Nebius research:
Nebius: The Company Russia Let Go, Now Renting Out the Future to Microsoft and Meta
- 1. Where This Company Actually Came From
- 2. The Megatrend, and Why Nebius Sits Right in the Middle of It
- 3. What a "Neocloud" Actually Is
- 4. The Three Floors of the Building
- 5. The Post-Training Factory
- 6. Letting Someone Else Pay for the Building
- 7. The Quarter, Plainly Stated
- 8. Nebius Against the Hyperscalers
- 9. Nebius Against Its Actual Rival
- 10. The Boring Slide That Matters More Than It Looks
- 11. The Bull Case: A Path to $410
- 12. The Bear Case: Burry's Depreciation Trap
- 13. What to Actually Watch From Here
- 14. The Two Clocks
- 15. What You're Actually Paying
- 16. What Could Go Wrong
- 17. Muffett's Take
1 Where This Company Actually Came From
Every so often a company's backstory does half the analytical work for you, and Nebius is one of those. Rewind to 2018: a cloud computing division quietly gets built inside Yandex, the Russian search giant sometimes called "Russia's Google." It builds GPU clusters, it builds data centers, it does the unglamorous plumbing work of infrastructure. Then in July 2024, in a $5.4 billion transaction, Yandex gets split cleanly in two. The Russian, domestic business stays in Russia. Everything else — including that cloud division — gets carved out, re-domiciled in Amsterdam, and handed to Arkady Volozh, the man who founded Yandex in the first place and who by then had relocated to Israel and taken Israeli citizenship. That international carve-out is Nebius. It listed on the Nasdaq, cleaned up its ownership structure, and started doing something no one particularly expected: building one of the fastest-growing AI infrastructure companies on the planet, eighteen months after essentially starting from scratch as an independent company.
We mention all of this not because we think you should buy or avoid a stock based on its founder's biography, but because it explains a few things that show up later in this note. It explains why Nvidia was an early and enthusiastic backer — a $700 million investment in late 2024, then another $2 billion in March 2026, plus priority early access to Blackwell and the next-generation Vera Rubin chips. It explains why Nebius has been unusually motivated to build clean, Western, audit-friendly infrastructure rather than cut corners. And it explains why, when Microsoft and Meta went looking for AI compute partners who weren't already fully booked by the giants, a scrappy, well-capitalized, newly independent operator with genuine technical pedigree and nothing to hide was an easy company to say yes to.
2 The Megatrend, and Why Nebius Sits Right in the Middle of It
We keep returning to the same idea across our technology coverage: every AI application in the world needs somewhere to actually run, and whoever owns that "somewhere" is sitting on the toll road, not just one of the cars driving on it. That's the megatrend. With DigitalOcean, we told that story from the bottom of the market — the small developer, the scrappy startup, the business that just wants AI to work without hiring a platform team. Nebius is the same story told from the top of the market: the AI labs and the trillion-dollar hyperscalers who need raw, massive, dedicated compute capacity, right now, and are willing to sign contracts worth tens of billions of dollars to lock it in.
There's a second megatrend layered on top of the first, and it's worth naming explicitly: the geopolitics of compute. The West wants AI infrastructure built by companies it can trust, audited by regulators it recognizes, and domiciled somewhere a subpoena actually works. A company that spent its early life inside a Russian conglomerate and then deliberately, formally, expensively extracted itself from that ownership to become a clean, Amsterdam-domiciled, Nasdaq-listed pure-play is about as good a fit for that anxiety as you're going to find. Microsoft and Meta didn't need to explain that part out loud when they signed contracts worth a combined $46 billion or more. It's simply part of why Nebius, rather than some anonymous GPU reseller, got the call.
3 What a "Neocloud" Actually Is
"Neocloud" is one of those words that sounds like marketing until you actually see the picture, and then it clicks immediately.
Picture the hyperscalers — Amazon, Microsoft, Google — as a sprawling, decades-old city that was built for cars and pedestrians and now has to somehow also handle freight trains. It works, mostly, but it's got traffic, detours, and infrastructure fighting itself. A neocloud like Nebius is a city built yesterday, purely for freight, with rail lines running straight through it and nothing else competing for the track. That's the practical difference: no virtualization overhead, no multi-tenant noise, networking built natively around the ultra-low-latency InfiniBand standard that large language model training actually needs. It is a narrower business than what AWS does. It is also, for the specific job of training and running enormous AI models, a faster and cheaper one.
4 The Three Floors of the Building
The part of this business we find genuinely persuasive isn't the raw compute. Raw compute is a commodity, and commodities get their margins ground down eventually, always. What convinced us is what Nebius built on top of the compute.
Picture it as a three-story building. The ground floor is the GPU compute and network — the NVIDIA chips, the bare-metal clusters, the InfiniBand wiring. That's the part everyone assumes is the whole business, and it's the part that eventually turns into a price war, because a chip is a chip no matter whose logo is on the data center. The middle floor is platform services: managed Kubernetes with GPU support built in, a workload manager Nebius built itself called Soperator, databases, storage. The top floor is AI services proper — a "post-training factory" for customizing models, serverless AI endpoints, the tools developers actually touch every day. Here's the thing about owning all three floors instead of renting out just the ground level: customers who move in on the top floor don't want to move out. They've built their workflow around your tools. That's a business you can charge more for, and a customer who's much harder to lose to the next company offering a marginally cheaper GPU.
5 The Post-Training Factory
If there's one slide in the source materials that made us sit up, it's this one, because it's the clearest evidence that Nebius understands exactly what game it needs to be playing.
Renting out a GPU is a commodity business. Taking a customer's raw idea, fine-tuning one of thirty-plus open-source models to fit it, compressing that model down into something three to five times faster to run without losing its reasoning ability, hardening it for real enterprise use cases like structured JSON or SQL output, and deploying it with one click — that's a software business wearing a hardware company's clothes. It's the same instinct we liked in DigitalOcean's shift toward software and managed services, except Nebius is running it at a much larger scale, for much larger customers, with a much more technically demanding product. Every step in that factory is a step where Nebius earns a margin that a company just renting out bare GPUs never will.
6 Letting Someone Else Pay for the Building
Here's the single cleverest thing in this entire deck, and it directly answers the most obvious objection anyone would raise about a company burning $5.7 billion of capital expenditure in a single quarter: where does all that money come from, and who's actually on the hook for it?
Nebius doesn't always build and own every data center itself. In a meaningful and growing share of its footprint, outside developers and regional operators put up the capital for the physical building, the power infrastructure, and the hardware. Nebius brings the software stack, the technical architecture, and — crucially — the customers already lined up to fill the capacity. Everyone gets paid through revenue-sharing and licensing arrangements instead of Nebius footing the entire bill alone. It's the same logic a hotel brand uses when it lets a property developer build the actual hotel: the brand keeps the high-margin part — reservations, loyalty, pricing power — while someone else carries the construction risk. We just watched Nebius announce exactly this kind of arrangement with Vantage Data Centers in South Wales, expanding UK AI capacity, on the very day we're writing this note. It's a real, live pattern, not a slide deck aspiration.
7 The Quarter, Plainly Stated
Strip away the narrative for a moment and just look at what the business actually did in the second quarter of 2026, because the numbers alone are startling enough.
Revenue of $582.3 million for the quarter, up 454% from a year earlier. Almost all of it — 98% — came from the AI cloud business specifically, which grew 514% on its own. Gross margin of 68.6%. Adjusted EBITDA margin of 41%, up from 32% just one quarter earlier, which tells you the business is getting more efficient as it scales, not less, which is exactly the direction you want to see and is not something every fast-growing infrastructure company can claim. Fifty to sixty percent of new capacity is being paid for upfront by customers themselves, which covers the majority of that eye-watering capital expenditure before Nebius even has to reach into its own $8 billion cash pile. And in a detail that says more about pricing power than any paragraph of commentary could, Nebius ran a capacity auction for its newest Blackwell-generation chips during the quarter and got the highest price ever recorded for that hardware — a 15% premium over what it had been charging before. When demand lets you raise prices and customers still show up, that's not hype. That's leverage.
8 Nebius Against the Hyperscalers
It's worth being honest about where Nebius sits relative to the giants, because the comparison cuts both ways.
Oracle and Amazon Web Services both post real, chunky, north-of-30% operating profit margins, because they're enormous, diversified, decades-old businesses with every kind of customer imaginable. Nebius doesn't have that yet — it is still scaling into profitability at the operating line even as its adjusted EBITDA margin expands nicely. What Nebius does have, and what neither of those giants can claim in the same category, is close to pure-play exposure to the fastest-growing part of the entire technology sector, paired with a genuinely conservative, cash-heavy balance sheet rather than the kind of leverage that usually accompanies a young company trying to grow this fast. The honest read is this: Nebius isn't a cheaper Oracle. It's a much riskier, much faster-growing, much more concentrated bet on the same underlying trend Oracle only gets a slice of.
9 Nebius Against Its Actual Rival
The more useful comparison isn't the hyperscalers at all. It's CoreWeave, the other well-known neocloud, because that's the fight investors are actually having in real time.
CoreWeave is bigger today, has more GPUs deployed, a larger contracted backlog, and got there first, which counts for something in a market where being early to a customer relationship matters. But CoreWeave got there by borrowing over $10 billion, and debt that size comes with interest payments and refinancing walls that don't care how good your growth story is. Nebius chose the slower, harder path of building out more conservatively, backed by equity investment from Nvidia and a genuinely large cash balance rather than a mountain of leverage. Both companies offer roughly the same triple-digit growth exposure to the same underlying AI infrastructure boom. Only one of them can survive a bad year without a lender making decisions for it. We know which side of that trade we'd rather be on.
10 The Boring Slide That Matters More Than It Looks
We almost skipped past this one, because "sustainability metrics" is usually the part of a deck where our eyes glaze over. This one is different, because it's not really about sustainability at all. It's about margin and about permits.
A data center that runs more efficiently uses less electricity to do the same job, and electricity is one of the largest recurring costs in this entire business, which is a direct, structural contributor to that 68.6% gross margin we mentioned earlier. But the water number is the one we found genuinely clever: 0.018 liters per kilowatt-hour, against a US data center average of nearly half a liter. In a world where new AI data centers are increasingly running into local opposition over water and power usage — this is a real, growing bottleneck across the industry, not a hypothetical — a facility with almost no meaningful water footprint gets through permitting and zoning meaningfully faster than a thirsty one. Being the easiest data center operator for a town to say yes to is a genuine competitive edge when speed of capacity buildout is exactly what determines whether you can honor a $19 billion Microsoft contract on time.
11 The Bull Case: A Path to $410
Let's lay out the optimistic case as cleanly as it deserves, because it isn't a fantasy — it's built on contracts that are already signed.
Start with demand visibility most companies would kill for: somewhere between $17 billion and $19 billion of contracted business with Microsoft, and a separate $27 billion infrastructure agreement with Meta, together guaranteeing that Nebius's capacity gets used well into 2027 regardless of what the broader economy does in the meantime. Layer on margin expansion — AI cloud margins already at 49.7% and climbing, driven by that Post-Training Factory software layer rather than bare hardware rental. Add the asset-light capital structure that lets Nebius scale past 4 gigawatts of committed capacity without taking on the kind of debt that sank so many previous infrastructure booms. Put it together, and Wall Street analyst Piper Sandler has modeled a path to $25 billion or more of annual recurring revenue by 2028 — nearly ten times where the run-rate stands today. That's the case that took the stock up 34% in a single session after this earnings report. It isn't a story pulled from thin air. It's a story built from contracts that are already signed and margins that are already expanding.
12 The Bear Case: Burry's Depreciation Trap
Now the other side, and we want to give it the same fair hearing, because it comes from a source worth taking seriously. Michael Burry — the investor who famously called the 2008 housing crash — has publicly disclosed a short position against Nebius, built on essentially this exact argument.
The core of it is an accounting argument, and it's a fair one: GPUs realistically wear out and become technologically obsolete within two to three years as newer, faster chips arrive, but standard accounting practice spreads their depreciation cost over five to seven years. If that mismatch is real, then some portion of today's headline 41% EBITDA margin is an accounting illusion — profits today, funded by a hardware replacement bill that hasn't hit the income statement yet but absolutely will. Layer on a valuation that, even using the most forgiving numbers, sits somewhere north of a thousand times trailing earnings with a forward P/E that isn't even positive yet, and you have a stock priced for flawless execution with essentially no room for a stumble. And there's a specific, concrete stumble already on the table: reported regulatory and inspection delays at Nebius's Vineland, New Jersey facility, a site that matters directly to hitting the delivery schedule underneath that $17 billion-plus Microsoft contract. We won't pretend this case is weak. It's the same discipline we'd apply to any high-multiple growth story — the bear case here isn't noise, it's a specific, falsifiable set of things that could actually go wrong, and Burry has put real money behind believing they will. So far, the stock has gone up anyway. That doesn't make him wrong. It makes the argument still open.
13 What to Actually Watch From Here
Rather than treating this as a binary bet you make once and forget, here's the actual sequence of checkpoints that will tell you which case is winning, well before the final answer arrives.
The near-term test is simple: does the 40%-plus adjusted EBITDA margin hold up as capacity keeps scaling, or does it start slipping as the easy wins get used up. The critical mid-term marker is Vineland itself — first power delivery there, on schedule, is the single event that would most directly rebut the bear case's execution-risk argument. By year-end 2026, management is targeting $7 billion to $9 billion of annualized run-rate revenue; landing at the top of that range would validate the aggressive growth trajectory the bulls are pricing in. And further out, full-year 2027 is about proving the asset-light partnership model actually works at scale — deploying more than a gigawatt of new capacity globally without the balance sheet damage that model is specifically designed to avoid. None of these are abstract. Every one of them is a date on a calendar with a clear pass-or-fail outcome attached.
14 The Two Clocks
Strip away everything else in this note and you're left with one genuinely elegant way to think about this stock, and it's the framing we keep coming back to ourselves.
There are two clocks running at the same time, and the entire investment case is a bet on which one ticks faster. Clock One is the speed of AI adoption — and right now, that clock is sprinting, evidenced by $46 billion-plus of contracted demand from just two customers and a capacity auction that just set a record price. Clock Two is the speed at which GPU hardware ages, gets commoditized, and loses pricing power — a clock that runs on its own schedule regardless of how excited anyone is about AI. If Clock One keeps winning for another three to five years, Nebius's asset-light, software-heavy model throws off enormous cash, and today's price looks cheap in hindsight. If Clock Two catches up first, the depreciation trap Burry is describing closes shut, and today's price looks like exactly the kind of top that gets talked about for years afterward. We don't think anyone — including us — can tell you with certainty which clock wins. What we can tell you is which one we'd bet on, and how much we'd risk finding out.
15 What You're Actually Paying
Let's be plain about the price tag, because this is where discipline matters most. Nebius trades today at $249.57, down about 3.7% today after a genuinely wild stretch: the stock rallied 34% in a single session on August 12 after this earnings report, closing that day near $259, before giving some of it back. The 52-week range is $62.01 to $299.86 — meaning this stock has been worth less than a quarter of today's price and nearly 20% more than today's price, both within the last twelve months. That is not a typo, and it is not a low-volatility holding by any definition.
On traditional measures, this is one of the most expensive stocks we cover: trailing earnings multiples running into the hundreds, by some measures well over a thousand, because net income of roughly $42 million is spread across a market capitalization north of $68 billion. A traditional value investor would simply close the file here, and we understand why. But a trailing P/E on a company scaling this fast, off a base this small, tells you almost nothing useful — it's measuring last quarter's rounding error against next year's business. The more honest way to size this up is against the contracted backlog: somewhere north of $46 billion in signed commitments from Microsoft and Meta alone, against a market cap of $68 billion. Wall Street's own average price target sits at $355, implying over 40% upside from here, though we'd flag that coverage is thin — only a couple of analysts publish targets on this name today, and the Street-high target of $410 comes from a single, recently very bullish note. We'd treat any of these numbers as informed guesses rather than gospel.
16 What Could Go Wrong
- The depreciation trap is a real accounting question, not just a bear-case talking point: if GPUs genuinely need replacing faster than the books currently assume, today's margins are partly borrowed from the future. We'd want to see management address useful-life assumptions directly, not just point to EBITDA.
- Customer concentration is severe: Microsoft and Meta together represent the overwhelming majority of the contracted backlog underpinning the bull case. A renegotiation, a delay, or a strategy shift at either customer would matter far more here than it would at a company with a broader base.
- Vineland is a single point of failure with a hard deadline attached: a facility that's already reportedly facing regulatory and inspection delays sits directly on the critical path for honoring the Microsoft contract's capacity schedule. This is the single most concrete, trackable risk in this entire note.
- The valuation leaves almost no room for a stumble: at well over a thousand times trailing earnings, this stock is priced for near-flawless execution. Any meaningful deceleration in growth, or any sign the AI infrastructure buildout broadly is cooling, would likely produce outsized multiple compression here specifically.
- Capital intensity is genuinely enormous: $5.7 billion of quarterly capital expenditure and over $8.1 billion in the first half of 2026 alone is a lot of money to keep raising and deploying correctly, quarter after quarter, even with the asset-light partner model absorbing some of the burden.
- The stock's own volatility is a risk in itself: a 52-week range spanning $62 to $300, including a 34% single-day move, means position sizing matters enormously here. This is not a stock to own with money you need to stay calm about.
17 Muffett's Take
We like the megatrend. We like that Nebius sits at the genuine intersection of two things we care about — the AI infrastructure buildout, and the broader shift of critical compute capacity toward companies the West can actually trust and audit, a company that quite literally extracted itself from Russian ownership to earn that trust. We like the balance sheet discipline relative to CoreWeave, the asset-light partner model that shifts real capital risk off Nebius's own books, and a management team that just posted a 41% adjusted EBITDA margin while growing revenue 454% — numbers that, put next to each other, most companies never get to claim at the same time.
What we don't love is the price of admission, and we think it would be dishonest to pretend otherwise. This is a stock priced for the bull case to be substantially correct, with very little cushion if the bear case — Burry's depreciation trap, or a Vineland delay, or a Microsoft renegotiation — turns out to have real teeth. Both cases in this note are argued by serious people with real money behind their conviction, and we respect that this genuinely is a two-sided argument rather than an obvious call in either direction.
