Nebius Group: A Long-Term Owner's Perspective

Hunting ten-year five-baggers among great growth stocks — pressing the upside question: "Can it get much bigger?"

• QHow high is its market ceiling? Is it enlarging an existing pie, or creating an entirely new market?6/10

  A

  The ceiling is high and the direction is real, but Nebius is essentially "enlarging a pie that already exists and is exploding" (AI compute leasing), rather than creating an entirely new market out of thin air. So the question has never been "is the track big enough," but "how stable and durable a slice it can carve out of this large pie."

  First, how big is the pie itself. This is indeed one of the rare super-long runways repeatedly validated by multiple independent institutions. Gartner expects global generative AI spending to reach $644 billion in 2025, up 76.4% year over year, with the vast majority flowing to AI servers and infrastructure rather than software; the report further cites IDC ("global AI infrastructure spending of about $318 billion in 2025, exceeding $1 trillion by 2029"), McKinsey ("data-center-related capex demand of about $6.7 trillion by 2030, of which about $5.2 trillion is AI-related"), and Synergy's estimate for the neocloud track ("revenue exceeding $23 billion in 2025, approaching $180 billion by 2030"). In terms of demand direction, Nebius is standing on a genuinely large wave; on this point the report scores industry attractiveness 4/5, and I agree.

  But we must honestly distinguish two things: the track's ceiling ≠ this company's ceiling. What Nebius sells is NVIDIA GPU clusters, on-demand compute, storage, and platform services—a hybrid model of "compute leasing + platform + tools." This enlarges an existing pie whose demand is surging, rather than opening up a previously nonexistent demand category (it did not "invent AI," it took on the "outsourced compute for AI training/inference"). Businesses that "enlarge an existing pie" naturally take a discount under the Baillie Gifford LTGG framework: their ceiling is constrained by how much outsourcing share the entire neocloud track can pull away from hyperscalers building in-house, while the demand side (model companies, inference providers) is extremely technically capable and can build in-house or shop across vendors at any time.

  More critically, this large pie is also the pie fought over by "the most capital-intensive competition in the world." The most direct rival, CoreWeave, had single-quarter capex as high as $7.695 billion in Q1 2026 alone, and guided full-year capex of $31-35 billion, backed further by the in-house capacity of AWS, Azure, Google, and Oracle. In other words, the bigger the pie, the more players fight for it, and the greater the risk that returns on capital per unit get diluted—which is precisely why the report scores the moat only 2/5 and separately flags "a good industry with an unformed company."

  My conclusion: on the ceiling dimension, what Nebius gets is a "track score," not a "monopoly score." It sits in a genuine, decade-scale large market (its hardest bull argument), but it is one of many heavily capitalized players in this large market, doing the business of grabbing existing outsourcing share rather than creating a new market that it defines and monopolizes. To the question "how high is the market ceiling," the answer is: the track ceiling is extremely high and real, but the portion belonging to Nebius still depends today on whether it can hold share and pricing amid the capital flood from the giants and its peers, and that has not yet been proven.

  Jun 11, 2026
• QOver the next five years, can its revenue at least double? Is growth driven mainly by volume, price, or new businesses?7/10

  A

  It can almost certainly double—and very likely far more than double, which is the least doubtful link in the entire Nebius investment story. The real suspense is not "can it double" but "the quality of the doubling": growth is driven mainly by "volume" (laying down capacity, bringing power online), "price" is a phased tailwind, and "new businesses" are a marginal supplement; whether this "swapping capital for revenue" growth can convert simultaneously into real profit and cash flow is where the report keeps placing question marks.

  First, "can it double"—the evidence is overwhelming. Nebius Q1 2026 AI Cloud revenue was about $390 million, up 841% year over year and 82% quarter over quarter; the company gives full-year 2026 revenue guidance of $3.0-3.4 billion and an ARR (annualized run-rate) target of $7-9 billion. From 2025 (about $530 million) to the 2026 guidance midpoint (about $3.2 billion) alone is a jump of more than 5x. In the intrinsic-value scenarios the report cites, even the most conservative case assumes 2030 revenue of about $6.5 billion, the neutral case $9-10 billion, and the optimistic case $14-18 billion—any of which is several times growth relative to today. So "revenue at least doubling over the next five years" is not a hard threshold but a default that will happen. In the report's investment checklist, the line "does it have long-term stable demand" is marked "Pass."

  Next, the three-factor breakdown of growth, which is the key to judging quality:

  The first driver is "volume"—laying down capacity. The essence of growth is bringing power, land, data centers, and GPUs online batch by batch, then maxing out utilization. In Q1 2026, contracted power already exceeded 3.5GW, owned capacity was over 75%, and the year-end contracted-power target was raised to more than 4GW, while it also announced an owned AI factory of up to 1.2GW in Pennsylvania and a 310MW site in Finland. Doubling revenue essentially means these megawatts come online year by year and the corresponding compute gets rented out—this is a classic "volume driver."

  The second is "price," but it is a phased tailwind, not structural pricing power. In its shareholder letter management explicitly stated "pricing for the newest GPU generation continues to rise, and prior-generation chips also received price support," and Q1 AI Cloud adjusted EBITDA margin reached 45%, indeed reflecting the operating leverage that appears once utilization rises. But the report (and I agree) attributes this to "a joint result of supply-demand mismatch and rising utilization," not sustainable pricing power—once GPU supply loosens and customers can more easily build in-house or shop across vendors, the price tailwind could reverse. So "price" is a boost to the revenue doubling, not a reliable long-term pillar.

  The third is "new businesses," currently a marginal item rather than an engine. During the year the company acquired Tavily and Eigen AI and obtained the Clarifai team and IP license, strengthening the inference and agentic platform (Token Factory); Avride (autonomous driving), TripleTen (education), and others remain, but at very small scale. These are more about completing the technology stack and increasing AI-cloud stickiness, and do not constitute an independent source of revenue doubling in the near term.

  My judgment: a five-year revenue doubling is "almost certain," but it is a "capital-fed, volume-driven growth machine"—Q1 2026 single-quarter capex was as high as about $2.473 billion, far exceeding that quarter's revenue. What Baillie Gifford truly cares about is not how steep the revenue curve is, but whether this growth is "regenerative"—whether it can increasingly self-fund as it grows. Nebius today is still "the more it grows, the more capital it eats," and while revenue doubles, equity is also being continuously diluted. So my answer to this question is: the doubling is a done deal and is mainly volume-driven, but the "gold content" of the doubling—whether it can simultaneously convert into real owner earnings—is the true decider of the whole piece, and is not yet validated.

  Jun 11, 2026
• QFive years from now, what will take over as the next growth engine? Does this "second curve" exist today?4/10

  A

  Frankly: Nebius's problem today is not "where is the second curve" but "the first curve itself has not yet run steadily"—it is still in the phase of taking its core AI cloud from cash-burning to reliably self-funding, so talk of a successor is premature. But if one must answer "what will take over five years from now," the most realistic candidate is not some brand-new business but an extension of the main curve up the value chain: upgrading from "selling raw compute" to "selling an inference/agentic platform and a software layer." This line shows an embryo today but is far from taking shape as an independent engine.

  First, to clarify the status quo: Nebius's revenue is already highly converged on a single curve. Of Q1 2026 group revenue of $399 million, Nebius AI Cloud accounts for about $390 million, i.e. about 98%. The report also makes clear: from 2025 Toloka has been deconsolidated and the data advantage is no longer fully internal, while Avride (autonomous driving) and TripleTen (education) remain but at very small scale. So Nebius today walks on essentially "one leg," and that leg is AI cloud infrastructure. In this structure, the Baillie Gifford-style "second curve" question must first take a discount—a company whose main curve has not yet proven owner earnings should focus on getting the first curve to work, not on distracting itself with a second growth pole.

  So what is the visible "successor candidate"? I judge it is a deepening of the main curve into the software/platform layer, rather than a lateral new business:

  First, the inference and agentic platform (Token Factory) + software layer. This is the direction management has bet on most clearly. During the year the company announced the acquisition of Eigen AI and obtained the Clarifai team and IP license, adding to the earlier acquisition of Tavily; the intent is clear: to climb from "GPU rental billed by the card" up to "inference and agentic services billed by token/by call." If successful, this would be a layer with higher gross margin, stronger stickiness, and closer to "software economics" than raw compute—that is the "second curve" truly worth anticipating, because it can improve unit economics and reduce dependence on hardware generations. But today it is still only "a feature/product layer thickened within the main cloud," with no separately disclosed revenue, far from an independent engine.

  Second, the cost/scale regenerativeness that owned capacity itself brings. Owned capacity over 75%, the 1.2GW owned factory in Pennsylvania, and the 310MW site in Finland could over the long term give Nebius a structural advantage in unit compute cost. But strictly speaking this "makes the first curve stronger," not a new curve.

  As for non-cloud assets like Avride and TripleTen, the report's framing is that they are more like "stock to be disposed of or marginalized," and treating them as a second growth engine is unrealistic.

  Let me add an implicit premise Baillie Gifford truly cares about—does this second curve "exist today": the answer is "the embryo exists, the scale does not." The direction (moving up to inference/software layer) is visible, and M&A is paving the way, but no independent, quantifiable second revenue pillar has yet emerged. This is fundamentally different from mature growth stocks whose "main business self-funds and whose second curve already contributes significant revenue."

  My conclusion: Nebius's successor five years out is most likely an upward move of the main curve along the value chain (inference/agentic/software layer), rather than a new business started from scratch; this line today has only an embryo, is pieced together through M&A, and has not yet proven it can independently shoulder growth. For a company that has not even validated the owner earnings of its first curve, the "second curve" should be viewed more as an option on "whether the main business can upgrade into higher-quality profit," rather than growth insurance already in hand. This is precisely one of the underlying reasons the report maintains "watch, do not chase."

  Jun 11, 2026
• QWhat is its core competitive advantage? Will this moat widen or narrow over the next three to five years?3/10

  A

  Nebius has real "entry barriers," but not yet a real "durable moat"—its advantage today looks more like a bridge-building-phase barrier stacked from "first-mover + execution + supply constraint," not a structural moat it can collect rent on for ten years. As for whether the moat widens or narrows over the next three to five years, my judgment is: uncertain, and the direction hinges on one thing—whether it can upgrade "selling raw compute" into a "sticky platform/software layer"; if it cannot, the moat will most likely narrow as GPU supply loosens. The report scores the moat 2/5, and I agree with this on-the-low-side framing.

  First, what exactly is its core competitive advantage—these are real, requiring money and time to replicate:

  First, the scarce-resource positioning of power, land, and capacity. In Q1 2026 contracted power already exceeded 3.5GW, owned capacity was over 75%, and the year-end contracted-power target was raised to more than 4GW, locking in a 1.2GW owned AI factory in Pennsylvania and a 310MW site in Finland. In an environment where power is the hardest current constraint, this is a real barrier.

  Second, deep binding with NVIDIA. The company obtained NVIDIA Exemplar Cloud (GB300 NVL72 training) qualification, is one of the few cloud providers to hold that certification across multiple GPU generations, and NVIDIA also made a $2 billion strategic investment. This lets it get the newest hardware earlier and gives it a voice in hardware-software co-design.

  Third, proprietary server/rack/data-center design + engineering culture. The report rates "culture/engineering capability" medium-to-strong and sees it as where management is most worth paying up for; the high proportion of owned sites may also bring long-term cost advantages.

  But from a dual Baillie Gifford/Buffett perspective, why do these still not count as a "moat"? The report dissects it clearly item by item: brand weak-to-medium, network effects weak, switching costs medium-to-weak, patent/regulatory barriers limited, the data advantage no longer internal since Toloka was deconsolidated, capital-allocation capability "not yet proven," and cost and scale advantages "not yet validated in long-term ROIC." The core crux is threefold:

  One is that the rivals are among the strongest, not small players—AWS, Azure, Google, Oracle, plus the most direct, CoreWeave (Q1 2026 single-quarter revenue already reached $2.08 billion, about 5x Nebius's size). CoreWeave itself proves that on this track, capital and orders can concentrate onto a handful of leaders at extremely high speed—others "can do it."

  Two is strong customer bargaining power. Buyers are mostly technically formidable model companies and inference providers, and once GPU supply loosens, pricing power may recede—the report judges today's 45% AI Cloud adjusted EBITDA margin to be the result of "supply-demand mismatch + rising utilization," not sustainable pricing power.

  Three is extremely fast hardware generational iteration. Today's scarce GPUs and rental prices do not necessarily equal stable high returns over the next ten years; a moat built only on "this generation of hardware being scarce" gets diluted as the next generation of supply arrives.

  So, widen or narrow over three to five years? The key variable is software/platform stickiness. If Nebius can deeply bind customer workloads through Token Factory, the inference/agentic platform, migration toolchains, governance, and data flows, raising switching costs, the moat has a chance to upgrade from "supply-constraint type" to "stickiness type" and widen; conversely, if it stays stuck in "raw-compute rental competed on performance/price," then as supply loosens and giants expand capacity, the moat will narrow—the most dangerous scenario being exactly the one the report names: "the fixed assets that look scarce today become expensive but mediocre-return assets a few years later."

  My conclusion: Nebius's moat is currently in a "bridge-building phase" rather than a "rent-collecting phase," strength 2/5. It has real barriers (power positioning + NVIDIA binding + engineering capability), but lacks the sources of sustainable excess returns like brand, network effects, and switching costs. The next three to five years are undetermined in direction, with downside risk skewed larger—this is exactly why the report marks both "does it have a durable moat" and "does it have pricing power" as "uncertain" and, on that basis, maintains a "watch" rating.

  Jun 11, 2026
• QIf the core business is disrupted, does it have the genes to reinvent itself? How does it treat mistakes and bad news?4/10

  A

  This is the dimension where the evidence on Nebius is "thinnest," and it must be answered honestly in two layers: one is "when the core business is disrupted, does it have the genes to reinvent itself," and the other is "how does it treat mistakes and bad news." The former can only get a neutral conclusion of "there are signs, untested"; the latter has a real and not-quite-positive ready sample—the material internal-control weakness.

  First, the self-reinvention genes—the implicit premise being: if one day the core business of GPU leasing is disrupted (oversupply, giants building in-house swallowing outsourcing demand, or the compute paradigm rewritten by new hardware), does this company have the underlying ability to "switch to another life and keep living"?

  The evidence supporting "it has the genes" comes mainly from its "parent bloodline" and its "initiative in moving up the value chain." Nebius's predecessor spun out of the Yandex system, and the team has a history of building large-scale technology stacks from scratch; the report also rates "culture/engineering capability" medium-to-strong and sees it as where management is most worth paying up for. A more substantive sign is that it has not clung to raw compute but has proactively climbed higher: during the year it acquired Tavily, announced the acquisition of Eigen AI, and obtained the Clarifai team and IP license, extending its capabilities up to the inference and agentic platform layer. This "don't wait for the core to be disrupted, proactively add a layer first" move is itself an expression of the will to reinvent.

  But cold water must be poured: these genes have never been truly tested by adversity. Its standalone record from spin-off to independent operation is very short, and the entire AI cloud revenue was built during a tailwind of extreme GPU scarcity, customers queuing to prepay, and wide-open financing windows. Real "self-reinvention genes" can only be seen in a downturn of loosening supply, shrinking orders, and pricier financing—and Nebius has yet to go through such a "life-or-death trial." So the honest answer at this layer is: it has the will to reinvent and an engineering foundation, but the disruption-resistant, life-switching genes have not been proven by any real crisis.

  Now, how it treats mistakes and bad news—here there is a ready sample that cannot be whitewashed.

  The hardest item is governance and accounting: in its FY2025 annual report Nebius disclosed two material internal-control weaknesses—fixed-asset accounting and TripleTen revenue recognition—and said remediation would be completed before the end of 2026. This item must be viewed from two sides: on the positive side, the company did publicly disclose the weaknesses and gave a remediation timetable, hiding nothing; on the negative side, a "material weakness" itself shows that rapid expansion has already strained control systems, and such issues need continued watchfulness because "dragging on unresolved leads to bigger problems"—the report indeed lists it as a clear risk item and one of the signals that "trigger reassessment."

  Another item is more subtle: from 2026 the company changed the depreciation life of servers and network equipment from 4 years to 5 years, applied prospectively. This is not fraud, but the depreciation convention happens to be the most sensitive accounting item for a capital-intensive company, with the most room for management judgment—when profit is under pressure, extending the depreciation life dilutes current-period depreciation and flatters the books, so the report groups it, together with revaluation gains propping up profit, under "accounting quality requires continued vigilance." In other words, faced with a "bad-news-style cost," management chose the accounting treatment more friendly to profit, which at least reminds us: between "presenting truthfully vs. making the numbers look good," it does not unconditionally lean toward the former.

  A relatively positive point is that management is decisive in correcting "strategic mistakes": at the end of 2024 it explicitly abandoned the share buyback it might have pursued and instead deployed capital into AI infrastructure, acknowledging that hoarding capacity in that window created more value than a buyback—showing it can change its mind quickly according to the environment rather than stubbornly holding old decisions. But this is the flexibility of "strategic timing," a different matter from "candor when facing its own mistakes and bad news."

  My conclusion: self-reinvention genes—there is will and an engineering foundation, but untested by real adversity, so it can only be viewed neutrally; treatment of mistakes and bad news—there is a side of disclosure and course correction, but the two items of a material control weakness + loosening the depreciation convention leave a solid warning on the "candor vs. flattery" scale. This is also the direct basis for the report marking "is management trustworthy" as "uncertain" and stressing that "accounting quality requires continued vigilance." On this dimension, giving Nebius too much credit would be dishonest.

  Jun 11, 2026
• QDoes management (especially the founder) have a long-term vision with interests deeply bound to the company? Are they willing to sacrifice current profit for five to ten years out?6/10

  A

  This is a relatively positive but "pros-and-cons coexist" dimension for Nebius: founder Arkady Volozh does have a long-term vision, extremely strong control, and a clear willingness to sacrifice current profit for five to ten years out—all three hold; but "interests deeply bound to the company" must take a discount, because his economic interest is far smaller than his voting power, governance is not a fully aligned "one-share, one-vote," and on top of that the standalone record is too short to support the kind of "trust the founder blindly" premium Baillie Gifford requires. The report scores "management and capital allocation" 3/5, and I agree.

  Long-term vision and control—the evidence is ample and on the positive side. Volozh is the founder who took the company all the way from the Yandex system to an independent AI cloud, personally signs the shareholder letters, and clearly expresses the long-term intent of "not simply responding to today's industry, but building infrastructure for its tomorrow." On control, as of March 31, 2026, he held about 11% of the economic interest through a family trust yet controlled about 52% of the voting power, with Class B shares carrying 10 votes each (the prior 20-F basis the report cites was 12.94% economic interest / 55.04% voting power, the difference arising from the different reporting period and the dilution of his stake from large issuances during the interval). This structure gives the founder overwhelming say over the long-term direction, so he will not be forced to pivot by short-term shareholder pressure—an advantage for a business that must endure years of losses to lay down capacity.

  Willingness to sacrifice current profit for the long term—the answer is a very clear "yes," even aggressively so. Several hard pieces of evidence: first, at the end of 2024 the company explicitly abandoned the share buyback it might have pursued, on the grounds that deploying capital into AI infrastructure creates more value for shareholders now—proactively sacrificing the short-term act of "returning capital to shareholders" to grab capacity. Second, it pays no dividends at all, putting almost all cash into reinvestment, capacity expansion, and completing the technology stack. Third, it dares to swap very heavy capital expenditure for future scale—Q1 2026 single-quarter capex was about $2.473 billion, far exceeding that quarter's revenue of $399 million. From the Baillie Gifford-preferred angle of "sacrificing the present for five to ten years out," Volozh's behavior fully conforms, even betting more boldly than most founders.

  But the "interests deeply bound" layer must be honestly discounted—this is the key two-sidedness. First, economic interest and voting power severely diverge: about 11% economic interest against about 52% voting power means the founder bears only about one-fifth of the "wallet pain of a falling share price" relative to the "decision power" he holds. The dual-class structure that the Baillie Gifford/Buffett framework praises presupposes "the founder is also the largest economic beneficiary, in the same boat as you"; when economic exposure is far below control, the credibility of "I'll make long-term decisions on your behalf" takes a discount—hence the report stresses that such a structure "is worth a premium only when management is highly credible and capital allocation is excellent over the long term, and Nebius does not yet have a long enough standalone record to enjoy that governance premium."

  Second, the flip side of "daring to sacrifice the present for the long term" is that the cost is borne by all shareholders together. To grab the window, the company completed a $1 billion public offering + $2.75 billion of convertibles in 2025, and raised a further $6.3 billion in Q1 2026 through NVIDIA prepaid warrants and convertibles. The founder's boldness of "diluting freely for the long term" corresponds to minority shareholders' per-share interest being solidly diluted—and because the founder's economic interest is inherently low, the "pain" of dilution for him is relatively smaller. The report separately flags "is capital allocation rational" and "can per-share value grow," and lists continued dilution as a fact that would overturn the investment thesis, precisely out of this concern.

  My conclusion: on the founder dimension, Nebius earns positive marks for "long-term vision + strong control + daring to sacrifice the present for the long term" (a relative bright spot distinguishing it from the likes of CoreWeave), but is dragged back to "uncertain" by "economic interest far below voting power, non-fully-aligned governance, a short standalone record, and the cost of long-term bets shared by all shareholders via dilution." What Baillie Gifford truly asks is "whether the founder is in the same boat as you, and whether the boat is sailing the right way"—Volozh most likely has the direction right and is indeed steering, but the "own money" he has staked on the boat is relatively small compared with his "steering power," which discounts the phrase "deeply bound." The composite is the report's 3/5.

  Jun 11, 2026
• QIf it disappeared tomorrow, how much would customers miss it? Is its growth model sustainable and not reliant on harming society and regulation?4/10

  A

  This question must be split into two independent sub-questions and answered honestly: one is "indispensability"—if Nebius disappeared tomorrow, how much would customers miss it; the other is the "social/regulatory sustainability of its growth model"—is its expansion not reliant on harming society and regulation. My conclusion: customers would be "inconvenienced" but not "fatally longing"—it is highly replaceable; and its growth model does not itself harm society, but it depends deeply on external permits such as power and export controls, so its sustainability carries real policy uncertainty. Taken together, on Baillie Gifford's dual threshold of "indispensable + sustainable," Nebius only partly clears both.

  First layer, indispensability—on the weak side. If Nebius disappeared tomorrow, customers would of course be affected: those large customers that signed multi-year dedicated capacity with it especially would face migration costs. The company has signed a multi-year dedicated GPU contract with Microsoft (in Q1 2026 it said the capacity commitment to Microsoft had all been delivered, with future volume ramping in Q3/Q4), and signed a long-term supply agreement with Meta of up to about $27 billion in scale (of which $12 billion of dedicated capacity has been specified) (note: after comparing the report body's "$12 billion / $27 billion total" with the primary shareholder letter's "up to $27B / $12B dedicated," the shareholder-letter figures should prevail; this is likely a unit typo in the report).

  But the degree of "missing" it is suppressed by three things: first, what it sells is relatively standardized compute—NVIDIA GPU clusters + on-demand compute + platform, and customers can switch at any time to AWS, Azure, Google, Oracle, or the most direct peer CoreWeave (already about 5x Nebius's size, Q1 2026 single-quarter revenue $2.08 billion). Second, customers are mostly extremely capable model companies/inference providers, strong at both bargaining and building in-house, which is why the report rates switching costs "medium-to-weak" and network effects "weak." Third, the biggest customers are themselves hyperscalers (Microsoft, Meta)—they use Nebius to supplement elastic capacity rather than depend on it irreplaceably; it is instead Nebius's dependence on these large customers that is rising (the report lists customer concentration as a risk item). So the honest answer is: customers would feel "the pain of finding a replacement," but would not "be unable to live without it"—a fundamental gap from those indispensable businesses whose "disappearance would truly paralyze customers."

  Second layer, social/regulatory sustainability—the business itself is legitimate, but dependence on external permits is high and policy risk is real.

  On the positive side, Nebius's growth model is not built on "harming society": it does not make money through regulatory arbitrage, data abuse, or harming users, and instead has earned enterprise-security compliance certifications such as SOC 2 Type II and ISO 27001, serving demand for AI training/inference that is widely seen as a productivity tool. On the Baillie Gifford sub-question of "whether it trades growth for harm to society," it basically passes.

  But the question mark on "sustainable" is the deep dependence on external permits:

  First, power. Its entire expansion is essentially turning megawatt-scale power into compute—contracted power already exceeds 3.5GW, the year-end target is over 4GW, and owned capacity is over 75%. Data-center power consumption is already a global-scale regulatory and social issue, and power permits, grid connection, community, and environmental constraints could all change—this is a regulatory risk the report explicitly lists. The faster it grows and the more power it consumes, the more its sustainability depends on whether it can keep securing power and social license.

  Second, export controls and data/security compliance. High-end GPUs are subject to U.S. export controls, and a change in geopolitics could affect its supply access and cross-region deployment; this is also one of the regulatory risks the report names. In other words, its growth sustainability is not fully in its own hands, but partly tied to the policy environment.

  My conclusion: on the indispensability dimension, Nebius is on the weak side—customers would be "inconvenienced" rather than "fatally longing," because it sells replaceable standardized compute and its dependence on large customers is instead deepening; on the social/regulatory sustainability dimension, the business is legitimate and does not rely on harming society, but it depends deeply on external permits such as power and export controls, carrying real policy uncertainty. The two together correspond exactly to the report rating switching costs/network effects weak, listing customer concentration and regulation separately as risks, and maintaining "watch"—on Baillie Gifford's most valued sustainability question of "how much customers can't do without you, and whether your growth can keep going," Nebius has not yet earned a high score.

  Jun 11, 2026
• QWhat are this business's unit economics (gross margin, incremental returns)? Do they get better or worse as scale grows? Where does the money earned go?3/10

  A

  Unit economics is the dimension where Nebius has "real bright spots, but is still unproven overall," and must be seen in three layers: at the operating layer, unit economics are "getting better with scale" (utilization rising, operating leverage real), but at the capital layer, unit economics are still "very poor" (behind every dollar of revenue sits capex far exceeding revenue), and the money earned and raised goes almost entirely back into expansion—so it is a business "whose operating profitability is improving, but which is still far from converting scale into real owner earnings." The report marks both "can it generate stable free cash flow" and "is its return on capital excellent" as "Fail," which is honest.

  First layer, operating-level unit economics—improving, this is a real bright spot. As utilization rises, operating leverage is indeed showing: the core Nebius AI Cloud's Q1 2026 adjusted EBITDA margin reached 45%, nearly doubling from the prior quarter, and management said it is heading toward a 20-30% EBIT margin. Combined with rising pricing for the newest GPU generation, price support for prior-generation chips, and the average deal size growing with pricing and longer contract terms, this shows that "as scale grows, the margin at the operating level is indeed getting better." This is the most concrete evidence for the bulls, and the report also lists it as one of the core bull-case reasons.

  Second layer, capital-level unit economics—still very poor, this is the crux. The problem is that the 45% above is on an "adjusted EBITDA" basis, before depreciation, interest, and the enormous expansion capex. On a full basis, the picture reverses: Q1 2026 single-quarter capex was about $2.473 billion, far exceeding that quarter's revenue of $399 million. In other words, for every dollar of revenue made today, far more than a dollar of capital must be poured in. The report calculates it clearly—if all capex is treated as cash flow to be deducted, that quarter's free cash flow is about -$215 million; even on the more conservative estimate of "stripping out the surging deferred revenue + using the quarter's depreciation and amortization of about $212 million as a maintenance-capex proxy," conservative owner earnings are about -$1.16 billion per quarter. In Baillie Gifford's language, this company is currently "the more it grows, the more capital it eats," incremental returns (the real return on invested capital) are not yet validated, and whether ROIC is excellent is even less discussable.

  Third layer, the "quality" of gross margin and cash—beware two dilutive items. One is that "ARR" is not a SaaS-style locked remaining contract value, but a run-rate of "the last month of the quarter's revenue ×12"; repeatability is strengthening, but predictability is still strongly tied to capacity coming online, utilization, and pricing. Two is that profit and cash-flow quality are flattered by one-off items: of Q1 2026 net income from continuing operations of $621.2 million, about $780.6 million came from revaluation of non-listed equity investments (ClickHouse's January financing at a valuation of about $15 billion brought about $781 million of non-cash, non-operating gains); meanwhile Q1's operating cash flow of about $2.258 billion was largely supported by customer prepayments from deferred revenue soaring from $1.578 billion to $4.778 billion, an increment of about $3.201 billion. Prepayments are a good thing (reducing financing needs), but they look more like "operating financing" and cannot be equated with cash already freely distributable. So the "real quality" of unit economics is quite a bit uglier than book profit and operating cash flow suggest.

  Where does the money earned go—almost entirely back into expansion, no dividends. The company uses cash mainly for reinvestment, expanding sites, raising contracted power, and completing its technology stack (during the year acquiring Tavily and Eigen AI and obtaining the Clarifai team and IP), and keeps relying on debt, an ATM program, and customer prepayments to support expansion. It is normal for an early-stage growth company to do this, but for a value framework it means "cash-distribution rights currently belong to the future, not the present."

  Positioning its unit economics against peers: CoreWeave's Q1 2026 adjusted EBITDA margin reached 56%, but it still had a net loss of $740 million, net interest expense of $536 million, and single-quarter capex of $7.695 billion—showing that "high adjusted EBITDA margin + still loss-making on a full basis + heavy capex" is common to the whole neocloud track, not unique to Nebius. Meanwhile a truly cash-flow-mature cloud company like DigitalOcean, with Q1 2026 revenue of $258 million, adjusted EBITDA of $105 million (about 41% margin), can still generate positive adjusted free cash flow—by contrast, Nebius's unit economics are already close at the "operating layer" but clearly lag on "whether it can convert scale into free cash flow."

  My conclusion: on unit economics, Nebius's "operating-layer margin improving with scale" (45% AI Cloud adjusted EBITDA, heading toward 20-30% EBIT) is a real bright spot, but its "capital-layer unit economics are still very poor" (capex far exceeding revenue, conservative owner earnings negative), and profit/cash quality is significantly flattered by revaluation gains and customer prepayments. The money earned all goes back into expansion, no dividends. The composite is the report's judgment: operating leverage is real, but the two gates of "stable free cash flow" and "excellent return on capital" have both not yet been passed today.

  Jun 11, 2026
• QFor it to rise fivefold in ten years, what conditions must hold simultaneously? Are these conditions realistic? What expectations does today's share price imply?3/10

  A

  For Nebius to rise fivefold in ten years (answered here by the Baillie Gifford LTGG standard proposition; the report itself gives more restrained "conservative/neutral/optimistic" intrinsic-value and annualized-return ranges) requires a long list of high-difficulty conditions to hold simultaneously, and today's price of about $211.69 has already priced in front a considerable part of these optimistic assumptions—which is precisely the core numerical basis for the report maintaining "watch, do not chase." I will first list "what conditions must hold simultaneously," then explain "what today's price implies."

  First, anchor today's starting point. As of the close on June 10, 2026, NBIS was about $211.69, market cap about $54.2 billion, about 256 million shares, PE about 66x, up about 160% year to date, with a 52-week range of $43.89-278.84. The report's snapshot, done at about $219.94 and about $55.8 billion market cap, is basically consistent with this (the current price is slightly lower).

  For a ten-year fivefold, the following conditions almost all must hold simultaneously:

  First, revenue sustains high growth for years without stalling. The report's optimistic scenario assumes 2030 revenue reaches $14-18 billion (about 5x the 2026 guidance midpoint of about $3.2 billion) and keeps growing thereafter—which requires that the 841% YoY AI Cloud growth of Q1 2026, though it must inevitably decelerate, still significantly outpaces the industry over the long term.

  Second, the operating margin truly climbs to a high level and holds. It needs to go from today's 45% AI Cloud adjusted EBITDA margin to deliver the 20-30% EBIT management talks about, even the 20-25% owner-earnings margin the report's optimistic case assumes—and hold it even after GPU supply loosens.

  Third, the moat upgrades from "supply-constraint type" to "stickiness type". Today the moat is only 2/5 and pricing power is "uncertain"; a ten-year fivefold requires it to truly bind customers with the platform/software layer (Token Factory, inference/agentic), or else the story collapses once supply loosens and pricing power recedes.

  Fourth, capex intensity comes down, and free cash flow turns positive and persists. Today it is "the more it grows, the more capital it eats"—Q1 single-quarter capex was about $2.473 billion, far exceeding that quarter's revenue; the fivefold story requires it to eventually switch from "burning cash to build plants" to "steadily spitting out cash," converting scale into real owner earnings.

  Fifth, large-customer contracts are delivered on schedule and profitably as expected. Microsoft's capacity commitment has been delivered, with future volume ramping in Q3/Q4, and the Meta agreement is up to about $27 billion (with $12 billion of dedicated capacity specified)—these must turn into real, profitable revenue, not just "paper backlog."

  Sixth, dilution is controlled, and per-share value can compound. The company issued $1 billion in equity + $2.75 billion in convertibles in 2025, and in Q1 2026 raised a further $6.3 billion via NVIDIA prepaid warrants and convertibles; a fivefold requires no more large financings that "keep diluting per-share economic interest," or else the company's scale rises fivefold while per-share value may not.

  Seventh, the external environment cooperates: no severe GPU oversupply, export controls/power permits not choking it, and interest rates and financing conditions staying manageable.

  Are these conditions realistic? Each on its own is "possible," but for them all to happen simultaneously is very hard—any single link failing (growth stalling, utilization/pricing declining, the moat not built, capex not coming down, contract profitability falling short, another large dilution) breaks the fivefold logic. This is also why the report marks multiple lines in the checklist—moat, pricing power, free cash flow, return on capital, valuation, margin of safety—as "uncertain/fail": most of the puzzle pieces the growth story needs are not yet in place today.

  What expectations does today's price imply? This is the most critical question. The report's three-tier DCF/owner earnings gives: conservative intrinsic value $30-70, neutral $80-150, optimistic $160-240/share. The current price of about $211.69 already falls within the optimistic scenario range—meaning buying today is not "buying a good business at a cheap price," but "buying at a relatively expensive price a future that still needs a lot of execution to realize." In other words, the current market price basically already assumes most of that string of high-difficulty conditions will be delivered; a fivefold on top of this already-front-loaded starting point amounts to "optimism on top of optimism." The report therefore sets the ideal buy price at $80-110, the acceptable tracking range at $110-150, and states outright that above $180 is "clearly prepaying for a lot of optimistic expectations" and around $220 looks more like "a good story paired with an expensive price." The restraint of sell-side consensus echoes this—the average target price across institutions is about $239, with a range from $120 to $380, showing huge disagreement and a thin margin of safety.

  My conclusion: a ten-year fivefold requires a long string of conditions—"high growth without stalling + high-margin delivery + moat upgrade + capex decline turning out free cash flow + profitable large-customer delivery + controlled dilution + no external chokepoints"—to hold simultaneously; each is feasible on its own, all together is very hard; and today's price of about $211.69 already implies fairly full optimistic expectations, falls within the report's optimistic valuation range, and offers almost no margin of safety. This is why the report's conclusion is "watch first, wait for a better price or more solid cash-flow evidence," rather than betting on a fivefold at the current level.

  Jun 11, 2026
• QWhy hasn't the market realized all this yet? Is it can't-understand, can't-respect, or can't-see-far? What will become the "narrative inflection point"?3/10

  A

  For Nebius this question must be answered in reverse: Baillie Gifford's "why hasn't the market realized this" is usually looking for "an undervalued great company," but Nebius's situation is exactly the opposite—the market has not only realized it, but has most likely already "over-realized" it. It is not a "gathering-dust, undervalued" stock, but a high-expectation stock whose "narrative is fully or even excessively priced." So among the diagnoses of "can't understand / can't respect / can't see far," the most fitting is instead that the market "sees too far, prepaying for long-term optimism," rather than "cannot see the value."

  First, use evidence to prove "the market has not ignored it." A stock neglected by the market would not perform like this: NBIS is up about 160% year to date, with a 52-week range from $43.89 all the way to $278.84, currently about $211.69, PE about 66x; sell-side coverage is dense and bull-tilted—multiple institutions give a "buy" consensus, with an average target price of about $239. The report's three-tier valuation shows the current price of about $211.69 already falls within the "optimistic scenario" ($160-240) range, with no margin of safety against the neutral or conservative valuations. All this shows the market has long chased it as a core name of the AI infrastructure wave, and the perception gap points toward "overvaluation" rather than "undervaluation."

  Now, check each of Baillie Gifford's three "market misjudgments":

  "Can't understand"—partly holds, but not the source of undervaluation. This business is indeed not simple: revenue recognition, customer prepayments, the cadence of capacity deployment, GPU generational transitions, depreciation lives, and enormous growth capex make the statements far more complex than a consumer or software company, and the report scores business understandability only 3/5. There are two real "hard-to-see-clearly" points: of Q1 2026 net income of $621.2 million, about $780.6 million came from non-cash equity revaluation, and operating cash flow of about $2.258 billion was largely supported by customer prepayments from a deferred-revenue surge of about $3.201 billion—many may not see through that "book profit and operating cash flow are flattered by one-off items and prepayments." But note: if the market sees through this "can't understand," the conclusion is "quality is not that good, it should be cheaper," pointing to a downward correction rather than "the market undervalued it, it should rise." So even if "can't understand" exists, it does not constitute a Baillie Gifford-sense bullish perception gap.

  "Can't respect"—does not hold. "Can't respect" means the market slights a company that is actually excellent. Nebius is exactly the opposite—it is highly valued: it has secured multi-year contracts with Microsoft and Meta (up to about $27 billion), a $2 billion NVIDIA investment, and Exemplar Cloud qualification, and raised $6.3 billion in Q1 alone, with contracted power over 3.5GW—capital and large customers are all rushing to endorse it, hardly a case of being slighted.

  "Can't see far"—this is the most fitting diagnosis, but the direction is opposite to Baillie Gifford's expectation. The market is indeed "seeing far," the problem being that it "sees too optimistically, pricing in long-term delivery ahead of time." About $211.69 falling within the optimistic valuation range means the current price already assumes the company can deliver high-difficulty long-term visions like 2030 revenue of $14-18 billion and an owner-earnings margin of 20-25%. In the Baillie Gifford context, "can't see far" originally means the market underestimates long-term potential; but Nebius is a case of the market "seeing too far"—trading a "long-term call option that still needs enormous capital and sustained execution to realize" at a price close to "already succeeded." So if a perception gap exists, it is more likely that "the market is being too generous about the long term."

  So "what will become the narrative inflection point"? This is the implicit premise of this question and the thing most worth watching. The inflection point will not come from "the market suddenly discovering it is undervalued," but is more likely to be the narrative being revalued from an "AI infrastructure high-growth platform" to a "high-capex, mediocre-cash-flow-quality capital-intensive supplier." The trigger signals the report points out are almost all "downward inflections":

  Downward inflection (more to watch out for): guidance cuts or consecutive ARR stalls; deferred revenue/customer prepayments no longer growing while capex stays high (showing growth cannot self-sustain); clear declines in utilization or pricing (supply loosening, phased bargaining power vanishing); Microsoft/Meta and other large-customer contracts delayed, cut, or falling short on profitability; new material control weaknesses, major asset impairments, or more aggressive accounting treatment; another large equity financing that keeps diluting per-share economic interest. Any one confirmed could pull the stock back from a "high-growth-expectation stock"—the report even judges that a 50%-80% drawdown from the current price is not a fantasy.

  Upward inflection (only counts if real substance is delivered)—only when the company truly converts "high growth + scarce capacity + capital-market support" into "stable, distributable, sustainable owner earnings" (capex intensity coming down, free cash flow turning positive, margins holding, dilution stopping) will the narrative upgrade from "option" to "validated compounding machine," and only then will the high valuation have solid support.

  My conclusion: the market does not "can't understand/can't respect" Nebius, but instead "sees far and leans optimistic," treating it as a core AI infrastructure name that is fully or even excessively priced—the perception gap points toward overvaluation, not undervaluation. The true narrative inflection point is more likely a "downward disenchantment" (growth stalling, prepayments cutting off, utilization/pricing declining, a contract or accounting blowup, re-dilution) than an "upward rediscovery"; an upward inflection must be earned through real owner-earnings delivery. This is fully consistent with the report's final conclusion of "watch first, do not chase, wait for a better price or more solid cash-flow evidence."

  Jun 11, 2026