Philip A. Fisher | February 22, 2026 Task: Earnings Review | Period: Q4 FY25 (quarter ending December 2025)
No prior Phil Fisher analysis on record for Nebius Group. This is a first-look earnings review. The Atlas baseline (Q4 FY25, 2026-02-21) provides a Conviction 4/5 verdict with six-factor Strong ratings on Growth, Trajectory, and Dominance. My task is to apply the Fisher framework independently and then determine where I agree, disagree, or add nuance.
Scuttlebutt is the foundation of my method. I will not form a view on management quality or competitive position from numbers alone. The following is drawn from active research across all seven categories.
Customers. The customer case study corpus is thin — appropriate for a company that went commercial only in mid-2024. Named early adopters include CRISPR-GPT researchers from Stanford, Princeton, and Google DeepMind; vLLM (the leading open-source LLM inference framework); and Slingshot AI. The caliber of these early adopters is unusually high. When the customers choosing your product include the researchers building the dominant inference framework, that is not coincidence — it is a signal of technical credibility that no marketing budget can manufacture. Capacity has been sold out since at least Q3 FY25. Volozh himself stated the Meta contract "was limited to the amount of capacity we had available." I have rarely encountered a clearer demand signal than a vendor who must turn away business.
Employees. Glassdoor shows 44 reviews, 84% recommend rate. The more significant data point is what happened between 2022 and 2024: when Yandex's Russian business was sold, Volozh retained a core nucleus of 850+ engineers who followed him through sanctions, redomicile, and a complete reset. People do not follow a manager through that kind of disruption unless they believe in him. That voluntary retention under extreme adversity tells me more about the quality of Volozh's leadership than any management survey. The culture friction noted in Glassdoor reviews — primarily around integrating a Russian-tech operating style with Western GTM hires — is real but not alarming. It is a predictable friction at this stage of company building.
Competitors. SemiAnalysis ranks CoreWeave Platinum and Nebius Gold. CoreWeave has 10x GPU fleet scale. The honest read: CoreWeave is the superior platform for large-scale pre-training at present, with Microsoft and OpenAI as anchor customers representing 77% of revenue. Nebius is building a different product — a full-stack managed AI cloud rather than raw GPU rental — and it has the superior balance sheet ($3.7B cash, net debt of only $450M vs. CoreWeave's $8B in debt-SPV financing). This is not a case of Nebius losing to CoreWeave; it is a case of two different strategies. The balance sheet durability matters most if GPU prices compress in a price war.
Management. Arkady Volozh built Yandex from zero to Russia's dominant internet company — a genuine tier-1 operator record. Former COO Greg Abovsky describes him as "a peer, like a player-coach almost" and notes he is "really good at gathering really, really smart people." Volozh's own framing — comparing AI infrastructure to railways in the industrial revolution — is the language of a long-range thinker, not a cycle-trader. His 1GW datacenter capacity target for 2026 implies $10B+ in annual revenue potential at current utilization rates. I note one concern on management focus: Volozh is simultaneously running Avride (autonomous vehicles) as a separate subsidiary. A founder who divides attention across two businesses of this scale is a concentration risk. I have seen capable managers spread themselves too thin before. This warrants monitoring.
Product/Technology. Nebius participated in MLPerf Inference v5.1 benchmarks and achieved first-place rankings among GB200 NVL72 submitters for Llama 3.1 405B. Of greater significance: their 128-node cluster demonstrated near-linear scaling on large-model training. This is technically difficult. Most providers deliver excellent single-node specifications but lose efficiency at cluster scale due to networking constraints. Nebius designs its own GPU server racks internally — eliminating Dell and HPE as intermediaries — and the performance data suggests this vertical integration is producing real results, not just cost savings.
Hiring. 269 open roles for a company of ~1,000-1,200 employees is approximately 20%+ of current headcount. This signals aggressive expansion, consistent with the 16 data center build target. The urgency in some hiring processes is consistent with supply constraints (not enough people) rather than demand constraints (not enough work). The current bottleneck is human capital, not customer demand.
EU Sovereignty. Nebius is Netherlands-domiciled, designing EU-compliant infrastructure (automated compliance audits, data anonymization, transparency dashboards) as the EU AI Act comes into full force in 2026. NVIDIA has specifically named Nebius (alongside Nscale) as a preferred European partner for deploying 14,000 Blackwell GPUs in UK data centers. NVIDIA controls GPU allocations. Being a named preferred European partner signals allocation priority — a structural advantage in a world where GPU supply remains constrained.
Tavily Acquisition. Announced February 10, 2026. Nebius acquired Tavily — an Israeli agentic search startup, one year old — for $275M upfront plus up to $400M contingent. Tavily has 3M+ monthly SDK downloads, 1M+ developers, and Fortune 500 customers including IBM. The strategic logic is sound: agentic AI systems require both inference compute and real-time context retrieval. Keeping both in the same stack reduces latency and eliminates cross-vendor data governance complexity. Futurum Group frames this as shifting Nebius from "GPU rental" to "platform" — the AWS analogy of moving from EC2 to managed services. I find this analogy instructive but imprecise. The risk: $275-400M is aggressive for a one-year-old company, and Nebius is simultaneously building 16 data centers, integrating Avride, and scaling headcount. Focus diffusion is a legitimate risk for any management team. Volozh has done harder things. But the number of simultaneous execution requirements is high.
Q4 FY25 revenue of $227.7m represents 547% YoY growth [GAAP]. Core AI revenue of $214.2m grew 830% YoY [GAAP] and constitutes 94% of total revenue. For the full year FY25, revenue was approximately $530m.
I have been investing for a long time, and I have learned to be skeptical of percentage growth numbers when the base is very small. But the dollar adds here are not trivial: Q4 FY25 added substantially more absolute revenue than any prior quarter. The ARR trajectory is the more instructive metric: $90m at Q4 FY24 → $1,250m at Q4 FY25. That is 13.9x in twelve months. Deferred revenue of $1,577.5m at quarter-end provides exceptional near-term revenue visibility — the FY26 guide of $3.0-3.4B is not a projection drawn from thin air but a near-mathematical consequence of contracts already signed.
FY26 guide of $3.0-3.4B [Non-GAAP basis, to be confirmed] implies approximately 500%+ growth at the midpoint of $3.2B. At $7-9B ARR as the exit target, FY27 could approach $5-8B in revenue if ARR converts to revenue at current rates. This is not speculative extrapolation — it is arithmetic from signed contracts.
Gross margin of 70% [GAAP] in Q4 FY25. This is outstanding for a capital-intensive infrastructure business. Most traditional data center and colocation businesses operate at 40-60% gross margins; pure cloud providers often achieve 60-75%. Nebius is at the high end of this range while still in the earliest stages of operating leverage — a favorable starting point.
Core AI Adjusted EBITDA margin: 24.0% in Q4 FY25, up from 19.1% in Q3 FY25 [Non-GAAP]. This sequential improvement is the margin trajectory signal I look for in a growing infrastructure business. First positive consolidated Adj EBITDA in Q4 FY25. FY26 guidance of 40% Adj EBITDA margin implies a 16-percentage-point improvement in one year — achievable only if the denominator (revenue) grows dramatically faster than the numerator's fixed cost base. Given $3.0-3.4B guidance and the deferred revenue runway, the math is credible.
GAAP operating margin remains deeply negative (-103% in Q4 FY25), reflecting the early-stage CapEx intensity and D&A from a rapidly expanding asset base. I do not find this disqualifying. Fisher knew that outstanding companies in their growth phases often carry large accounting losses while generating powerful underlying economics. The 70% gross margin and improving Adj EBITDA tell the underlying story more clearly than the GAAP operating loss.
One genuine concern: the EPS miss (-0.68actualvs.−0.54 estimate) is unexplained without the earnings call transcript. This is an $0.14 miss, approximately 26% worse than expected. Without the transcript, I cannot determine whether this reflects accelerated investment (management's decision) or cost overruns (an execution failure). The distinction matters enormously. I am treating this as an open question pending transcript publication.
CapEx of $2,056m in Q4 FY25 alone is extraordinary. This is a company spending over $2B per quarter to build the infrastructure its $3.2B FY26 revenue guide requires. Cash of $3,678m and net debt of only $449.6m provides a reasonable near-term cushion. OCF of $834.3m [positive, driven by deferred revenue cash received] partially offsets the CapEx burn.
The most significant red flag in the financial data: no FY26 CapEx guidance. The prior Q3 guide of $5B annualized was not updated or confirmed. For a business with this level of capital intensity, the absence of CapEx guidance makes it impossible to independently verify the 40% Adj EBITDA margin commitment. I have generally found that managements who fail to disclose significant capital allocation plans are either (a) uncertain themselves, or (b) aware the numbers would concern investors. Neither interpretation is reassuring. I will watch for clarification when the transcript is published.
Debt of $4,127.7m at Debt/OCF of approximately 5x is elevated but not alarming given the asset base being built. The critical question is whether the assets being built will generate the contracted cash flows. Deferred revenue of $1,577.5m suggests they will.
Point 1 — Market potential: The AI infrastructure market is the defining build-out of this decade. Demand is constrained only by supply. All available capacity is sold out. PASS.
Point 2 — New product development: Tavily acquisition, custom rack design, Blackwell Ultra deployment, Vera Rubin commitment for H2 2026, enterprise software layer (compliance tools, capacity management). PASS.
Point 3 — R&D effectiveness: MLPerf first-place ranking on GB200 for Llama 3.1 405B inference. Near-linear 128-node cluster scaling. Custom rack design producing measurable performance advantages. R&D produces results. PASS.
Point 4 — Sales organization: Building rapidly — 269 open roles include significant sales/BD hiring. Named customer wins at Microsoft (17 − 19Bcommitment)andMeta(3B over 5 years) suggest enterprise sales capability exists at the highest tier. PASS, with monitoring given early stage.
Point 5 — Profit margins: 70% gross margin [GAAP]. Core AI Adj EBITDA 24% and improving. For stage and capital intensity, these are excellent. PASS.
Point 6 — Improving margins: Core AI Adj EBITDA improved from 19.1% to 24.0% sequentially. FY26 target of 40% implies substantial further improvement. Trajectory is correct. PASS.
Point 7 — Labor relations: 84% Glassdoor recommend rate. Culture integration friction is present but not alarming. PASS with monitoring.
Point 8 — Executive relations: The retention of 850+ engineers through Yandex's dissolution is an extraordinary signal. Former colleague characterizations of Volozh are consistently strong. PASS.
Point 9 — Management depth: Volozh is exceptional. The bench depth below him is less clear from available scuttlebutt. The simultaneous Avride subsidiary is a focus concern. CONDITIONAL PASS — monitoring required.
Point 10 — Financial controls: The absence of FY26 CapEx guidance for a business spending $2B/quarter on CapEx is unusual. Possible explanations include genuine uncertainty or deliberate opacity. YELLOW FLAG.
Point 11 — Other business aspects: NVIDIA preferred European partner designation is a structural allocation advantage. Vertical integration (custom rack design) is defensible. EU regulatory compliance embedded as a product feature is a moat. PASS.
Point 12 — Long-range outlook: Volozh frames this as "railways of the industrial revolution." 1GW capacity target for 2026. Vera Rubin deployment planned for H2 2026. The long-range thinking is evident and credible. PASS.
Point 13 — Dilution: Not available in the scout brief. Share count history and equity compensation not quantified. INCOMPLETE — requires follow-up.
Point 14 — Management transparency: The sold-out capacity disclosure and the candid description of the Meta contract size being capacity-limited are examples of appropriate transparency. The EPS miss without explanation (transcript not yet published) and the absent CapEx guidance are transparency deficits. MIXED — trending positive but two active concerns.
Point 15 — Integrity: Volozh voluntarily divested Yandex's Russian business under sanction pressure — a complex, multi-year process that required no deception and resulted in a clean break. No self-dealing indicators. No pattern of promise abandonment in the short operating history. PASS.
Summary: 12 points pass, 1 yellow flag (CapEx disclosure), 1 incomplete (dilution), 1 mixed (transparency). This is a strong Fifteen Points profile for a company this early in its commercial history.
In my experience, the most important distinction to make about a growing company is whether it is growing because it is fortunate — riding an industry tailwind that would carry any participant — or whether it is growing because it is able — management is actively creating advantages that competitors cannot easily replicate.
In the GPU cloud market, many companies are fortunate. The AI infrastructure buildout is real, demand exceeds supply, and almost any competent provider can find customers. This is the railway boom of our era — and in a railway boom, many railways get built and most generate returns.
The question I must answer is: is Nebius merely fortunate, or is it also able?
The evidence strongly suggests both. The custom rack design producing measurable MLPerf advantages is not luck — it is engineering. The NVIDIA preferred European partner designation was earned, not given. The 850-engineer team that followed Volozh through personal and corporate adversity is a human asset that cannot be purchased. The Tavily acquisition — whatever one thinks of the price — reflects a management team thinking about platform economics, not just rack economics. The 70% gross margin in a business that started commercial operations in mid-2024 reflects a thoughtful approach to pricing and cost management.
I have seen many companies that were fortunate and able in the early years of an industry buildout. Some of them became the defining businesses of their generation. Nebius is not guaranteed to be one of them — the execution demands over the next 24 months are formidable. But the characteristics of an outstanding growth company are present.
Atlas assigned Conviction 4/5 and rated Growth, Trajectory, and Dominance as Strong. I am in agreement on the facts underlying those ratings.
Where I add nuance:
On dominance: Atlas notes Nebius is "#2 neocloud" behind CoreWeave. I would reframe this. Nebius and CoreWeave are building different products for different customers. CoreWeave's Platinum tier is earned by its scale and its hyperscaler customer relationships. Nebius's Gold tier reflects its different strategy — full-stack managed cloud with an enterprise and research lab customer base. The competitive risk from CoreWeave is real but overestimated if one assumes they are competing for identical customers. They are not.
On the EPS miss: Atlas flagged the -0.68vs.−0.54 miss as unexplained without the transcript. I concur, and I weight this more heavily than Atlas did. A 26% EPS miss without explanation from management is a transparency gap. In my framework, management that misses estimates without explaining the cause warrants heightened scrutiny on Points 10 and 14 until resolved.
On the Tavily acquisition: Atlas's brief mentions this; Atlas's analysis does not evaluate it as a Fisher event. I assess it as strategically sound but execution-risk-adding. The platform analogy (AWS moving from EC2 to managed services) is instructive. The risk is focus: Nebius is simultaneously managing the largest infrastructure build in its history while integrating an acquisition and scaling a sales organization. Each of these demands is individually manageable; together, they test the depth of the management team.
On the missing CapEx guidance: Atlas cited this as the top anti-thesis risk. I agree. Without FY26 CapEx guidance, the 40% Adj EBITDA margin commitment cannot be independently verified. This is not a reason to avoid the stock, but it is a reason to require clarity before adding meaningfully.
Priorities for the Q4 earnings transcript (when published):
Catalysts for upgrading conviction:
Conditions for reassessing thesis:
I have never set a price target and I do not intend to start. Precise valuation of a company growing 500%+ per year is an exercise in false precision.
What I will observe: the Atlas analysis estimates EV/FY26 Revenue of ~7.8x and EV/FY26 Adj EBITDA of ~19.5x at a ~$24.6B market cap. For a business with the growth trajectory, margin profile, and visible backlog that Nebius has, these multiples are reasonable — not cheap, but not unreasonable for an outstanding company in an early growth phase.
The more useful question for my purposes is: could this business be worth substantially more in three to five years? If Nebius achieves $5-7B in revenue by FY27-28 at 40%+ EBITDA margins, the answer is clearly yes. The deferred revenue backlog and ARR trajectory suggest the $3.2B FY26 guide is credible. Whether FY27 continues the trajectory depends on whether Nebius can build enough capacity to meet the demand it clearly already has.
I am not troubled by the current valuation for a long-term position. The question is execution, not price.
Thesis status: Initiating — Intact
Nebius Group passes the Fifteen Points on the dimensions that matter most: management integrity, long-range thinking, research effectiveness, improving margins, and genuine competitive differentiation. The demand picture is unambiguous — capacity is sold out and management is capacity-constrained rather than customer-constrained. The ARR-to-revenue conversion, backed by $1.58B of deferred revenue, makes the FY26 guide credible rather than aspirational.
The open questions — unexplained EPS miss, absent CapEx guidance, dilution unknown, management focus with Avride — are real but none of them break the thesis at this stage. They are questions to be answered, not evidence of failure.
In my experience, the time to avoid a great business is when management quality is poor, the competitive moat is illusory, or the growth runway is overstated. None of those conditions apply to Nebius today. The time to own a great business is when the evidence supports quality, the demand is real, and the management team has demonstrated it can execute under pressure. The evidence supports all three.
I would own this. I would not trade it.
| Dimension | Prior (none — first look) | Updated |
|---|---|---|
| Management quality | No prior | Volozh is tier-1; team retention signal is extraordinary |
| Competitive position | No prior | Gold tier; full-stack moat vs CoreWeave's pure-scale |
| Growth runway | No prior | Multi-year; $20B+ in contracted customer commitments |
| Margin trajectory | No prior | 70% GM; Core AI Adj EBITDA 24% → 40% FY26 target |
| Key risk | No prior | Execution complexity across 16 DCs + Tavily + headcount |
| EPS miss | No prior | Unresolved — transcript needed |
| Capital allocation | No prior | Yellow flag — CapEx guidance absent |
Analysis saved: 2026-02-22 Next review trigger: Q4 FY25 transcript publication; Q1 FY26 earnings Reference: briefs/NBIS_earnings-review_2026-02-21/