results; deployment starts late 2026 / early 2027, ramping through 2027; GB300→Rubin transition described as technically easier than the prior generational step.
- Still not disclosed: utilization %, customer concentration percentages, backlog composition by counterparty, acquisition prices, Token Factory revenue, depreciation-change quantification.
Product and launch scorecard- Capacity auction (new): first round cleared 15% above the highest price ever charged for Blackwell, 20% above pipeline pricing; winner plans to participate in future rounds. Real evidence of pricing power and a repeatable price-discovery mechanism; uses "a small portion" of capacity.
- Short-term scale training deals (new): 3–6 month dedicated GB300 clusters at $40–50M/MW+ for time-boxed training/post-training runs; one signed "just recently." Confirmed take-or-pay business. Revenue impact begins late 2026 — explicitly immaterial to 2026 guidance.
- Asset-light model (new): already contributing to Q2 revenue and margin per CFO; "dozens of inquiries" post-announcement; no signed partner, capacity, or revenue figures disclosed — early-stage but financially visible faster than expected.
- Token Factory: Eigen and Clarifai teams "fully integrated... already shipping inside our roadmap"; day-0 support for Nemotron Ultra, GLM 5.2, Kimi K3, DeepSeek Flash, Minimax 3; GLM 5.2 implementation scored 100% quality with leading Artificial Analysis performance; powers Nebius's own "Echo" infrastructure agent. Named users now include Revolut and MasterCard (agentic workflows). Still zero revenue, token-volume, or ARR metrics — four quarters in.
- Tavily: first full quarter consolidated; developer community 2.5M, up from 1M in February; launched keyless search for autonomous agents; enterprise certifications achieved. First hard adoption metric for any acquisition — though no revenue.
- Landmark AI cloud deals: Reflection and Cohere named; all four were competitive displacements of incumbent suppliers (in some cases hyperscalers), validated via POCs; expansion discussions (including Vera Rubin and Token Factory inference) already underway with all four.
- Pipeline: "stepped up again in Q2," includes multiple $1B+ opportunities across AI-natives, neolabs, enterprises.
Sell-side read-through- Named questioners: Ryan Lantz (Morgan Stanley, twice), Alex Duval (Goldman), Arsenije Matovic (Wolfe), Tyler Radke (Citi), James Kisner (WaterTower), Rob Oliver (Baird, twice), Brett Knoblauch (Cantor), Stefan Slowinski (BNP Paribas). Portal-curated format continues.
- Vineland got a fuller answer than last quarter — with new facts and a new risk: management disclosed the Bloom fuel-cell power switch (the actual cause of the site-plan amendment), confirmed building completion and tranche delivery to date, and put Vineland inside 2026 connected capacity. But the hearing adjourned without a vote, and "optimistic that once the public has been heard, this will move quickly" is not a schedule. The permitting risk is now concrete, not hypothetical.
- The 2027 question produced the call's most important forward statement: Lantz's framing question extracted the $20M+/MW, <2-year-payback baseline for 2027 pricing, the "could have sold out 2027 today" claim, and a promise of formal 2027 guidance later this year.
- Financing question (Matovic) was answered with specifics: SOFR+250 on the $775M facility, the $40B borrowable backlog, and an explicit hierarchy — but note the ATM was already used for $2.8B before this call, so "minimizing dilution" rhetoric now sits against executed dilution.
- Allocation question (Radke) revealed a priority ordering: existing customers → new logos → terms (price, then prepayment, then duration) — and the deliberate shortening of how far ahead capacity is sold. This is a real strategy disclosure, not boilerplate.
- xAI question (Kisner) got philosophy: new entrants "validate the market"; no direct comment on premium-pricing competition or renewal pricing.
- Not asked or not pressed: the Q2 margin beat versus the guided dip (nobody asked why the dip didn't happen), depreciation-policy impact on the 50% core margin, concentration percentages despite the $40B backlog disclosure, ClickHouse/Avride/TripleTen status, Microsoft delivery cadence (only Vineland-adjacent), asset-light margin structure and counterparty risk.
Management credibility- The margin bridge over-delivered: last quarter management guided a Q2 dip below 32%; it printed 41% group / 50% core. Beating your own conservative path is credibility-positive, but it also means the stated bridge (front-loaded costs, back-end capacity) was wrong within one quarter — the forecasting precision claimed last quarter didn't hold, favorably.
- Delivery claims continue to hold under direct questioning: all contracted tranches delivered to date; Vineland building complete; the Bloom pivot explained rather than concealed. The pattern of answering site-level questions with more specificity each quarter continues.
- Funding promises kept — but the mix changed: last quarter's "evaluating change" on the ATM became $2.8B of executed equity issuance at $224. Management now calls the ATM "an option in our arsenal rather than a commitment," which is accurate but softer than the "minimizing shareholder dilution" principle stated in the same answer. Dilution is no longer a tail risk; it happened.
- Disclosure improved materially where management chose: quarterly CapEx ($5.7B), backlog (>$40B), deal-level unit economics ($20–25M/MW midterm, $40–50M/MW short-term, 50–60% CapEx prepayment coverage), facility pricing (SOFR+250). The holdouts are now narrower: concentration, utilization, acquisition prices, Token Factory revenue, depreciation quantification.
- New looseness to note: (1) the "we could sell all of 2027 today" claim is unverifiable and conveniently supports withholding capacity for spot pricing; (2) the 800MW–1GW year-end target was reaffirmed while quietly adding that revenue activation slips into H1 2027; (3) cash fell $1.3B QoQ despite $2.3B operating cash flow and $2.8B ATM proceeds — the CapEx draw is real and accelerating; (4) the CFO's core-EBITDA figure was misstated on the call and corrected in the transcript.
- Consistency check: "we choose when to sell, to whom, and on what terms" is now backed by observable behavior (auction, withheld 2027 capacity, short-term premium deals) — the strategy and the evidence match this quarter.
What changed versus the prior quarter- The guided Q2 margin dip didn't happen: 41% group / 50% core vs the promised dip below 32%/45% — the 40% full-year guide now looks conservative.
- ARR adds accelerated: ~$1.1B in Q2 vs ~$650M in each of the prior two quarters.
- ATM moved from "evaluating" to executed: $2.8B raised at $224/share; 12.3M shares remain.
- First asset-backed facility done: $775M at SOFR+250 — the Meta/Microsoft-style contract-collateral financing thesis is now demonstrated in the market, at scale to come against $40B of backlog.
- Backlog disclosed for the first time: >$40B — ends (partially) a multi-quarter refusal, though composition is still hidden.
- Quarterly CapEx disclosed for the first time: ~$5.7B.
- Deal-level economics disclosed: $20–25M/MW midterm with 50–60% CapEx prepayment coverage; $40–50M/MW short-term — the first unit-economics transparency in the company's history.
- New go-to-market motions: capacity auction (15% above prior peak pricing) and short-term premium deals; deliberate shift to selling closer to deployment.
- Asset-light model launched and already revenue-contributing — a genuinely new business line appearing between quarters.
- Contracted power target raised again: ≥4GW → 5GW by end-2026; >1GW/yr build rate claimed from 2027.
- Vineland power source switched to Bloom fuel cells — the previously dodged site-delay story now has a concrete cause and a concrete open permitting step.
- Cash declined QoQ ($9.3B → $8.0B) despite $2.3B operating cash flow and $2.8B ATM — CapEx is now the dominant cash variable.
- Meta/Microsoft went unmentioned by name — no new hyperscaler contract news this quarter; the narrative rotated to AI-cloud deal flow.
Bull case- Pricing is rising and Nebius is proving it in public: auction cleared 15% above the all-time high, short-term deals at $40–50M/MW vs $20–25M/MW midterm, and management claims it could sell out 2027 today at <2-year paybacks — but is holding capacity for better prices. That is a seller's market with a disciplined seller.
- The margin story is ahead of plan: 50% core / 41% group in the quarter that was supposed to dip; asset-light, Token Factory, and M&A are already margin-accretive contributors, and owned-capacity margin improvement is still ahead (H2 2027).
- ARR adds accelerated to ~$1.1B while most 2026 capacity hasn't come online yet — the H2 capacity ramp plus late-Q4 deal activations stack on top of an already-raised run-rate.
- Financing is now demonstrated, not theoretical: $2.8B ATM at $224, $775M asset-backed at SOFR+250, >$9B of 2026 prepayments, $40B of borrowable backlog, near-zero corporate debt — five funding levers, four already used.
- Customer diversification is real and competitive: four $1B+ deals won against incumbent suppliers (including hyperscalers) via POCs — Reflection, Cohere, a neolab, a quant fund — plus multiple $1B+ opportunities in pipeline. The two-counterparty story is broadening.
- Asset-light is a free option on capacity: if partner capital funds facilities while Nebius captures high-margin platform revenue, the GW/yr ambition partially decouples from the balance sheet.
- Vera Rubin transition de-risked: already in labs, easier migration than GB300, deployment late 2026/early 2027 into a market where Nebius has early access via the NVIDIA relationship.
Bear case- Dilution happened: $2.8B of equity issued at $224 within weeks of "evaluating" language, with 12.3M shares still available and a >1GW/yr 2027 build to fund. The "minimizing dilution" principle is stated in the same breath as open-ended equity optionality.
- Cash is now declining: $8.0B vs $9.3B despite $2.3B operating cash flow and the ATM — Q2 CapEx of $5.7B annualizes to ~$23B, and the 2027 build is larger. The funding machine must keep running at full speed; any prepayment or debt-market hiccup forces more equity.
- Vineland permitting is an open item: hearing adjourned without a vote, and this site is inside 2026 connected capacity. The Bloom switch mitigates community concerns but adds a novel power technology to the critical path. Community opposition is no longer a prepared-answer topic — it's a live schedule risk.
- The 800MW–1GW year-end target now carries a revenue-timing caveat: "active throughout the first half of 2027" means the connected-power headline and the revenue are separated by months — the H2-weighted year still has execution risk compressed into Q4.
- "We could sell all of 2027 today" is unfalsifiable and strategically convenient: it justifies withholding capacity for spot pricing, which is a bet that prices keep rising. If the market softens, Nebius is long uncontracted capacity with a $20–25B annual CapEx program.
- Concentration disclosure was again avoided: the $40B backlog number was given without counterparty composition — Meta and Microsoft almost certainly dominate it, and the asset-backed financing model depends on those investment-grade cash flows.
- The 50% core margin still sits on the unquantified 4→5yr depreciation change, and nobody asked. The CFO also misstated core EBITDA on the call — small, but sloppy on the quarter's headline metric.
- Token Factory remains metric-free after four quarters of "primary inference product" status; asset-light economics (margin split, counterparty risk, capacity commitments) are entirely undisclosed.
Next-quarter watchlist- H2 capacity ramp execution: does Q3 deliver the "very significant" capacity additions needed to hit 800MW–1GW connected by year-end, and does the Q2-deployed capacity contribute to Q3 revenue as guided?
- Vineland: vote outcome on the amended site plan, Bloom fuel-cell deployment pace, and any slippage in the 2026 connected-capacity classification.
- 2027 guidance: management promised formal 2027 outlook "later this year" — capacity, pricing baseline ($20M+/MW), and the asset-light contribution are the key inputs.
- Funding cadence: size and pricing of the next asset-backed facilities against the $40B backlog; any corporate-level debt; whether the remaining 12.3M ATM shares get used; progress on GPU-as-asset-class financing.
- Auction repetition: second-round clearing price vs the first (15% above prior peak) — a direct read on whether AI compute pricing is still rising.
- Short-term premium deals: how much capacity gets allocated to $40–50M/MW deals vs midterm contracts, and whether the "sell closer to deployment" tactic holds if pricing flattens.
- Asset-light first signed partner: capacity committed, revenue/margin structure, and whether "dozens of inquiries" convert.
- Token Factory metrics: any revenue, token-volume, or ARR disclosure; conversion of Revolut/MasterCard-type usage into contracted inference revenue.
- Concentration and backlog composition: now that $40B is public, pressure for counterparty breakdown will build — four calls running without it.
- Depreciation quantification beneath the 50% core margin; Avride/TripleTen deconsolidation progress; Microsoft delivery cadence through year-end (unaddressed this quarter).
|
| May 13, 2026 | +15.72% | Q1 FY2026 | Read transcript briefingQuarter in one view- Q1 delivered on both axes: group revenue $399M, +684% YoY, +75% QoQ; Nebius AI (ex-TripleTen/Avride) revenue $390M, +841% YoY, +82% QoQ, 98% of group. Core ARR $1.9B at end-March, +50%+ from $1.25B — the Q4 momentum carried into Q1 rather than reverting to lumpiness.
- Margins inflected hard: group adjusted EBITDA $130M (32% margin) vs $15M (7%) in Q4; Nebius AI adjusted EBITDA margin 45%, up from 24% — nearly doubling in one quarter. Full-year ~40% group guide reiterated, now with a visible quarterly path (Q2 dip, Q3 back to Q1 levels, Q4 higher).
- CapEx raised again: $20B–$25B for 2026 (from $16B–$20B), explicitly for 2027 capacity with customer commitments already in place. Contracted power >3.5GW now, target raised to ≥4GW (from >3GW); new 1.2GW owned Pennsylvania site announced; owned capacity now >75% of contracted power.
- Balance sheet transformed: $9.3B cash (from $3.0B) via $4.3B converts (1.25% / 2.60% coupons), $2B NVIDIA equity investment, and record customer prepayments driving $2.3B Q1 operating cash flow. >90% of the *prior* CapEx range now secured by cash and contractual commitments.
- Second Meta contract disclosed: $27B, 5-year — $12B committed dedicated compute (delivery early 2027) plus a $15B capacity option where Meta backstops any capacity Nebius chooses not to sell to AI cloud customers at potentially higher prices. Management: "if the market remains strong, we should generate more than $27B."
- Net income $621M, flattered by a non-cash ClickHouse valuation uplift from its recent funding round.
What management is focused on- Four-dimension framework: capacity/scale, product/functionality, customers/demand, capital — CEO's organizing structure for the whole call; execution across all four is the stated job.
- Building for 2027, not 2026: the CapEx raise, Pennsylvania site, and Meta #2 are all framed around 2027 capacity "where we already have customer commitments in place." H1 2027 capacity additions will exceed all of 2026's.
- Full-stack / inference positioning: three acquisitions this year (Tavily, Eigen, Clarifai) all pointed at inference optimization and agentic; Eigen "#1 speed inference provider by NVIDIA"; Token Factory named the "primary inference product." Inference called "the fastest-growing segment in our stack."
- NVIDIA alignment as strategy: $2B equity investment, Exemplar Cloud status on GB300, early access to Vera Rubin/Vera CPU SKUs, and a line of sight to 5GW of capacity commitment by end-2030 — supply-chain certainty as competitive advantage.
- Financing architecture as a product of contracts: the Meta/Microsoft contracts are explicitly described as collateral that "unlocks billions" in asset-backed financing at rates tied to Meta/Microsoft credit ratings; corporate debt and the ATM (25M Class A shares, unused but "evaluating") round out the stack.
- Community/political risk management: a full answer on U.S. data-center opposition (transparency, town halls, Nebius Academy) — a new topic this call, signaling siting risk is now on the radar.
Key numbers and quarter mechanics- Revenue $399M (+684% YoY, +75% QoQ); Nebius AI $390M (+841% YoY, +82% QoQ). Non-core (TripleTen + Avride) implied at only ~$9M — now consolidated and named, a disclosure improvement.
- ARR $1.9B at March vs $1.25B at December → ~$650M of Q1 adds, matching the Q4 run-rate.
- Group adj. EBITDA $130M / 32%; Nebius AI margin 45% (from 24%); the group-core gap is Avride + TripleTen investment drag. Management intends to find partners and deconsolidate both eventually.
- Cash $9.3B; operating cash flow $2.3B (vs -$198M a year ago), "primarily upfront payments." Converts: $4.3B gross, 1.25%/2.60% coupons; NVIDIA equity $2B.
- CapEx $20B–$25B; component inflation impact on 2026 program only "low single digits" as % of spend (2026 supply largely locked in 2025) — the raise is volume/2027-driven, not cost-driven.
- Guidance reiterated: ARR $7B–$9B, revenue $3.0B–$3.4B, group adj. EBITDA margin ~40%. Explicit quarterly margin path: Q2 dips below Q1 (back-end-weighted capacity, front-loaded hiring/M&A costs), Q3 returns to Q1 levels, Q4 higher — "nonlinear progression."
- Meta #2 mechanics: $12B committed, delivery starting early 2027; $15B option at Nebius's discretion across the 5-year term; designed to be asset-backed-financeable while retaining upside to sell at market prices.
- Capacity schedule: 2026 heavily H2-weighted (Q3 "very significant," Q4 also); Q1 2027 brings Alabama and first Missouri online; Pennsylvania lights up end-2027 with 250–300MW, then ~300MW/year to 1.2GW by ~2030.
- Microsoft: first tranche delivered November; schedule runs to end-2026, ramping from midyear, most volume in Q3–Q4. Vineland NJ delay reports addressed only with "we delivered all our capacity commitments" — no direct confirmation or denial of site-level delays.
- Demand metrics: pipeline +3.5x QoQ (AI cloud only, excludes hyperscaler deals); 4+ customers competing per GPU; prices raised again in Q1 and still selling out "across all chip types"; durations extending; prepayments growing including from hyperscalers.
- Still not disclosed: Q1 CapEx spend, backlog/RPO, utilization %, customer concentration percentages, Microsoft/Meta #1 contract terms, acquisition prices.
Product and launch scorecard- Token Factory — first real traction evidence: named customers Revolut and monday.com (new wins), plus usage by AI-native builders; called the "primary inference product" with "good product-market fit." Still no revenue, token-volume, or ARR metrics — but this is the first quarter with named paying logos after two quarters of silence.
- Eigen AI + Clarifai (new M&A): inference optimization at model level (Eigen) and system level (Clarifai), folded into Token Factory; Eigen carries an NVIDIA #1 inference-speed recognition. No price, revenue, or integration metrics — rationale-stage, like Tavily last quarter.
- Tavily: referenced as proof of "proven developer adoption" M&A criteria; still no purchase price or contribution disclosure.
- Aether 3.5 launched this quarter — version cadence continues, no adoption data.
- NVIDIA partnership deliverables: Exemplar Cloud on GB300 (training), early Vera Rubin/Vera CPU deployment access, joint inference/agentic software work, physical-AI collaboration, 5GW-by-2030 line of sight. Vera Rubin selling starts H2 2026.
- Customer wins named: Sword Health (healthcare/life sciences), Rhoda and 1X Technologies (physical AI/robotics), Revolut and monday.com (Token Factory) — broadening beyond AI natives into fintech, healthcare, robotics, enterprise.
- Go-to-market build-out: new regional GMs (Americas, APJ, Middle East) — evidence of enterprise-sales scaling.
- Avride/Toloka: Avride now explicitly a consolidated EBITDA drag with a stated intent to find partners and deconsolidate — more transparency than prior quarters, but no operating metrics. Toloka absent again. ClickHouse appears only as a mark-to-market gain.
Sell-side read-through- Named questioners: Alex Duval (Goldman, twice), James Kisner (WaterTower, twice), Andrew Beale (Arete), Josh Baer (Morgan Stanley), Alex Platt, Tal Liani (BofA), Nehal Chokshi (Northland). Still IR-curated portal format.
- The margin-bridge question got a real answer this time: Kisner's 45%→40% step-down question produced a specific quarterly walk (Q2 dip from front-loaded hiring/M&A costs and back-end capacity, Q3 recovery, Q4 higher). This repairs last quarter's biggest credibility gap — though the depreciation-change effect remains unquantified and unasked.
- Vineland delay question was partially evaded: Baer asked directly about media-reported NJ delays; the answer recited delivery performance ("we delivered all our capacity commitments") and the Microsoft schedule without addressing the site report itself. Microsoft volume confirmed as Q3/Q4-weighted — slippage risk lives in H2.
- Meta option structure was the best disclosure of the call: Platt's question extracted the full $12B/$15B mechanics, including the arbitrage logic (finance against Meta's commitment, sell at market premium). Notably, management confirmed willingness to redirect Meta-optioned capacity to better-priced customers.
- Concentration question got philosophy, not numbers: asked directly, answered with "we are very intentional" and diversification narrative — zero quantitative disclosure of Meta/Microsoft share of ARR or contracted revenue. Third consecutive call avoiding this.
- CapEx raise attribution was clean: Korolenko explicitly separated volume (2027 commitments) from inflation (low single digits) — a direct, quantified answer.
- Not asked or not pressed: depreciation-policy impact on the 45% core margin, Q1 CapEx spend vs guide, utilization %, acquisition prices, ATM dilution math, Toloka, what "evaluating change" on the ATM means.
Management credibility- The margin bridge was rebuilt: last quarter's weakest point (7%→40% asserted) is now a stated quarterly path with named drivers, and Q1's 32% group / 45% core prints are ahead of the trajectory the 40% guide implies. Credibility repaired, though Q2 is now a self-declared down quarter to verify.
- Delivery claims continue to hold: Meta #1 fully delivered in Q1; Microsoft on schedule per management. The pattern of beating then raising targets continues (power: 2.5GW → 3GW → 3.5GW contracted, ≥4GW target).
- Funding promises kept ahead of schedule: last quarter's "~60% funded" became ">90% of the prior range secured," with converts done at 1.25%/2.60% and a $2B strategic equity check — the debt-market risk flagged last quarter is substantially reduced for the old plan, though the *incremental* $4–5B raise still needs financing.
- New looseness to note: (1) "evaluating change" on the ATM is a softening of last quarter's "no concrete plans" — equity dilution is back on the table; (2) net income of $621M is mostly a ClickHouse mark — headline profitability optics flatter operations; (3) the depreciation 4→5yr change still sits unquantified beneath the 45% core margin; (4) the Vineland question was answered around, not through.
- Disclosure improving at the edges: non-core revenue now separable (~2% of group), Avride/TripleTen drag explicitly quantified as the group-core margin gap, Meta #2 structure fully explained. Concentration, backlog, and utilization remain the holdouts.
- Transcript hygiene still imperfect (garbled names persist) but substance and specificity were high this quarter.
What changed versus the prior quarter- CapEx raised again within one quarter: $16–20B → $20–25B, now explicitly a 2027 build with commitments attached.
- Power targets raised twice in a quarter: >2GW contracted (Feb) → >3.5GW contracted, ≥4GW target; first owned gigawatt-scale site #2 (Pennsylvania, 1.2GW); owned share now >75%.
- Second Meta contract ($27B) — the single largest new fact; introduces the novel $15B backstop/option structure and confirms Meta revenue into 2027+.
- Funding executed: $4.3B converts + $2B NVIDIA equity + $2.3B operating cash flow → $9.3B cash; NVIDIA is now a shareholder with a 5GW-by-2030 alignment.
- M&A accelerated: from one acquisition (Tavily) to three (Eigen, Clarifai added), all inference/agentic-themed.
- Margin story moved from promise to print: 45% core margin makes the 40% group guide look conservative-leaning rather than heroic; Q2 guided dip is the new test.
- Token Factory got its first named customers; inference declared the fastest-growing segment.
- ATM language shifted: "no concrete plans" → "evaluating change on it."
- Avride/TripleTen strategy clarified: consolidate now, find partners, deconsolidate later.
- New risk topic surfaced: U.S. data-center community opposition warranted a prepared answer.
Bull case- ARR adds held the Q4 run-rate (~$650M again) — the delivery machine is now repeatable, not a one-quarter cliff-clearing event, and Q1 pipeline grew 3.5x QoQ excluding hyperscaler deals.
- The 40% margin guide is now de-risked: core already at 45%, group at 32% in Q1, with a credible quarterly bridge and pricing still rising across all GPU generations.
- 2027 is being pre-sold, not speculated: the CapEx raise is tied to signed commitments (Meta $12B starting early 2027), and the $15B Meta option creates a financed floor with market-price upside — "more than $27B" if demand holds.
- Financing risk largely cleared for the prior plan: >90% secured, converts at 1.25%/2.60%, $9.3B cash, prepayments at records — the customer-prepayment flywheel is compounding.
- NVIDIA's $2B equity + 5GW alignment + early Rubin access is a supply-chain moat and a third-party validation of the platform; Vera Rubin selling starts H2 2026 into a sold-out market.
- Demand broadening is visible in named logos: Revolut, monday.com, Sword Health, 1X — diversification beyond the two hyperscalers is actually happening at the AI cloud layer.
- Pricing power persists: prices raised again in Q1, sold out across all chip types, durations extending, 4+ bidders per GPU.
Bear case- CapEx has now been raised three times in ~six months (~$5B 2025 → $16–20B → $20–25B); the incremental $4–5B is unfunded and management is openly "evaluating" the ATM — dilution risk is live again despite the strong cash balance.
- Q2 is a guided margin down-quarter with H2-weighted capacity and Q3/Q4-weighted Microsoft volume — the year is structurally back-loaded, so any H2 site or delivery slippage (note the unanswered Vineland delay reports) compresses both revenue and the 40% margin into Q4.
- Concentration deepened materially: a second $27B Meta contract plus Microsoft means two counterparties anchor the financing stack, the prepayment flywheel, and 2027 revenue — and concentration/backlog disclosure was refused again when asked directly.
- Headline net income is mostly a ClickHouse mark ($621M, non-cash) — GAAP profitability optics will reverse if private marks do.
- The 45% core margin still sits on an unquantified depreciation extension (4→5 years); no analyst pressed it, and management didn't volunteer it.
- Pennsylvania's full 1.2GW lands ~2030 (250–300MW end-2027, ~300MW/yr) — headline GW numbers remain far ahead of connected reality; 2026 connected-capacity progress wasn't updated on the call.
- M&A is accelerating without disclosure: three acquisitions, zero prices or revenue contributions; software monetization remains narrative ("enabler," not a revenue stream) — Token Factory has logos but still no metrics.
- Community/political opposition is now significant enough to warrant a prepared response — siting/permitting is an emerging execution risk for the owned-site strategy.
Next-quarter watchlist- Q2 margin dip vs the stated path: does group/core EBITDA margin fall only modestly and recover to Q1 levels (32%/45%) in Q3 as promised? This is now the key credibility test.
- Microsoft H2 delivery cadence: tranche milestones, any Vineland/NJ site confirmation, and whether Q3/Q4 volume weighting holds.
- Funding the incremental $4–5B: first asset-backed facility against Meta/Microsoft contracts (size, pricing), corporate debt, and any ATM usage after the "evaluating change" language.
- Q1/Q2 CapEx spend disclosure against the $20–25B guide — still not given quarterly.
- Connected capacity: progress toward 800MW–1GW by year-end; status of Alabama and Missouri for Q1 2027; Pennsylvania groundbreaking milestones.
- Meta #2 mechanics in practice: first asset-backed financing against the $15B option; any evidence of capacity redirected to higher-priced AI cloud customers.
- Token Factory monetization: revenue, token-volume, or per-token pricing metrics; conversion of Revolut/monday.com-type wins into disclosed ARR.
- Eigen/Clarifai/Tavily integration: purchase prices, contribution, and whether inference revenue becomes a disclosed line.
- Concentration and backlog: any quantitative disclosure of Meta/Microsoft share of ARR or RPO — asked and dodged three calls running.
- Depreciation-change quantification and any further accounting-policy adjustments beneath the margin story.
- Avride/TripleTen: progress on strategic partners/deconsolidation; Toloka and ClickHouse monetization; community-opposition impact on any site timelines.
|
| Feb 12, 2026 | +1.26% | Q4 FY2025 | Read transcript briefingQuarter in one view- The Q4 cliff was cleared: core ARR hit $1.2B at December, exceeding the $1.1B high end of the exit guide — the ~$350M+ single-quarter ARR add that last quarter's setup demanded actually happened. Q4 group revenue $228M, +547% YoY, +56% QoQ; core AI cloud revenue +830% YoY, +63% QoQ.
- Profitability inflected: group adjusted EBITDA turned positive in Q4 (as guided), with core AI cloud adjusted EBITDA margin expanding 19% → 24%. Group Q4 adjusted EBITDA margin was only 7%, yet 2026 group margin is guided to ~40% — a very steep one-year ramp.
- Balance sheet finally disclosed: $3.0B cash at year-end; $834M Q4 operating cash flow, "primarily comprised of upfront payments from long-term agreements" — i.e., mega-deal prepayments are already funding the build.
- 2026 guidance issued: revenue $3.0B–$3.4B; ARR $7B–$9B reiterated; adjusted EBITDA margin ~40%; EBIT still loss-making; CapEx $16B–$20B (vs ~$5B in 2025) with ~60% of funding already in hand.
- Capacity targets raised again: >2GW contracted power already (February), 2026 target raised 2.5GW → >3GW; 800MW–1GW connected capacity by year-end reiterated; nine new data centers announced today.
- Mega-deal execution on track: Meta fully delivered (both tranches, early February, on time); Microsoft first tranche delivered in November, remainder through 2026, majority H2.
What management is focused on- Capacity acceleration as the whole game: nine new data centers announced on the call day; >2GW contracted already; "everything we build, we sell"; 2026 capacity "already sold out." The 3GW raise is framed as demand-driven, not speculative.
- Two-dimension strategy — scale and product: CEO explicitly frames capital allocation as split between data centers/GPUs (scale) and platform functionality (product), with M&A now an active lever ("hopefully not the last one").
- Tavily acquisition as the new product proof point: agentic search, ~700,000 developers, Fortune 500 customers — positioned as the first of ongoing tuck-ins to deepen developer stickiness.
- Financing architecture: COO Ophir Nadav (unusually) took the CapEx question personally — ~60% of 2026 CapEx funded from cash, operating cash flow, and prepayments from long-term contracts; the rest via a deliberate move from zero debt to corporate debt and asset-backed financing; ATM untouched and "no concrete plans" to use it.
- Demand quality narrative: average new-customer contract duration +50%; GPU prices flat even on prior generations; startups scaling from hundreds to tens of thousands of GPUs; Q1 pipeline creation tracking >$4B.
Key numbers and quarter mechanics- Revenue $228M (+547% YoY, +56% QoQ); core AI cloud +830% YoY, +63% QoQ. Non-core implied small but not broken out on the call.
- Core ARR $1.2B at December vs $551M at September — roughly $650M of Q4 adds, validating the "capacity timing, not demand" claim from Q3.
- Core adjusted EBITDA margin 24% (from 19%); group adjusted EBITDA margin 7% in Q4, guided to ~40% for full-year 2026 — the bridge is core mix growth plus shrinking relative drag from other businesses, which remain EBITDA-negative in 2026.
- Cash $3.0B; Q4 operating cash flow $834M, mostly customer prepayments — first cash disclosure in three quarters.
- 2026 guidance: revenue $3.0B–$3.4B; ARR $7B–$9B (reiterated, conviction "stronger"); adj. EBITDA margin ~40%; EBIT loss; medium-term EBIT margin target 20–30% maintained.
- CapEx $16B–$20B for 2026 — a 3–4x step-up from ~$5B in 2025. COO's split: <10% power securing, ~20% data center build, remainder GPUs deployed against visible demand.
- Accounting change: depreciation schedule extended 4 → 5 years starting Q1 2026, framed as reflecting market/utilization reality. This mechanically lifts EBITDA-adjacent optics and reduces depreciation — worth flagging as a non-operational margin tailwind.
- Meta mechanics: both tranches live early February → ~12 months revenue on tranche one, ~11 on tranche two in 2026. Microsoft: revenue ramps through 2026 (majority H2), full annual run rate only from 2027.
- Contracted power >2GW now, target >3GW by end-2026; connected capacity 800MW–1GW by year-end reiterated; nine new sites (mix of owned and colocation); colocations ramp from Q2, own large projects ramp 2027+.
- Still not disclosed: Q4/full-year CapEx spend, backlog/remaining contract value, utilization percentage, customer concentration, Microsoft/Meta contract terms.
Product and launch scorecard- Tavily (first M&A): agentic search connecting AI agents to the web; ~700,000 developers; claimed Fortune 500 customer base. Strategic logic (developer funnel, stickiness, LTV) is coherent; no price paid, revenue, or integration metrics disclosed — an acquisition rationale, not yet evidence.
- Token Factory: mentioned only in passing as proof of in-house capability. Still zero customer, revenue, or usage metrics — two consecutive calls without traction data.
- Software stack attach: CRO claims 100% of AI cloud customers use the software layer; new products named — Token Factory, Aether releases, Tavily, embedded storage for verticals (physical AI, media, healthcare). Monetization "early stages"; per-token pricing being explored. Attach rate is asserted, not quantified in ARR.
- Capacity delivery as product: Meta delivered fully and on time; Microsoft first tranche on time (November); New Jersey on schedule with "safety margin buffers." This is the strongest execution evidence in the quarter.
- Demand/pricing evidence: sold out Q3, Q4, and "already" 2026; Hoppers sold out with renewals moving to 12+ months at nudging-up prices; Q4 deals >12 months nearly doubled vs Q3; ASPs +50%+; prepayments for future capacity increasing. Specific and directional, though still no absolute price levels.
- Avride, Toloka, ClickHouse: absent again — second consecutive call of silence on the equity-stake portfolio (ClickHouse stake >25% and AV ride/Avride were mentioned only as future capital sources, with a claimed ~$15B ClickHouse valuation cited).
Sell-side read-through- Named questioners: Josh Baer (Morgan Stanley, twice), Alex Platt (D.A. Davidson, twice), Alex DeVal (Goldman Sachs, twice), Nehal Chokshi (Northland), James Kisner (WaterTower), Andrew Beal (ArrayTech — likely Arete garbled). Format still IR-curated via portal.
- Goldman's Q4-revenue-miss question was deflected, not answered: revenue "came in the middle of our guidance" — but the question noted it was light versus *consensus*. Management reframed to ARR as "north star" without addressing the consensus gap.
- The 40% EBITDA margin question got a mix-shift answer, not a bridge: from 7% in Q4 to 40% for 2026 with only "core margins are significantly higher and other businesses shrink as a share" — no quantified walk. This is the least-substantiated guidance on the call.
- Northland's capital-allocation question (build vs buy) got a philosophical answer; no M&A budget, return thresholds, or Tavily price.
- Power bridge question (800MW–1GW connected vs 3GW contracted) was answered with portfolio logic — colocations ramp from Q2, own large projects ramp 2027+ — implicitly conceding the 3GW is mostly a 2027+ revenue asset.
- Not asked or not pressed: depreciation-change impact on margins, customer concentration, Microsoft/Meta contract terms, Q4 CapEx spend, why revenue missed consensus, Avride/Toloka status, ATM vs debt sequencing beyond "no concrete plans."
Management credibility- The big promise was kept: last quarter's maintained $900M–$1.1B exit ARR guide required a heroic Q4; they printed $1.2B, above the high end. The "capacity timing, not demand" explanation is now validated by delivery.
- Meta delivery completed on time and Microsoft on schedule — the two execution obligations most at risk are tracking.
- Capacity over-delivery: 2.5GW-by-year-end target already exceeded in February (>2GW), prompting the 3GW raise. Pattern of raising targets after beating them is credibility-positive but also makes targets soft.
- New looseness to note: the 4→5 year depreciation change coincides with a 40% EBITDA margin guide — management calls it conservative and market-aligned, but it flatters the margin trajectory and was not quantified. The 7% → 40% margin bridge was asserted, not built.
- Transcript/sloppiness persists: speaker attributions and names are garbled throughout (CFO introduced as "Dado Alonso," operator voicing CFO lines, "AV Wright"/Avride, "ArrayTech"/Arete); the CEO's demand anecdotes include garbled customer names. Substance was strong this quarter; presentation hygiene was not.
- Financing posture is now consistent: unlike last quarter's surprise ATM, this call gave a coherent funding stack (~60% internal/prepayments, debt to come, ATM explicitly deprioritized) — a credibility repair versus Q3.
What changed versus the prior quarter- The Q4 ARR bet resolved positively: $1.2B vs $551M — the single biggest open risk from Q3 closed.
- Guidance regime shifted from ARR-only to full P&L: first full-year revenue guide ($3.0–3.4B), first group EBITDA margin guide (~40%), explicit EBIT-loss acknowledgment.
- CapEx stepped up again: ~$5B (2025) → $16–20B (2026), with a named funding structure replacing last quarter's reactive ATM.
- Cash disclosed for the first time in three quarters ($3.0B), plus $834M Q4 operating cash flow revealing mega-deal prepayment mechanics.
- M&A entered the toolkit: Tavily is the first acquisition; CEO signals more.
- Depreciation policy changed (4→5 years) — new this quarter, unquantified.
- ATM posture softened: from "filed the day after the call" to "not used, no concrete plans."
- Contracted power target raised mid-cycle again: 2.5GW → >3GW.
- Continued silences: Avride operations, Toloka, Token Factory metrics, concentration, utilization.
Bull case- Guide beat on the metric that mattered: $1.2B ARR vs $1.1B top end, with ~$650M of Q4 adds — the delivery machine works when capacity lands, and capacity is now landing on schedule (Meta done, Microsoft on track).
- 2026 is unusually de-risked for the growth rate: $3.0–3.4B revenue guide against $7–9B exit ARR, sold-out 2026 capacity, >$4B Q1 pipeline trajectory, lengthening contracts (+50% duration), rising ASPs, and prepaying customers.
- Self-funding flywheel is real: $834M of Q4 operating cash flow from customer prepayments plus $3B cash covers ~60% of a $16–20B CapEx year — the Microsoft/Meta contracts are literally financing the build, as promised.
- Margin evidence improved: core EBITDA margin 24% and rising; group positive; the 40% 2026 guide, while aggressive, sits on a disclosed core margin already above half that level.
- Power position is a moat in progress: >2GW contracted, >3GW targeted, nine new sites, <10% of CapEx securing power — the scarce input is being locked up cheaply ahead of 2027 demand.
- Pricing disproves the obsolescence bear case: prior-generation GPUs holding price, Hoppers renewing at longer terms and higher prices.
Bear case- The 40% EBITDA margin guide is the new heroic number: from 7% in Q4 to 40% for 2026 with no quantified bridge, in the same quarter depreciation was lengthened from 4 to 5 years — part of the "improvement" is accounting, and the mix-shift math was asserted, not shown.
- CapEx tripled-to-quadrupled again ($16–20B) one quarter after the ~$5B raise; ~40% of funding is not yet in hand and depends on debt markets the company has never tapped. Execution or credit-condition slippage directly threatens the plan.
- Revenue missed consensus (per Goldman's question) even in a beat-everything quarter — the revenue guide's "prudence" framing suggests H2-weighted delivery risk is being priced into guidance, and any Microsoft H2 slippage pushes revenue into 2027.
- Concentration is now the business: Meta fully delivered, Microsoft ramping, prepayments funding CapEx — counterparty health and contract performance of two customers drive the model, and no concentration, backlog, or contract-term disclosure was offered.
- 3GW contracted vs 800MW–1GW connected means most of the announced power is a 2027+ story; the gap between contracted headlines and connected reality is where expectations can overshoot.
- Product layer remains assertion: Token Factory has no metrics after two quarters; Tavily's price and contribution are undisclosed; "100% attach" is a definition, not monetization.
- Silent portfolio: Avride/Toloka/ClickHouse appear only as future piggy banks — if core execution falters, the fallback is selling stakes, not operating results.
Next-quarter watchlist- Q1 2026 revenue and ARR trajectory vs the $3.0–3.4B / $7–9B path; whether ARR adds stay at the Q4 run-rate or revert toward capacity-timing lumpiness.
- The 40% EBITDA margin bridge: quarterly margin progression from 7%, core vs group split, and explicit quantification of the depreciation-change effect (4→5 years) on reported margins.
- CapEx funding execution: first corporate debt or asset-backed facility (size, pricing, collateral), any ATM usage despite "no concrete plans," and quarterly CapEx spend disclosure against the $16–20B guide.
- Microsoft delivery cadence: tranche milestones through 2026, H2 weighting, any slippage language; Meta revenue recognition (~12/11 months) flowing as expected.
- Connected-capacity build: progress toward 800MW–1GW; colocation ramps from Q2; status of the nine new sites and the 3GW contracted target.
- Tavily integration: purchase price, revenue contribution, developer-funnel conversion into cloud ARR; any second acquisition.
- Token Factory / Aether / software monetization: first hard metrics (customers, ARR, per-token pricing launch) after two quarters of narrative.
- Concentration and backlog disclosure: Microsoft/Meta share of ARR and contracted revenue; remaining performance obligations.
- Pricing and contract-duration trends: whether +50% duration and rising ASPs persist as new capacity floods in during H2.
- Silent items: Avride operations, Toloka, ClickHouse monetization, and any further accounting-policy changes.
|
| Nov 11, 2025 | -7.03% | Q3 FY2025 | Read transcript briefingQuarter in one view- The story changed category: from "supply-capped neocloud ramping ARR" to "contracted hyperscaler supplier." Two mega deals now anchor the model: Microsoft ($17.4B–$19.4B, announced early September) and Meta (~$3B over five years, announced today). Management says Meta's size was capped by available capacity — "if we had more, we could have sold more."
- Q3 group revenue $146M, +355% YoY, +39% QoQ; core infrastructure ~90% of revenue, +400% YoY, +40% QoQ. September ARR $551M vs $430M in June — only ~$12M incremental ARR in the quarter (per Northland's question; down from ~$180M in Q2 and $159M in Q1), which management attributes entirely to capacity timing and promises reverses sharply in Q4.
- Core adjusted EBITDA margin expanded to ~19% (first margin disclosure of this kind). Group adjusted EBITDA still guided slightly positive by year-end, negative for the full year.
- Guidance reshaped: FY2025 group revenue tightened to $500M–$550M (from $450M–$630M), pacing to midpoint; ARR exit guide of $900M–$1.1B maintained; CapEx raised ~$2B → ~$5B; new 2026 ARR target of $7B–$9B, of which "more than half is already booked."
- Capacity targets escalated again: 2.5GW contracted power by 2026 (up from 1GW discussed in August) and 800MW–1GW of connected (fully built) power by end-2026 (vs 220MW connected by end-2025 previously).
- Financing posture flipped from "opportunistic" to active: three named sources — corporate debt, asset-backed financing (in process, "supported by the creditworthiness of our largest customers"), and equity via a new ATM program for up to 25M Class A shares, prospectus supplement to be filed November 12.
What management is focused on- Capacity as the sole bottleneck, and removing it: every demand answer routes back to "we bring capacity online, we sell all of it." The 2.5GW contracted / 800MW–1GW connected plan is the centerpiece; securing "several new large sites" expected to add "hundreds of megawatts."
- Mega deals as financing engine for the core cloud: CEO explicitly frames Microsoft/Meta economics as funding faster buildout of the core AI cloud — "this is our real future opportunity." Allocation across customer categories will be based on "individual economics of the deals."
- Staged CapEx discipline as the risk-management narrative: CEO laid out a three-stage model — land/power ~1% of CapEx, building connected power ~18–20%, GPUs ~80% — with the 80% GPU spend committed "only when we see real demand." This is the answer to both the bubble question and the oversupply question.
- Enterprise readiness: platform v3.0 "Aether" (transcript renders it "Ether") with compliance/security certifications, IAM, admin dashboards; enterprise sales leadership hires; verticals (healthcare/life sciences, physical AI, media) named as progress areas.
- Inference as the "next wave": Token Factory (transcript garbles it once as "Nebius Talking Factory") positioned for vertical AI builders, ISVs, enterprises — guaranteed performance, transparent cost per token, 99.9% uptime, open-weight models (OpenAI OSS, Qwen, DeepSeek, etc.).
Key numbers and quarter mechanics- Revenue $146M (+355% YoY, +39% QoQ); core ~90% of total. Implies non-core (Tripleten etc.) ~$15M.
- ARR $551M at end-September vs $430M June. The ~$12M net-adds figure came from the analyst question and was not disputed — a dramatic sequential slowdown that management says is purely capacity-timing and will be "significantly higher" in Q4. This is the single most important mechanical fact of the quarter: the $900M–$1.1B exit guide now requires ~$350M–$550M of ARR adds in Q4 alone, roughly 3–5x the best quarter to date.
- Core adjusted EBITDA margin ~19% in Q3, expanding QoQ (prior quarter disclosed only "positive").
- FY2025 revenue guide $500M–$550M, pacing to midpoint — the narrowing is attributed to "exact timing of when capacity comes online," i.e., a soft acknowledgment that the top half of the old range is out of reach.
- CapEx guide ~$5B for 2025 (from ~$2B) — a 2.5x raise in one quarter, covering hardware, power, land, and sites.
- Mega-deal revenue mechanics: Microsoft — first tranche delivered (New Jersey), no material 2025 revenue/ARR effect; remaining tranches in 2026, more than half in H2 2026; full annual run rate from 2027. Meta — deployments conclude within ~3 months; mostly full run rate during 2026.
- 2026 ARR target $7B–$9B, "more than half already booked" (CRO). Full-year 2026 revenue guidance promised next quarter.
- Pipeline: $4B generated in Q3, +70% QoQ (CRO) — new disclosure, but "pipeline" is undefined (no conversion-rate or coverage math given).
- GPU deployment lead time: 6–12 weeks from connected power to revenue (shorter at existing sites) — useful for modeling the Q4 ramp.
- Still not disclosed on the call: cash balance, Q3 CapEx spend, backlog/contract value beyond the two named deals, utilization, churn, concentration.
Product and launch scorecard- Blackwell: Israel live with B200s, UK live with B300s (capacity "coming online in the next week or so"), both presold before launch. GB300 launching in Finland in December — claimed first in Europe. Q4 capacity "nearly sold out"; selling "remnants of Q4" and pre-selling future quarters. This is the strongest launch evidence management has produced: presold, dated, and geographically specific.
- Hopper durability: renewals and Blackwell-upgrade paths both "typically selling immediately, often at better pricing than previously" — a direct, quantified-in-direction answer to the obsolescence concern, though still no price levels.
- Aether (Nebius 3.0): enterprise compliance/security certifications delivered "in a matter of months," IAM and admin tooling. No adoption or revenue metrics — a readiness claim, not a traction claim.
- Token Factory (inference): launched; Roman Chernin owns it as promised last quarter. Feature-level detail (dedicated endpoints, guaranteed latency, 99.9% uptime, cost-per-token transparency) but zero customer, revenue, or usage metrics — still pre-evidence.
- New AI-native logos: Cursor, Black Forest Labs, World Labs added in Q3; Shopify expanded. (CEO's "Courser" is presumably Cursor — transcript inconsistency.) These are credible, named, high-growth customers.
- New Jersey: first tranche handed over to Microsoft; expansion continuing "as planned." UK capacity "close to doubled" since June announcement; peak capacity at current UK facility by January; Israel fully live and presold, with government subsidies stimulating local demand.
- Avride, Toloka, ClickHouse: not mentioned at all on this call — a notable silence given last quarter's extensive stake-optionality narrative.
Sell-side read-through- Named analysts: Alex Platt (D.A. Davidson), Alex Duval (Goldman Sachs), Nehal Chokshi (Northland), Andrew Beal (Arete). Format remains curated via IR, with online questions interleaved.
- The hardest question of the call was Northland's: incremental ARR of ~$12M in Q3 vs ~$180M in Q2 — and the answer ("capacity bottleneck; Q4 will be significantly higher") confirms the entire 2025 guide is now a single-quarter bet, with no quantification of what "significantly higher" means.
- Goldman pressed the 2.5GW → revenue math ("fair to assume >$20B of revenue?"). Andrey answered "I guess it's fair to assume" then immediately walked it back to capital constraints and demand pacing — management let a very large number hang in the air without endorsing a timeline.
- Dilution question was asked directly ("you just completed a secondary, why an ATM?"). The answer — "more tools at our disposal," "dilution sensitive" — did not address why equity is needed now alongside claimed attractive asset-backed debt terms. No sizing, pricing, or use-of-proceeds detail.
- Greenfield specifics evaded: Arete asked for LOIs/site detail; answer was "robust pipeline... not in a position to say more." Last quarter's "two US greenfields near closing" were not confirmed as closed.
- Unasked/unpressed: cash balance, Q3 CapEx spend, customer concentration (Microsoft + Meta now dominate contracted revenue), Microsoft deal structure/term length, Avride/Toloka/ClickHouse status, depreciation policy, Dutch tax.
Management credibility- Delivered on the mega-deal promise: last quarter the CEO said large deals would be "incremental" and unguided; this quarter Microsoft ($17.4–19.4B) and Meta (~$3B) are signed, and "more than half" of the new $7–9B 2026 ARR target is booked. That is a major promise kept, ahead of any stated timeline.
- But the near-term numbers wobbled: ~$12M Q3 ARR adds is a sharp miss versus the trajectory implied in August, and the revenue guide was effectively trimmed at the top end ("pacing to midpoint"). Management's framing — timing, not demand — is plausible given presold capacity, but it means the Q4 print must be extraordinary (~$350M+ of ARR adds) to validate the maintained exit guide.
- CapEx credibility cut both ways: raising from ~$2B to ~$5B in one quarter signals conviction and contracted demand, but it also means the prior capital plan was obsolete within 90 days — and the funding plan (ATM + debt + asset-backed) was assembled after the fact.
- Consistency markers: the three-stage CapEx framework is coherent and was used consistently to answer bubble, oversupply, and CapEx-philosophy questions. Margin focus was repeated by both CEO and CRO. The 19% core EBITDA margin disclosure adds substance.
- Sloppiness persists: transcript shows product-name confusion (Ether/Aether, Talking Factory/Token Factory, "Courser"/Cursor, "Vineland"), and the CRO's "more than half already booked" is unquantified against the $7–9B range. The "I guess it's fair to assume" $20B+ answer was loose for a number that large.
- New CFO now owns guidance and capital structure; the ATM announcement one day after the call is decisive but reverses last quarter's "no immediate need" posture without explanation.
What changed versus the prior quarter- Mega deals went from unguided upside to the core of the story: Microsoft and Meta together represent ~$20B+ of contracted value; last quarter these were explicitly excluded from all projections.
- 2026 visibility introduced: $7–9B ARR target (>half booked) and a promise of full-year 2026 revenue guidance next quarter — the first multi-year financial target of this scale.
- Capacity targets roughly doubled: 1GW → 2.5GW contracted by 2026; connected power target of 800MW–1GW by end-2026 is new (prior: 220MW connected by end-2025).
- CapEx raised 2.5x (~$2B → ~$5B) after being held flat last quarter.
- Financing posture reversed: from "significant cash on hand, opportunistic raises, no immediate need" to an active three-channel funding plan including a 25M-share ATM filed the day after the call.
- ARR momentum decelerated sharply (~$12M Q3 adds vs ~$181M in Q2) even as the exit guide was maintained — the divergence between guide and run-rate is now extreme.
- Margin disclosure added (~19% core EBITDA margin); revenue guide tightened with a midpoint bias.
- Silences: no mention of Avride, Toloka, ClickHouse, or the inference platform's prior "pipeline" claims; no update on the two US greenfields "near closing"; no cash balance for the second consecutive call.
Bull case- Contracted revenue base transformed: ~$20B+ across Microsoft and Meta, with Meta explicitly capacity-capped ("if we had more, we could have sold more") — demand at the largest scale is proven, not projected. More than half of the $7–9B 2026 ARR target is already booked.
- Unit economics are showing: ~19% core EBITDA margin at $146M revenue, expanding QoQ, alongside the prior 2–3 year Hopper payback disclosure — the model profits at small scale before the gigawatt buildout.
- Everything new is presold: UK, Israel, Finland Q4 phases all sold before launch; Q4 "nearly sold out"; Hoppers renewing at equal or better pricing. The Q4 ARR cliff is a delivery problem, not a sales problem.
- Financing is now credible in structure if not in detail: asset-backed debt secured by "creditworthiness of our largest customers" (i.e., Microsoft/Meta contracts as collateral) is exactly how this asset class should be financed, and the staged CapEx model (1% land/power, ~20% build, 80% GPUs against visible demand) limits downside if demand pauses.
- Demand indicators are extreme: $4B pipeline in Q3, +70% QoQ; CRO "learning to say no to customers."
- 2026 guide is coming: full-year revenue guidance next quarter against >50% booked ARR gives an unusually de-risked setup for a company this early.
Bear case- The 2025 guide requires a heroic Q4: ~$350M–$550M of ARR adds in one quarter versus ~$12M in Q3 and a best-ever ~$181M. Any slip in New Jersey expansion, Finland GB300 launch (December), or UK ramp breaks the guide with zero buffer — and management just demonstrated (Q3 adds, midpoint pacing) that timing slips are already happening.
- Dilution is now concrete: a 25M-share ATM announced the day after the call, on top of a recent secondary, with "dilution sensitive" as the only guardrail. CapEx tripled to ~$5B with no disclosed cash balance — the funding gap is being filled reactively.
- Concentration risk inverted the thesis: Microsoft and Meta now dominate contracted value; the "diverse core AI cloud" is the stated priority but the booked 2026 ARR is majority mega-deal. Counterparty, renegotiation, and delivery-obligation risk (New Jersey tranches, H2-2026-weighted Microsoft deliveries) are all unquantified.
- Execution surface area exploded: 2.5GW contracted, 800MW–1GW connected, multiple new countries, greenfields, and a December GB300 launch — all simultaneously, with supply-chain and permitting constraints acknowledged. The 6–12 week power-to-revenue lag means H2-2026 Microsoft tranches leave little slack.
- Disclosure gaps widened at the worst time: no cash balance, no Q3 CapEx spend, no utilization/churn/concentration data, no quantification of "more than half booked," and total silence on the equity stakes that were last quarter's funding backstop.
- Management let a "$20B+ revenue from 2.5GW" framing pass with "I guess it's fair to assume" — if that number circulates as an implicit target, it sets an expectation the staged-CapEx model explicitly does not commit to.
Next-quarter watchlist- Q4 ARR print vs the $900M–$1.1B guide: the single decisive number. Requires ~$350M+ of adds; check whether "significantly higher" Q4 adds materialize and whether exit ARR is reported cleanly.
- December GB300 Finland launch (claimed first in Europe) and UK/Israel/New Jersey ramp status; any slippage language.
- 2026 full-year revenue guidance (promised) and quantification of the "more than half booked" portion of the $7–9B ARR target; Microsoft delivery schedule (H2-2026-weighted tranches) and Meta deployment completion within ~3 months.
- Financing execution: ATM usage (pace, price, dilution), asset-backed facility terms and size, corporate debt; first cash-balance and CapEx-spend disclosure in two quarters.
- 2.5GW contracted-power evidence: site announcements, LOIs, the previously "near closing" US greenfields; progress toward 800MW–1GW connected.
- Concentration and contract disclosure: Microsoft/Meta as % of ARR/backlog, contract terms, any performance-obligation detail on New Jersey.
- Token Factory and Aether traction: first customers, revenue contribution, enterprise pipeline conversion — currently assertion-only.
- Margin trajectory: whether ~19% core EBITDA margin holds as mega-deal revenue (potentially lower-margin) begins to mix in during 2026.
- Silent items: Avride (Dallas robotaxi was "later this year"), Toloka/ClickHouse monetization, depreciation policy, Dutch tax — any reappearance or continued absence.
|
| Aug 7, 2025 | +18.55% | Q2 FY2025 | Read transcript briefingQuarter in one view- First absolute revenue print on a call: $105.1M in Q2, +625% YoY, +106% QoQ, driven by core AI cloud (revenue >9x YoY) plus Tripleten. Core AI infrastructure hit positive adjusted EBITDA in Q2 — one quarter ahead of the prior "probably sometime in Q3" marker. Group GAAP net income was positive, but only via two items management explicitly called one-time: an equity-securities revaluation gain and a discontinued-operations gain.
- ARR guidance raised from $750M–$1B to $900M–$1.1B for year-end 2025, while FY revenue guidance was held — an explicitly back-end-loaded ramp, with most GPU installations landing in Q4. ARR disclosed at $249M in March → $430M in June, with July momentum continuing (no monthly updates going forward).
- Capacity story escalated: 220MW of connected power (active or GPU-ready) by end-2025, of which 100MW active; >1GW of power targeted by end-2026; two "substantial" US greenfield sites near closing. New Jersey has 200MW under construction; Finland adding 50MW this year; UK and Israel GPU clusters coming "early Q4" with claimed first B300 delivery in the UK.
- New leadership on the call: Dado Alonso (new CFO) and Marc Boroditsky (new CRO); Roman Chernin moved off go-to-market to new initiatives including a vertically integrated inference-as-a-service product.
- Note a discrepancy: the CEO said prior ARR guidance was "$700 million to $1 billion"; the prior-quarter record shows $750M–$1B. Likely a verbal slip, but worth verifying against the Q1 release.
What management is focused on- Capacity as the binding constraint, now removed as the excuse: CEO framed Q2 as "we could grow faster, but we were oversold on all of our supply of previous generation hoppers" — growth is now explicitly a supply story, and the entire call architecture (220MW, 1GW, greenfields, Q4 GPU installs) is built to argue the constraint is being eliminated.
- Enterprise traction as the new proof point: named Cloudflare, Shopify, Prosus as new large global tech customers (plus AI-natives HeyGen, Lightning.AI, Photoroom). Shopify uses Nebius infrastructure plus Toloka training data; Cloudflare uses Nebius for inference at the edge via Workers AI. This is the first quarter with named enterprise logos of this scale.
- Frontier-lab/hyperscaler deals as unguided upside: asked directly what it takes to win a large multiyear deal, the CEO answered "we are increasing our capacity" and reiterated that all current and midterm projections exclude such deals — "if or when they will come, it will all be incremental."
- Financing via stake optionality, not urgency: CFO said "significant cash on hand" and capital raising will be "opportunistic"; Tom Blackwell walked through Toloka (Scale AI comp at ~$30B), ClickHouse (reported ~$6B round, "potentially... several billion dollars" at a future liquidity event), and Avride (Waymo comp at $40–50B, citing a D.A. Davidson note). No raise announced; no cash balance disclosed on the call.
- Greenfield-first build strategy: ~20% lower TCO than market average claimed, control over design/phasing, no long-term lease lock-in — a deliberate differentiation from build-to-suit/colocation peers.
Key numbers and quarter mechanics- Revenue: $105.1M, +625% YoY, +106% QoQ. Implies Q1 revenue of roughly $51M (my arithmetic from the QoQ figure) — the first time an absolute base can be derived.
- ARR: $249M (March) → $430M (June); ~$181M net adds in the quarter. Note this restates the prior quarter's "April $310M" data point onto a cleaner March baseline of $249M — the two series are consistent only if April added ~$61M; the prior summary's $220M March marker is superseded by the $249M figure.
- Guidance: ARR exit $900M–$1.1B (raised); core business revenue $400M–$600M (maintained); group revenue $450M–$630M, now ex-Toloka (Toloka's $50M–$70M removed); FY adjusted EBITDA negative, group positive "by the end of the year" and positive for full-year 2026 (new); CapEx ~$2B maintained.
- Capacity: 220MW connected by YE2025 (100MW active); >1GW secured by end-2026; New Jersey 200MW under construction (part 2025, rest H1 2026); Finland +50MW this year; UK and Israel clusters early Q4.
- Unit economics (new disclosure): Hopper GPUs expected to break even in ~2–3 years on a gross-profit basis including hardware and opex, excluding higher-margin software/services upside; Blackwells priced "at a premium," specifics withheld.
- GAAP profitability mechanics: equity revaluation gain + discontinued-operations gain drove net income; management flagged both as one-time. No cash balance, no Q2 CapEx spend, no backlog/contract-value figures given on the call (shareholder letter referenced but not included).
- CFO slip: "1 gigabyte target" when meaning 1 gigawatt — trivial, but the second consecutive call with a units stumble from an executive.
Product and launch scorecard- Blackwell: B200s deployed and "actively selling through"; Grace Blackwell (GB300/Blackwell Ultra) interest cited with implementation "later this year"; UK B300 cluster claimed as first-to-market in that geography, early Q4. The raised ARR guide is explicitly predicated on Blackwell Ultra delivery in Q4 — still no quantified presale values, though management now says "a significant portion" of the ARR guide is "already under contract" (unquantified).
- Hopper economics: sold out at "peak utilization"; pricing "relatively stable" even against Blackwell alternatives — directly addresses the obsolescence concern; 2–3 year gross-profit payback disclosed for the first time.
- Inference-as-a-service: new enterprise-grade, vertically integrated platform for open-weight models (Llama, Qwen, Flux, OpenAI's new open models), targeting latency/GPU-bottleneck pain points; Roman Chernin dedicated to it. Early-stage — no revenue or customer metrics.
- Platform performance: network speed doubled; MLPerf Training v5.0 results submitted for Llama 3.1 405B claiming linear scaling and bare-metal-comparable performance in cloud; auto-healing/health-check software improved MTBF. MLPerf is the most externally verifiable technical claim management has made.
- Partnerships: Mistral, Baseten, SkyPilot, Lightning AI, Anyscale integrations; NVIDIA AI Enterprise added; launch NCP partner for DGX Cloud Lepton — Lepton described as generating "quite a significant pipeline" that converts to direct relationships (qualitative).
- Avride: H-E-B grocery delivery (Texas), Mitsui Fudosan indoor robots (Japan), Grubhub campus expansion, Uber robotaxi launch in Dallas "later this year"; Hyundai road tests expanding. Toloka-structure deal (cede control, retain economics) confirmed as the target model. Still no revenue figures.
- Toloka: deconsolidated effective Q2; Bezos Expeditions-led round closed; majority economic interest retained; Scale AI's ~$30B valuation cited as comp.
Sell-side read-through- Format remains curated (IR reads consolidated questions), but named analysts are now attributed: Alex Duval (Goldman Sachs), Nehal Chokshi (Northland), Alex Platt (D.A. Davidson), Andrew Beale (Arete) — marginally more transparent than Q1.
- Repeated/pressing themes: funding (asked in two forms again), the ARR-up/revenue-flat guidance divergence, and the path to hyperscaler-scale deals (Arete). Management's answers were consistent but non-committal on timing and size.
- Notable evasion: the direct utilization question ("trends in the quarter or by GPU family") was answered with a forward-selling narrative ("shifting to selling against future requirements") — no utilization number given beyond the earlier "peak utilization" claim. No churn, concentration, contract-length, or cash-balance disclosure.
- The guidance-divergence question (ARR raised, revenue flat) was answered cleanly with a Q4-weighted capacity timing argument — coherent, but it concentrates execution risk into a single quarter.
- Unasked/unpressed: Q2 cash balance and CapEx spend, depreciation policy (the Q1 transcript garble), Dutch tax status, Toloka transaction economics, customer concentration.
Management credibility- Delivered ahead of schedule: core-business positive adjusted EBITDA in Q2 vs "probably Q3" guidance — the first time management has beaten its own profitability marker. ARR of $430M in June is consistent with the trajectory implied by prior disclosures.
- Guidance raised on ARR but held on revenue and CapEx — a disciplined pattern (raise only what demand supports, hold what timing governs) rather than across-the-board inflation.
- Candor markers: one-time nature of GAAP profit items explicitly flagged; "we could grow faster, but we were oversold" admits supply-capped Q2; Blackwell pricing "still early to comment"; tariffs answer remains a hedge ("a bit early to say anything definitive").
- Credibility risks: the funding answer is now softer than Q1's multi-source narrative — "significant cash on hand" and "opportunistic" raises, with no cash figure disclosed and stake monetizations explicitly pushed to "no immediate need"; the 1GW-by-2026 target is a large new commitment with no capital plan attached; CEO's "$700M" misquote of prior guidance and CFO's "gigabyte" slip are minor but sloppy on the numbers that matter most.
- New CFO and CRO are both weeks/months in — guidance ownership is transitioning; watch whether the new CFO maintains the prior team's guidance cadence and disclosure granularity (no monthly ARR updates is a slight disclosure tightening).
What changed versus the prior quarter- Disclosure stepped up: absolute revenue ($105.1M) given for the first time on a call; ARR series restated on a quarterly basis ($249M March, $430M June) replacing the ad-hoc monthly markers; Hopper payback economics (2–3 years gross profit) disclosed for the first time.
- Profitability inflected early: core adjusted EBITDA positive in Q2 vs Q3 guide; group positive now promised for full-year 2026 — new.
- Guidance mix shifted: ARR exit raised $750M–$1B → $900M–$1.1B; group revenue restated ex-Toloka ($450M–$630M); CapEx held at ~$2B after last quarter's raise.
- Capacity targets escalated: from ">100MW operational by end-2025" to 220MW connected/100MW active, plus a firm >1GW-by-end-2026 target and two US greenfields near closing; UK added as a new market (London-area cluster, early Q4, B300s).
- Customer mix evidence upgraded: first named enterprise-scale logos (Cloudflare, Shopify, Prosus) with use cases described; Q1's enterprise talk was aspirational.
- Funding posture softened: Q1's explicit "monetize stakes + capital markets" narrative became "no immediate need," with stake sales framed as multi-year optionality at comp valuations (Scale AI ~$30B, Waymo $40–50B) rather than near-term funding events.
- Org changes: new CFO (Dado Alonso), new CRO (Marc Boroditsky), Roman Chernin to inference/new initiatives; VP of Sales Strategy & Operations and regional GMs being hired — go-to-market is being rebuilt for enterprise.
- Prior-quarter open items not addressed: depreciation-policy wording, Dutch tax, Toloka deal economics, Blackwell presale values.
Bull case- The ramp is now contracted, not just projected: $430M June ARR to a $900M–$1.1B December exit requires ~2.1–2.6x in six months, and management states a "significant portion" is already under contract — versus last quarter when the guide rested on unquantified Blackwell presales.
- Core EBITDA positive a quarter early at ~$105M revenue scale suggests the unit economics work at much smaller scale than the 1GW ambition implies; Hopper 2–3 year gross-profit payback plus stable Hopper pricing alongside premium Blackwells undercuts the rapid-obsolescence bear argument.
- Demand is supply-capped, not demand-capped: sold-out Hoppers, "peak utilization," "if we had more capacity we probably would have sold more" — and 220MW/1GW directly attacks that constraint, with greenfield TCO claimed ~20% below market.
- Enterprise validation arrived: Cloudflare (edge inference via Workers AI), Shopify (with Toloka data), Prosus — category leaders with scaling relationships, plus a stated pipeline of "other major technology companies."
- Frontier-lab/hyperscaler deals are explicitly excluded from all guidance — any signing is pure upside to both the ARR guide and the midterm "several billion" target.
- Balance-sheet optionality remains real: $4B+ raised to date, no debt mentioned, and three externally validated stakes (Toloka/Bezos round, ClickHouse at reported ~$6B, Avride with Uber/Hyundai) as contingent capital.
Bear case- The entire guide now hinges on Q4: most GPU installations land in Q4, revenue and ARR are "back-end weighted," and the raised ARR number depends on Blackwell Ultra delivery and sell-through in a single quarter — any supply, commissioning, or deployment slip pushes the miss into year-end with no buffer.
- Funding opacity increased: no cash balance disclosed on the call, ~$2B CapEx maintained, 1GW-by-2026 newly committed, and the funding answer regressed to "opportunistic" raises plus stake sales explicitly deferred ("no immediate need," "coming years"). The gap between ambition and disclosed liquidity is wider than the narrative suggests.
- Utilization and durability still unproven: the direct utilization question was deflected; no churn, concentration, or contract-length data; "peak utilization" is asserted, not measured. The Q4 pattern of large customers rolling off has not been shown to be resolved.
- GAAP profitability is cosmetic this quarter (one-time revaluation and discontinued-operations gains) — group adjusted EBITDA remains negative for the full year.
- Guidance comparability is muddy: group revenue is now ex-Toloka, the CEO misquoted the prior ARR range, and the ARR series baseline shifted (March $249M vs the prior "at least $220M") — reconstructing a clean growth series requires the shareholder letter.
- Greenfield strategy concentrates execution and capital risk (construction, power delivery, phasing) versus asset-light colocation, and the ~20% TCO claim is unaudited management assertion.
Next-quarter watchlist- Q4-weighted delivery: evidence that GPU installations are landing on schedule — Blackwell Ultra/GB300 deployment status, New Jersey construction progress against the 200MW plan, Finland +50MW, UK and Israel clusters live "early Q4."
- ARR trajectory: Q3 ARR print vs the $430M June base; the gap to $900M–$1.1B must narrow to ~2x or less by Q3 for the guide to be credible; whether "significant portion under contract" gets quantified.
- Cash and CapEx: Q2/Q3 cash balance and CapEx spend from the press release (absent from this call); any equity/debt raise; progress or silence on ClickHouse/Toloka/Avride monetization.
- Group EBITDA crossover: positive group adjusted EBITDA "by the end of the year" and the new full-year-2026 positive commitment — margin trajectory in Q3.
- Frontier-lab/hyperscaler signal: any large multiyear deal announcement — explicitly unguided, so any signing is a material positive surprise; watch for capacity pre-commitments.
- Greenfield announcements: the two US sites "nearly closed" — size, power, capital, timeline; path to 1GW secured by end-2026.
- Inference-as-a-service: first customers, pricing, and whether it begins to show the software-margin contribution management keeps promising.
- Avride: Dallas robotaxi launch with Uber "later this year"; any Toloka-style strategic investment.
- Disclosure clean-up: depreciation policy wording, Dutch tax status, Toloka transaction economics, and reconciliation of the CEO's "$700M" prior-guide misquote and the restated March ARR baseline.
|
| May 20, 2025 | +4.21% | Q1 FY2025 | Read transcript briefingQuarter in one view- Strong Q1 print versus the prior quarter's soft Q4: revenue grew "nearly 400%" year-over-year and annualized run-rate revenue "nearly 700%," with April ARR disclosed at $310M — comfortably above the prior call's "at least $220M" March marker and signaling ~41% ARR growth in a single month (my inference from management's two data points). No absolute Q1 revenue dollar figure was given on the call.
- All guidance reiterated: FY2025 revenue $500M–$700M, December 2025 exit ARR $750M–$1B, full-year adjusted EBITDA negative. EBITDA timing was refined: core infrastructure business positive adjusted EBITDA "probably sometime in the third quarter," group positive "at some point in the second half."
- CapEx guidance raised to ~$2B for 2025 from $1.5B — attributed to Q4 CapEx slipping into Q1 plus opportunistic expansion (Israel). $544M already spent in Q1; cash fell to $1.44B from $2.4B at year-end (a ~$1B sequential decline including the $180M Dutch tax payment flagged last quarter — management did not break out the components on this call).
- New midterm framing introduced: "mid-single-digit billions" of revenue in "a few years," 20%–30% EBIT margins (potentially "well north of 30%" longer term), and a capacity pipeline ambition beyond 1 gigawatt.
- Structural portfolio change: Toloka to be deconsolidated following Bezos Expeditions and Mikhail Parakhin's investment (voting stake below 50%; majority economic interest retained); Q2 reporting and guidance will be restated ex-Toloka. Avride "active discussions" with strategic investors confirmed.
What management is focused on- Four-segment demand roadmap articulated by the CEO: (1) AI-native startups (current revenue base, mostly US); (2) frontier AI labs — explicitly "not in our revenue yet," requiring much larger data centers; (3) enterprises — called the most promising, underpenetrated sector; (4) national/sovereign AI projects — Israel framed as "first but probably not the last" national AI factory.
- Funding narrative shifted center stage: with cash ($1.44B) now below 2025 CapEx plans (~$2B), management built an explicit financing story around monetizing non-core stakes — ClickHouse (28%, citing press reports of a ~$6B fundraise), Toloka, Avride — plus "classical" capital markets, while pledging minimal dilution and low debt. No debt currently.
- US and global capacity build-out: New Jersey build-to-suit (first capacity late summer), Kansas City (Blackwells deploying for Q2), Iceland (live), Finland (phase one late Q3, phase two near year-end), Israel (new, revenue contribution mostly 2026). >100MW operational by end-2025 reiterated.
- Technology differentiation: ~50 products shipped in Q1 across AI cloud and AI Studio; NVIDIA relationship deepened (Blackwell Ultra AI Factory first-wave partner, Dynamo launch partner, 1 of 5 NVIDIA Cloud Partner reference platforms, DGX Cloud Lepton marketplace support at launch).
Key numbers and quarter mechanics- Revenue: "nearly 400%" YoY growth; ARR "nearly 700%" YoY. No absolute Q1 revenue disclosed on the call (press release not included in transcript).
- ARR: April 2025 $310M vs March 2025 "at least $220M" (prior call) — implies ~$90M net ARR added in one month. Q1 ended with "record high number" of managed customers; "hundreds" of customers total.
- CapEx: $544M spent in Q1 against new ~$2B FY2025 plan (implies ~$1.46B remaining across Q2–Q4). Prior guide was $1.5B.
- Cash: $1.44B at quarter-end vs $2.4B at year-end 2024. No debt.
- EBITDA: FY2025 negative reiterated; positive "at some point in the second half"; core infrastructure alone positive "probably sometime in Q3" — first time management has given a quarter-level marker, and only for the core business, not the group.
- Capacity: >100MW operational by end-2025 reiterated; pipeline ambition >1GW. Note: CEO said "more than 100 gigawatts" in one answer — almost certainly a misstatement for 1 gigawatt given every other reference; worth clarifying.
- Midterm targets (new): mid-single-digit billions revenue "in a few years"; 20%–30% EBIT margins medium-term, "well north of 30%" longer-term scenarios. Margin drivers cited: utilization, software mix, shift toward inference workloads.
- Depreciation: management claims a "full year depreciation schedule" versus peers' 5–6 years. This is likely a transcription garble — a four-year schedule would make sense in context (conservative vs peers); the literal wording is implausible. Flag for verification against the press release/20-F.
- Contracts: ranging from several months to a year and beyond; Blackwell fleet expected to enable longer-term contract discussions. No disclosed contract values, backlog, or customer concentration.
Product and launch scorecard- Blackwell: B200 deploying in Kansas City part two, available "a bit later in Q2"; GB200 deployment starting; Blackwell Ultra (GB300 NVL72) first deployments in Q3, with Nebius claiming it will be among the first to stand up GB300 NVL72 instances. Slightly more specific than last quarter's B200 Q2 / GB200 Q3 framing, and deployment is now visibly underway — but still no quantified Blackwell presale value, which remains the key unproven pillar of the back-half ARR guide.
- Platform reliability: Slurm-based upgrades (auto node recovery, proactive health checks) drove a quantified ~5% improvement in nodes available for commercial use — the most concrete operational metric management has given.
- Software: ~50 product launches in Q1 (MLflow and JupyterLab to GA, enhanced object storage, integrations with Metaflow/dstack/SkyPilot, storage partnerships with DDN/VAST/WEKA). Management again concedes standalone software revenue contribution is "relatively small" today while positioning it as the primary long-term margin driver.
- NVIDIA ecosystem: Blackwell Ultra AI Factory first-mover status, Dynamo launch partner, 1 of 5 reference NCP partners, DGX Cloud Lepton marketplace support — strong vendor validation, though none of these carries disclosed commercial value.
- Customer proof points: named wins Captions (AI video, Mirage model) and Quantori (biopharma, 3D molecular generation); vertical momentum cited in healthcare/life sciences, media/entertainment, financial services. Qualitative only — no deal sizes.
- Toloka: Bezos Expeditions and Mikhail Parakhin (Shopify CTO) investing; customer list disclosed (Amazon, Anthropic, Microsoft, Poolside, Recraft, Shopify); deconsolidating from Q2 reporting. Genuine external validation, but the company loses consolidated revenue — the stated Q1 revenue growth figures presumably include Toloka; ex-Toloka comparability will need reconstruction.
- Avride: partnerships now cited as Uber, Hyundai, Grubhub, Rakuten (Hyundai and Rakuten are new names versus last quarter's disclosure); confirmed "active discussions" with strategic investors. Still no revenue figures.
- ClickHouse: 28% stake; management cites press reports of a ~$6B fundraise — implies stake value ~$1.7B on those reports (my arithmetic), which is how the funding narrative gets its weight. Not confirmed by ClickHouse itself.
Sell-side read-through- Q&A remained management-curated: IR (Neil Doshi) read and assigned consolidated webcast questions to executives; no named analysts pressed live. External pressure-testing on the call remains minimal.
- Question themes covered: midterm definitions, Q1 ARR drivers, EBITDA timing, CapEx raise rationale, funding sources, growth segments, Toloka deconsolidation, capacity/Israel, GPU rollout, tariffs, customers, contracts, NVIDIA, software monetization, Avride.
- Notably absent or unpressed: absolute Q1 revenue, current utilization rates, customer concentration, quantified Blackwell presales/backlog, churn progression after the Q4 departures, the magnitude/timing of any equity raise, and the revenue mix shift from Toloka deconsolidation. The $1B sequential cash decline was addressed only through CapEx ($544M) — the remainder (including the known $180M tax payment) was not reconciled on the call.
- The repeated funding questions (asked twice in different forms) signal investor focus on the gap between $1.44B cash, $2B 2025 CapEx, and multi-year build-out ambitions; management's answer leans heavily on unquantified, unannounced stake monetizations.
Management credibility- Positive: the March ARR marker ("at least $220M") was beaten within weeks (April $310M); capacity milestones from the prior call delivered on schedule (Kansas City live — explicitly the last Hopper deployment; Iceland fully operational); guidance reiterated rather than stretched despite the strong start; EBITDA breakeven guidance became more specific (core infra Q3, group H2).
- Candor markers: software revenue contribution still "relatively small"; frontier labs explicitly "not in our revenue yet"; tariff answer hedged ("don't believe... major changes... very dynamic situation").
- Credibility risks: CapEx guidance raised ~33% within one quarter, with a quarter-slip explanation that accounts for part but not all of the increase; midterm targets ("mid-single-digit billions," 20–30% EBIT margins, 1GW+) are new, undefined in years ("a few years... we'll go as quickly as we can"), and unfunded on current cash; the funding plan relies on stake monetizations that are "in discussions" or based on press reports; the CEO's "more than 100 gigawatts" slip echoes last quarter's corrected Avride stumble — small, but worth noting on an otherwise polished call.
- Language to track: "well on track" (ARR guide), "probably sometime in the third quarter" (core EBITDA), "minimizes dilution... prudent in terms of debt" (funding) — all directional, none contractual.
What changed versus the prior quarter- Execution trajectory reversed: Q4 was explained as timing misses (customer completions, migration drag, slow sales ramp); Q1 delivered the claimed recovery — record managed customers, April ARR $310M vs the $220M March marker, demand "strengthening each month."
- CapEx posture shifted: $1.5B → ~$2B for 2025, and cash fell from $2.4B to $1.44B — the funding question moved from "opportunistic future raises" to an explicit multi-source financing narrative (stake sales, capital markets, low debt).
- Guidance granularity improved: EBITDA breakeven refined from "at some point during the year" to core-infrastructure Q3 / group H2; new midterm revenue and margin targets introduced for the first time.
- Portfolio structure changed: Toloka deconsolidation announced (external investors, voting <50%, majority economics retained) — Q2 reporting will be restated ex-Toloka; Avride strategic-investor talks moved from "planned" to "active discussions"; ClickHouse stake explicitly positioned as a funding source at a cited ~$6B valuation.
- Footprint: Israel added (first Middle East site, national-AI-factory framing); New Jersey build-to-suit named as the US Blackwell facility with late-summer first capacity; Kansas City and Iceland now live.
- Blackwell moved from presales talk to physical deployment (Kansas City), with GB300 Ultra added to the near-term roadmap; still no presale value disclosed.
- Dutch tax item not mentioned this call — the $180M payment was flagged last quarter; final settlement status unconfirmed in this transcript.
Bull case- Momentum is now evidenced, not promised: April ARR of $310M implies the $750M–$1B December exit requires roughly 2.4–3.2x growth over eight months versus the 3.4–4.5x required as of last quarter's $220M marker (my arithmetic) — the ramp is de-risking on schedule.
- Demand signals: Q1 revenue ~+400% YoY, record managed customer count, sales ramp (presales/solution architects/24x7 support) visibly contributing, SemiAnalysis ClusterMAX Gold recognition, and the DeepSeek H200 demand spike handled in "cloud manner" as proof of fleet flexibility.
- Capacity is delivering on time (Kansas City, Iceland live; Finland on track; New Jersey late summer; Israel added) with >100MW year-end reiterated and a >1GW pipeline ambition.
- NVIDIA alignment is deep and multi-vector (investor, Blackwell Ultra first-wave, Dynamo, reference NCP partner, Lepton marketplace) — supply access and co-validation that most neoclouds lack.
- Funding optionality: $1.44B cash, no debt, and non-core stakes (ClickHouse at a cited ~$6B round, Toloka now externally validated, Avride in investor talks) that could fund growth with limited dilution.
- Market expansion vectors are credible and sequenced: AI-natives (current), frontier labs, enterprises, sovereign AI factories — each requiring the scale now being built.
Bear case- Cash burn is steep and the funding plan is soft: ~$1B cash decline in one quarter; $1.44B remaining vs ~$2B 2025 CapEx and a stated multi-year build toward 1GW+. The plan rests on monetizing stakes at valuations supported only by press reports and "active discussions," plus unspecified capital-markets raises — dilution or leverage timing and terms remain unknown.
- The ARR guide is still back-half loaded: $310M April ARR to $750M–$1B by December depends on Blackwell deployments landing on time (GB200s only now deploying, GB300 in Q3) and converting "discussions" on longer-term contracts into signed revenue — presale values remain undisclosed.
- Guidance quality: CapEx rose ~33% in one quarter; midterm targets are deliberately undefined ("a few years"); no absolute Q1 revenue was stated on the call, and Toloka deconsolidation will make Q2-onward growth rates non-comparable to the headline figures given this quarter — creating a disclosure air pocket.
- Depreciation-policy claim as transcribed ("full year depreciation schedule") is likely an error but, if verified as stated, would flatter margins versus peers on 5–6 years — either way it needs checking before crediting the 20–30% EBIT margin target.
- Concentration and durability unknown: revenue skews to US venture-backed AI-natives; no utilization, churn, or concentration metrics disclosed; the Q4 pattern (two large customers rolling off) could recur.
- Toloka deconsolidation reduces reported scale; Avride remains capital-intensive with funding dependent on third parties; tariff/regulatory answers remain "we don't believe"-level hedges.
Next-quarter watchlist- Q2 reporting ex-Toloka: restated revenue base, updated segment disclosure, and whether headline growth claims survive the deconsolidation; absolute Q1/Q2 revenue dollars from the press release (absent from this transcript).
- ARR trajectory: monthly/quarterly ARR updates vs the $310M April print; evidence the H2 ramp is contracting the gap to the $750M–$1B December exit.
- EBITDA milestones: core-infrastructure positive adjusted EBITDA in Q3 (first hard profitability test); group H2 crossover; margin progression.
- Blackwell conversion: B200 availability "later in Q2," GB300 first deployments in Q3, and — critically — any quantified presale values, contract lengths, and pricing for the Blackwell fleet.
- Funding events: ClickHouse round confirmation and valuation; Avride strategic investor announcement; any equity/debt raise; cash balance versus the ~$1.46B remaining 2025 CapEx.
- Capacity milestones: New Jersey first capacity late summer; Finland phase one late Q3; Israel build details (size, timing, capital, customer commitments); progress beyond 100MW.
- Depreciation policy verification: reconcile the "full year" transcript wording against the 20-F/press release to confirm the actual useful-life assumption.
- Toloka terms: stake retained, transaction economics, and any revenue/contribution disclosure before deconsolidation takes effect.
- Customer metrics: replacement of Q4 churned large customers, contract-length mix shift with Blackwell, utilization disclosure, enterprise/lab pipeline evidence.
- The CEO's "100 gigawatts" remark — confirm it was a slip for 1GW.
- Dutch tax: confirmation the $180M payment closes the matter (unaddressed this call).
|
| Feb 20, 2025 | +3.17% | Q4 FY2024 | Read transcript briefingQuarter in one view- Nebius Group's second earnings call as a public company (Q4/FY2024, call dated 2025-02-20). Core message: 2025 ARR guidance of $750M–$1B (December 2025 exit) reiterated, anchored on a stated March 2025 annualized run-rate revenue of "at least $220 million" from contracts already in place.
- FY2025 revenue guidance reiterated at $500M–$700M; EBITDA guided negative for the full year with breakeven crossed "at some point during the year" — no specific timing given.
- Q4 revenue/run-rate underperformance was attributed to timing: longer customer lead times/POCs, two large customers completing engagements in Q4 without replacement before year-end, resources consumed by migrating all customers to the new AI cloud platform, and late-quarter sales hires.
- Capacity: ~38,000 GPUs total by end of March (20,000 H200s); Finland tripling on schedule (first delivery Q3, remainder by year-end); France facility live; Kansas City (first US cluster) deploying, live by end of Q1; new Iceland H200 cluster available March; new build-to-suit US facility for Blackwell signed. Guidance: 100MW in operations by end-2025, with secured sites scalable to 300MW+.
- Cash: $2.4B at year-end including the $700M December raise (NVIDIA, Accel, Orbis among investors). Dutch tax liability settled at $180M paid in February vs up-to-$400M previously flagged — ~$220M freed for expansion, though "not 100% done."
- Other businesses: Toloka revenue +140% in 2024; TripleTen revenue +250%, ~14,000 students (figure marked phonetic in transcript); Avride signed Uber (one of two autonomous partners alongside Waymo/Zoox per management) and Grubhub (~100 robots at Ohio State, ~1,000 deliveries/day in week one); 28% ClickHouse stake retained, unconsolidated.
What management is focused on- Three stated 2025 priorities: (1) full-stack AI technology build-out (data centers in Finland, France, Iceland, Kansas City; first owned-design US data center to be announced; AI cloud and AI Studio inference platform enhancement); (2) capital for growth ($2.4B cash, public-company financing access, "opportunistic" future raises); (3) corporate/sales infrastructure, especially US go-to-market.
- US market emphasis is repeated throughout: sales hires, San Francisco and Dallas offices, Kansas City cluster, build-to-suit US Blackwell facility, US customer concentration.
- Blackwell transition is the central 2025 revenue thesis: H200 deployments finish late March/early April, then B200 (Q2), GB200 (Q3), GB300 (later in year). Management calls Blackwells "maybe the main source of our income."
- Full-stack differentiation (data center engineering/PUE, in-house servers/racks/cooling, bespoke AI cloud platform, value-added software services) framed as the moat versus "Neo Cloud" peers and commoditizing bare-metal offerings.
- IR/visibility build-out: zero sell-side coverage acknowledged; new Head of IR (Neil Doshi, SF-based) prioritizing analyst coverage initiation, conferences (Goldman Disruptive Technology Symposium, March), possible data-center tours.
Key numbers and quarter mechanics- ARR: March 2025 run rate "at least $220M" on existing contracts; December 2025 exit ARR guide $750M–$1B reiterated. Implied ~3.4–4.5x run-rate growth over nine months — heavily back-loaded, with management explicitly stating most ARR/revenue lands in H2, "weighted more towards the end of the year."
- FY2025 revenue guide $500M–$700M; EBITDA negative for the year, breakeven crossed intra-year, timing unspecified.
- GPU fleet: ~38,000 GPUs by end of March, of which 20,000 H200s; H200 deployment began late October/November 2024.
- Capacity: 100MW operational by end-2025 guided; 300MW+ secured/scalable; greenfield expansion in US and Europe targeting "multiply times" current capacity; CEO has floated gigawatt-scale ambition conditional on demand.
- Cash $2.4B at year-end 2024; $700M December PIPE oversubscribed.
- Dutch tax: $180M paid February 2025 vs prior up-to-$400M exposure; ~$220M delta earmarked for expansion; final settlement not fully closed.
- ROIC framing: H200 payback period estimated at 2.5–3 years; Blackwell payback "too early to say," with claims that data-center efficiency, in-house hardware, and high-margin software/services accelerate returns. Software layer admitted to be "very small today."
- Q4 mechanics: revenue softness explained as timing — longer POC-driven sales cycles (framed as industry-wide and a function of larger, multi-thousand-GPU deals), two large customer completions, platform migration drag, and sales hires arriving late in the quarter (six-month ramp to full productivity).
- Pricing: H100 price pressure in Q4 acknowledged ("as all the market did"); mitigation claimed via cost structure and mix shift to higher-priced Blackwells.
- Churn: acknowledged at the lower end (self-service, AI Studio, on-demand) and framed as structural to the model; DeepSeek drove a late-January spike in demand for "several thousand" H200 inference chips.
Product and launch scorecard- AI cloud platform relaunch: completed in Q4 with full customer-base migration — a real execution milestone, but management concedes it consumed resources and contributed to the Q4 revenue timing miss. No adoption or revenue metrics disclosed for the new platform.
- AI Studio (Inference-as-a-Service): launched; validated circumstantially by the DeepSeek-driven H200 inference demand spike; no usage or revenue figures.
- Blackwell presales: launched end of 2024; management reports "positive" demand, pre-commitments at "multiple thousands of GPUs," contract lengths of one year and longer, and expects to pre-sell "a proportion" of initial deliveries. No signed-value disclosure — the key proof point for the back-loaded ARR guide remains unquantified.
- Capacity deliveries: France live; Kansas City on track for end-Q1; Iceland H200 cluster for March; Finland expansion first phase Q3. Near-term milestones are specific and checkable.
- Avride: strongest third-party validation on the call — Uber partnership (robots live on Uber Eats in Austin, Dallas, Jersey City; cars expected in service later in 2025), Grubhub campus deal with ~100 robots at Ohio State doing ~1,000 deliveries/day in week one (CEO initially misspoke the numbers; corrected on the call), Japan certification received. Co-investor(s) planned given capital intensity; nothing specific yet.
- Toloka: +140% FY2024 revenue; pivot from crowdsourced labeling to expert-driven GenAI work (red teaming, reasoning-model evaluation, coding/math training); claims "virtually all" leading AI labs as customers. No absolute revenue disclosed.
- TripleTen: +250% revenue; ~14,000 students (transcript flags the number as phonetic); doubled new-student additions. No absolute revenue disclosed.
- ClickHouse: 28% stake retained, flagged as "significant potential source of value," outside consolidation.
Sell-side read-through- No named sell-side analysts asked questions; the Q&A was run by management (Tom Blackwell) reading submitted/chat questions — consistent with the disclosure that the company currently has zero covering analysts. This itself is a read-through: the story is not yet being independently pressure-tested on the call.
- Recurring question themes (per management's own selection): capacity expansion beyond guidance, confidence in the ARR guide, Q4 revenue dynamics, sales ramp timing, customer mix and contract length, competition, DeepSeek impact, full-stack advantage, Blackwell deployment timing, revenue cadence through the year, 2026+ capital raising, ROIC/payback, other-business progress, AI Diffusion rule/UVEU status, EU tariff risk and possible US relocation, Stargate/European mega-projects, H100 pricing, GPU counts, guidance reiteration, Q4 churn, analyst coverage, Dutch tax.
- Notable management-selected soft spots: no question was taken pressing the Q4 miss magnitude in absolute dollars, no question on current utilization rates, no question on customer concentration, no question quantifying the two large Q4 customer completions, and no question on the size/timing of the next capital raise beyond "opportunistic."
- Regulatory answers were reassuring but hedged: UVEU qualification "anticipated," tariff risk "no material risk" — both explicitly caveated as early reads.
Management credibility- Positive markers: the Dutch tax item resolved at $180M vs the up-to-$400M prior flag — a concrete, favorable resolution of a disclosed liability. Guidance was reiterated rather than walked back after a soft Q4, and management gave a specific near-term marker (March ARR ≥$220M) that is verifiable within weeks.
- Candor markers: Q4 churn and H100 price pressure acknowledged directly; the CEO's Avride delivery-number stumble was corrected on the call; software-layer contribution admitted to be "very small today"; Blackwell payback declined as "too early to say."
- Credibility risks: the Q4 shortfall is attributed entirely to timing with no quantification of the miss or the churned customers; the $750M–$1B exit ARR requires steep H2 acceleration that rests on (a) unquantified Blackwell presales, (b) sales hires reaching productivity, and (c) on-time data-center deliveries — each individually plausible, collectively unproven. The "gigawatt" rhetoric sits uneasily next to a 100MW year-end target and an acknowledged need for capital "beyond what we currently have."
- Language to track: "well within reach" (ARR guide), "reasonably confident" (tax finality), "we don't anticipate any material negative impact" (regulatory) — all hedged; none are commitments.
- Note on context: the supplied prior-quarter summary covers Yandex FY2021 (Q4 2021) — the predecessor business, three years and a corporate separation removed. It is not a comparable sequential quarter for Nebius Group; continuity items (Avride/Rovers, Toloka, ClickHouse, cloud) are the only meaningful threads.
What changed versus the prior quarter- Because the supplied summary is Yandex Q4 2021, a true sequential comparison is not possible; the call itself states this is only the second quarterly call as Nebius. Changes identifiable from within this transcript and its own references to the prior (Q3 2024) call:
- Guidance posture: 2025 ARR ($750M–$1B exit), revenue ($500M–$700M), and EBITDA (negative FY, intra-year breakeven) guidance all reiterated — no raise despite the claimed March $220M run rate and Blackwell presales.
- Capital structure transformed: resumed NASDAQ trading in October 2024; $700M oversubscribed PIPE in December with NVIDIA, Accel, Orbis entering the register; $2.4B year-end cash.
- Dutch tax overhang: moved from an open up-to-$400M liability to $180M paid, pending final confirmation.
- Capacity footprint: from Europe-only (Finland, France) to adding Iceland and the first US cluster (Kansas City), plus a signed US build-to-suit Blackwell facility — a clear geographic pivot toward US demand.
- Platform: new AI cloud platform launched and full customer migration completed in Q4; AI Studio launched.
- Avride: from the 2021-era "Rovers with US contracts" concept stage to named commercial deployments (Uber Eats in three cities, Grubhub at Ohio State) and Japan certification — the most tangible multi-year progression visible across the two documents.
- Disclosure evolution: management now gives GPU counts, MW targets, and a near-term ARR marker — more operational specificity than the legacy Yandex-era calls, though absolute revenue, utilization, and segment P&L for Toloka/TripleTen/Avride remain undisclosed.
Bull case- Contracted momentum: March ARR of at least $220M already in hand, with the full-year $750M–$1B exit guide underpinned by capacity (100MW by year-end, 300MW+ secured), undeployed H200s, and reserved Blackwell supply.
- Demand signals are current, not historical: DeepSeek triggered a several-thousand-GPU H200 inference spike in late January; Blackwell presales show multi-thousand-GPU pre-commitments at one-year-plus terms; deal sizes are shifting to multi-thousand-GPU clusters.
- Cost/moat argument is coherent: in-house data-center engineering (PUE), own racks/servers/cooling, and a purpose-built AI cloud stack support the claimed 2.5–3-year H200 payback and resilience through H100 price declines; the software/services layer offers margin upside if it scales.
- Balance sheet: $2.4B cash, an oversubscribed raise with strategic investors (NVIDIA), public-market financing access, and a $220M tax saving versus prior expectations.
- Portfolio optionality: Avride with live Uber/Grubhub deployments and Japan entry, Toloka at +140% with top AI labs as customers, TripleTen at +250%, and a 28% ClickHouse stake — multiple shots on goal beyond core compute.
- Structural tailwinds: Stargate-scale government AI infrastructure programs ($500B US, $100B France cited) position Nebius as a potential builder; management claims best-in-class positioning for this work.
Bear case- The guide demands a steep H2 ramp: from ≥$220M March ARR to $750M–$1B by December requires roughly 3.4–4.5x growth in nine months, dependent on on-time data-center construction, GPU delivery/deployment, Blackwell pricing, and a sales force that management itself says takes six-plus months to reach productivity. Any slippage compounds.
- Q4 already showed execution fragility: two large customers rolled off and weren't replaced in time; churn is structural at the low end; POC-lengthening sales cycles could persist regardless of the "timing issue" framing.
- Capital intensity is open-ended: management concedes 2026+ plans require "significant capital beyond what we currently have" — dilution or leverage is coming, timing and terms unknown, and the gigawatt rhetoric implies the appetite is large.
- Pricing risk is real: H100 prices fell in Q4; the mitigation thesis (Blackwell mix at "higher pricing") assumes next-gen premiums hold as supply across Neo Clouds and hyperscalers expands.
- Disclosure gaps: no absolute Q4 revenue figure discussed on the call, no utilization rates, no customer concentration, no quantified Blackwell presale value, no segment financials for Toloka/TripleTen/Avride, no EBITDA breakeven timing. The $220M March ARR is "at least" language on contracts "already in place" — durability and pricing of those contracts are undisclosed.
- Regulatory/geopolitical exposure is answered with optimism, not analysis: UVEU status "anticipated," tariff risk dismissed, Amsterdam HQ retained — all subject to policy implementation not yet settled.
- Zero sell-side coverage means limited external scrutiny and potential volatility as coverage initiates.
Next-quarter watchlist- March ARR print vs the "at least $220M" marker — the first hard test of the reiterated guide; any revision or silence is a signal.
- Blackwell cadence: B200 availability in Q2, GB200 in Q3; conversion of "pre-commitments" into signed contracts with disclosed values, lengths, and pricing.
- Capacity milestones: Kansas City live by end of Q1; Iceland H200 cluster in March; Finland first-phase delivery in Q3; details of the signed US build-to-suit Blackwell facility and the promised first owned-design US data center announcement.
- Revenue cadence: evidence of H1 traction vs the guided H2-weighted ramp; Q1 revenue against the $500M–$700M FY range; any update on intra-year EBITDA breakeven timing.
- Churn and customer metrics: replacement of the two large Q4 departures, customer count progression ("hundreds" toward "thousands"), contract mix shift toward longer terms, utilization disclosure.
- Capital raising: size, structure, and timing of the next raise; any debt financing; further detail on 2026+ capex scope.
- Dutch tax final confirmation that $180M closes the matter.
- Avride: Uber car deployments "later this year," Grubhub campus expansion beyond Ohio State, Japan activity, and the promised co-investor announcement.
- Toloka/TripleTen: any absolute revenue or profitability disclosure to replace growth-rate-only reporting.
- Regulatory: implementation of the AI Diffusion framework and UVEU qualification; any EU tariff developments affecting US-served-from-EU workloads.
- Sell-side coverage initiations and the Goldman March conference appearance — first external framing of the story.
|
| Oct 31, 2024 | -3.08% | — | Transcript briefing unavailable |
Corbanu