Most people focus on the visible front end of AI the chatbots, the video tools, the model announcements. But none of it works without massive compute underneath. That is where Nebius Group N.V. NBIS comes in. It is not trying to build the next GPT. It is building the infrastructure that GPT depends on. It is betting that while models come and go, the demand for compute will grow forever. The future of intelligence runs through data centers. And Nebius wants to own that future.
This is one of the most fascinating businesses in the AI ecosystem today. The kind of company that could one day become a quiet compounder, like the broadband utilities of the last generation. But it is not there yet. Right now, Nebius sits between a compelling opportunity and a set of real execution risks. To understand if investors should participate in its rise, we need to dig beneath the hype and into the numbers, the risks, the economics, and the cold hard truth of capital-intensive growth. Because in this business, timing is everything.
A Business That Grows Because AI Cannot Slow Down
Every major leap in AI requires exponentially more compute than the leap before it. Model parameters are exploding. Training cycles stretch longer. GPU clusters that seemed massive a year ago now look insufficient. That demand curve never dips. It doesn't care about interest rates or election cycles. The world keeps building smarter systems and Nebius gets paid every time intelligence scales. Nebius operates high-performance data centers filled with NVIDIA NVDA H100 and Blackwell GPUs. They are purpose-built for model training, inference, and workloads that depend on maximum throughput. Any AI company that needs speed, scale and availability can pay Nebius instead of trying to build a private GPU farm that costs billions. The idea is simple. The execution is not. For Nebius, scale equals survival. And survival depends on filling those GPUs with paying workloads fast enough to outrun the capital bill that built them.
The Microsoft Deal That Put Nebius on the Map
Every great infrastructure business starts with an anchor client. Nebius secured one of the biggest possible.A five-year USD 17.4 billion core commitment with up to USD 2 billion in optional expansion bringing the total potential Microsoft MSFT contract value to USD 19.4 billion. This scale is unusual for a company still early in its public life. A relatively young public company is now supplying a portion of Microsoft's AI compute pipeline.
Microsoft needed fresh GPU capacity outside its existing hyperscale data centers. Nebius needed a client that could absorb large-scale infrastructure as fast as it was built. Both needs aligned in one of the most surprising partnerships in the AI era so far. This deal did three things. It validated Nebius as a real alternative to AWS and Google Cloud. It guaranteed early revenue scale. It made Nebius instantly relevant in the global supply chain for AI compute. But it also introduces risk. When one relationship supports so much of the story, it becomes a single point of failure. Microsoft is not just a customer. It is a dependency.
One reason this deal is so extraordinary: Nebius secured local approvals in Vineland, New Jersey by partnering directly with Vineland Municipal Electric Utility a rare municipal utility that controls its own generation. That gave Nebius behind-the-meter power and drastically shorter permitting timelines. Ironically, this is the exact bottleneck that even Microsoft's own teams struggle with under U.S. federal/state utility rules. That operational detail is the difference between theoretical scale and actual buildout.
The Moat Nebius Is Trying to Build While Running Full Speed
Nebius is not competing in a traditional cloud market. It is operating in a segment where performance, scale, and capital access determine who stays relevant. Its moat is forming from a combination of hardware physics, software integration, access to capital, and geopolitical positioning. First, performance switching costs. AI training infrastructure is not easy to move. Once a company tunes its workflows to Nebius's GPU clusters, ripping and replacing is painful, slow and expensive. Nebius wins every month a customer stays and does not think about leaving.
Second, sovereignty. Nebius is headquartered in Amsterdam and expanding across Europe and the Middle East. Nations increasingly want their intelligence infrastructure located on home turf, not under the full policy reach of the United States. Nebius becomes the preferred partner where digital sovereignty matters. Third, supply allocation. Nvidia controls the pace of AI progress. GPUs go first to trusted strategic buyers. The best chip in the world is the one you can actually get. Nebius secured its access early, including Blackwell GPUs that will define the next generation of model training. Fourth, credibility through partnerships. Microsoft didn't choose Nebius by accident. Neither did Nvidia or Accel Partners. Their presence signals operational competence. Trust layered on trust. The moat exists as potential. It is forming. It is not yet unbreakable.
Nebius' Structural Electricity Advantage: The Hidden Driver of GPU Economics
One structural advantage deserves far more emphasis: Nebius' position inside the Vineland Municipal Electric Utility system. Because Vineland operates as a municipal utility with its own generation, Nebius avoids several layers of cost that private-utility data centers must absorb including PJM transmission charges, ISO congestion fees, PJM capacity charges, and the return-on-equity markup embedded in investor-owned utilities. Even using conservative assumptions, this creates a 410/kWh structural power cost advantage, which translates into tens of millions of annual savings for a 24/7 hyperscale AI load. This matters directly to the GPU-hour spread underpinning Nebius' business model: lower delivered power costs expand the contribution margin and increase the probability that Nebius reaches self-funding scale before its debt load becomes restrictive.
Vineland's structure also accelerates Nebius' deployment cycle. Because the municipal utility controls its own grid and interconnection, Nebius avoids the PJM interconnection queue that can delay large-load data centers by three to seven years in neighboring states. This advantage is not theoretical it is a real compression of build-time risk, which shortens the path to utilisation and reduces the carry cost of capital. In a model where timing and economics determine survival, cheaper surcharge-free electricity and faster grid access meaningfully improve Nebius' odds of outrunning its capital intensity.
The Financial Profile: Hyper-Growth Meets Hyper-Burn
Nebius is growing at a rate rarely seen in infrastructure businesses. Revenue in 2024 hit USD 117.5 million, a 462% jump from 2023. Q1 2025 brought in USD 55.3 million, and Q2 followed with USD 105.1 million. That's not year-over-year growth. That's quarter-over-quarter doubling. The company's annualized revenue run-rate now tops USD 400 million, with management targeting USD 900 million to USD 1.1 billion by year-end 2025. Investors pay close attention to curves like that.
The right way to think about Nebius isn't GAAP revenue. It's GPU-hour spread. A GB300 slot fully loaded with networking and liquid cooling costs about USD 58,000. Economic life is three years. That's ~USD 2.20/hour of depreciation. Power + cooling add ~USD 0.39/hour. O&M adds ~USD 0.08/hour. All-in economic cost: ~USD 2.67 per GPU-hour. Microsoft is paying Nebius the equivalent of ~USD 4.43/hour (USD 19.4B / 5 years / ~100,000 GPUs / 8,760 hours). That is ~USD 1.76/hour of economic contribution margin. That spread not revenue growth is the real business.
And when you annualize that spread over 100,000 GPUs, Nebius generates roughly USD 1.54 billion of economic cash flow per year before debt service. This is the heart of the model. This is not a tech trade it is a financing trade. If Nebius can borrow at ~5% cost of capital instead of ~10%, that single variable alone moves net cash flow by ~USD 300 million per year. Cost of capital isn't a side factor it is the determinant of whether this becomes a compounding machine or a slow-motion insolvency.
So here's another side to that growing revenue chart. The losses are massive. Nebius posted a USD 394 million net loss from continuing operations in 2024. In just the first half of 2025, its accumulated net losses reached USD 235 million. These losses are not accidents they are deliberate investments to achieve scale. This is a business intentionally built to lose money now to dominate later if scale and efficiency catch up in time.
Under the hood, the cash flow statement tells the real story. Operating activities consumed roughly USD 317 million in the first half of 2025, after adjusting for a non-cash USD 597 million revaluation gain from equity investments. Working capital ballooned receivables, VAT recoverables, and prepayments all soaked up liquidity while supplier payments accelerated. Nebius is pouring cash into infrastructure, partnerships, and the future. And then comes capex the other half of the capital load. Property, equipment, lease assets, and intangibles swallowed another USD 1.18 billion. Add that to the operating burn, and free cash flow lands around USD 1.5 billion for just six months.
This is capital intensity at full throttle and Nebius is fully committing to it. Nebius isn't trying to preserve cash; it's trying to build a moat wide enough that competitors struggle to match its scale and power economics. The balance sheet confirms it: over USD 1 billion in new long-term debt, liquidity still near USD 1.7 billion, and equity climbing on paper from revaluation gains.
But dig a little deeper and the story gets more complicated. Nebius's total debt load now exceeds USD 4 billion, including USD 3.1 billion in convertible notes that accrete at 115% of principal a significant source of potential dilution if the stock underperforms. The company also holds about USD 1.5 billion in value from its 28% stake in ClickHouse. That's a strategic cushion, but not cash an asset, not oxygen. The balance sheet isn't broken; it's built for a company racing to reach self-funding before credit markets lose patience. As Buffett likes to ask: If a recession hits tomorrow, does this business survive? Nebius has visibility for months not years. The window is real Nebius must scale into cash flow before capital markets tighten
Indeed, this is one of the most demanding execution paths in tech scale must materialize before the balance sheet becomes restrictive. If it does, this will be remembered as one of the boldest capex sprints in recent memory. If not, it'll be an example of how capital-intensive models can face pressure if utilisation lags.
The Revenue Model: Renting Out the Most Valuable Real Estate in the Digital World
Nebius earns money the way landlords do. It builds capacity. It rents it out. But instead of apartments or office floors, its rentable units are blocks of GPU compute. Contracts vary from on-demand usage to multi-year reserved capacity. The Microsoft deal functions like a core master lease a guaranteed tenant paying premium rates.
As utilization increases, margins rise dramatically. Cost of revenue fell from 53 percent to 29 percent as data centers filled and waste shrank. Adjusted EBITDA losses narrowed sharply as more revenue flowed over a mostly fixed cost base. This is the same financial transition AWS went through long ago. But AWS had time. AWS had Amazon AMZN. Nebius has urgency.
Nebius is not fighting scrappy peers. It is fighting companies with trillion-dollar market caps and near-limitless investment budgets. AWS. Google Cloud. Microsoft Azure Plus specialized rivals like CoreWeave (CRWV), which has shipped more than 250,000 GPUs compared to Nebius's roughly 20,000 deployed or contracted. Nebius does not need to win. It only needs to own a meaningful piece of the fastest-growing compute market ever created. If it builds regional monopolies in Europe and the Middle East, it becomes a strategic necessity. If it loses that race, the business risks becoming high-fixed-cost infrastructure without attractive returns.
Every infrastructure utility transforms the same way. Year 1 to 3: heavy construction. Year 4 to 6: filling capacity. Year 7 onward: profit machine. Nebius is caught between the first two phases. Still building and filling at the same time. As its installed GPU base ramps, depreciation stops crushing margins, revenue gets smoother, and capex eventually tapers. Free cash flow can flip from negative hundreds of millions to positive hundreds of millions fast if scale arrives. Management believes that inflection hits around 2027. Investors must decide whether they believe that too. Because the difference between 2027 and 2029 is the difference between compounding and collapsing.
Valuation: What the Market Is Pretending Not to See
The three valuation outcomes below are not random scenario sketches they are direct functions of the GPU-hour spread we just calculated. If Nebius keeps utilisation above ~80%, holds realised pricing near Microsoft levels, and keeps capital cheap, the ~USD 1.76/hour contribution margin compounds. If utilisation slips, pricing softens, or capital costs spike the margin collapses and valuation collapses with it. This company is valued based on the durability of that spread, not the speed of reported revenue growth.
At USD 31 billion enterprise value today, Nebius trades at nearly 30 times next year's revenue. That is a premium that assumes everything works. Let's define the futures more clearly.
Base Case: USD 10 billion revenue by 2030. 25 percent EBITDA margin. 20x EBITDA exit multiple. Intrinsic value around USD 50 billion. Discount that at 10 percent and the fair value falls around today's valuation.
Bull Case: USD 15 billion revenue by 2030. 30 percent EBITDA margins. Intrinsic value above USD 50 billion. With the 30x EBITDA multiple and 10% discount rate, Nebius should be worth around $83 billion, 2.7 times its current valuation.
Bear Case: Revenue stalls near USD 5 billion. EBITDA Margins peak at 20 percent. At 15x EBITDA valuation and similar 10% discount rate, its fair value drops toward USD 9.3 billion, 30% of the current price only. Downside from here could be 70 percent.
The market isn't pricing the risk. It is pricing the dream. Buffett would not pay for perfection. He would wait for the moment when expectations are lower and cash flow is visible. He would wait for proof. Nebius must keep four critical plates spinning. Microsoft must stay committed. Nvidia must keep allocating chips. Credit markets must stay open. Clients must keep filling clusters. If even one falters, the financial model shakes. If two falter, the model breaks. This is not a stable utility yet. It is a leveraged bet on flawless execution. And yet, this is what makes the upside extraordinary. Risk builds return. But only if risk is managed, not ignored.
Management: The Founder Who Has Done the Impossible Before
Arkady Volozh turned Yandex into one of the greatest tech ecosystems outside the United States and China. He knows scale. He understands data. He has scars and victories behind him. He plays long-term games. Via his family trust LASTAR Trust, he effectively owns 12.9% economic interest in the company, while maintaining 55% voting power. Insiders collectively own more than 18.4 percent of the company, and strategic investors sit on the cap table for a reason. Leadership does not guarantee success. But Nebius is not led by dreamers without proof. It is led by an operator with a history of beating the odds. The question is whether he can beat capital costs too.
The Investor Verdict: The Right Idea at the Most Unforgiving Time
Nebius checks every box of a future compounder. Unavoidable demand. High switching costs. Global tailwinds. Trusted strategic partners. Leadership with conviction. It also carries the classic risks of early-stage infrastructure: customer concentration, supplier dependence, enormous debt, negative cash flow, and a valuation that assumes perfection.
This is not a stock for the risk-averse. It is a stock for investors willing to believe that inevitability can outrun leverage. One day, Nebius could become the AWS of the AI era the utility behind intelligence. But AWS earned that title with cash, not promises. If Nebius flips its cash flow engine on schedule and survives its balance sheet test, this could become a compounding machine. A business you buy once and hold through a decade of exponential demand for intelligence.
Until then, patience is not optional. And the investment takeaway for GuruFocus readers is brutally clean: monitor three levers cost of capital, utilisation, and realised GPU pricing. If Nebius keeps cheap capital and runs its GPUs hot, the model works. If capital tightens or utilisation stalls, the story breaks. Everything reduces to those three variables.