Intro:

Nebius is what Yandex turned into after the Ukraine war. The parent company was forced to sell the Russian search engine in exchange for U.S. listing, and so they did. What’s left, aside from a few exciting equity investments across robotics and more, is the Nebius AI Cloud. As a neocloud, it builds vast data centers, populates them with server racks packed with Nvidia chips and rents that compute power. 

A big part of its pitch is that it also offers a plethora of cloud and inference services on top of that, driving differentiation, cross-selling potential and customer loyalty beyond cyclical chip scarcity. Differentiated value comes from how efficiently customers can run training and inference and how useful the ancillary services built around it are. It’s about the ecosystem… not commoditized bare metal GPU rentals. While this is a pretty abstract idea, it becomes more tangible and compelling when you consider things like best-in-class performance for most benchmarks on MLPerf. SemiAnalysis also praises its platform quality and pricing power. According to public list prices from late August, a 64-GPU B200 cluster costs about 20% more at CoreWeave and roughly double at AWS and Google Cloud. According to CTAIO, for H100s, Nebius’s on-demand rate is about 27% below CoreWeave’s, even after raising it 17% in October. Mega-deals like Microsoft’s and Meta’s are priced privately, so those gaps mainly apply to the long tail renting by the hour.

What are its various products?

Nebius AI Cloud is the platform used to rent access to fully managed compute clusters via various AI-related services. It designs its own servers and features table stakes cloud offerings like storage and networking. Customers can use it to train and post-train their models, run managed inference through its Token Factory (dedicated inference service) and access a slew of software services that enrich and augment the performance of its hardware. Important software services to know include:

Generally speaking, tools like Soperator, Managed MLflow and others are useful in driving perpetual performance and efficiency gains for the NBIS platform and its customers. That gets at why offering a true platform is so important. It’s how you drive differentiation beyond acting as a compute middleman.

Different ways in which those products are delivered?

Giant customers like Microsoft and Meta make up the lion’s share of NBIS’s business and generally demand multi-year, massive and dedicated compute clusters in their contracts. The long tail of customers representing pretty much everyone besides mega-caps uses the managed, shared Nebius AI Cloud infrastructure and software. These customers are able to rent access to compute by the hour through the Nebius AI Cloud, while others can purchase the aforementioned managed inference service from Token Factory. When they do this, they pay by the unit of AI output, rather than the unit of time. It’s the easiest, fastest and cheapest way for enterprises to get started and is gaining rapid traction. For context, it’s approaching a multi-hundred-million-dollar run rate less than a year post-launch. Commonly, capacity-starved customers come specifically for its Token Factory and eventually find all the other value-add services that help them conduct their AI work better. That’s when the up-selling and contract stickiness start to kick in. Nebius sells all of this from one unified platform.

In terms of different formats for capacity:

Quick Thoughts on Why This One?

First, why a neocloud?

As Max readers know, this replaced some of my SOXX exposure recently while I boosted overall allocation to pure-play AI infrastructure from 6% to 7%. Why? We continue to see hefty token pricing pressure. Models are leapfrogging each other left and right and custom chips are successfully altering market share dynamics. Things are rapidly evolving alongside competitive dynamics. But what isn't changing? The expectation is that we are going to need a lot more compute. Hyperscaler CapEx commentary for 2027 and beyond is uniformly positive. Consumer agents are just now beginning to scale. Robotics and autonomous mobility are in their infancy. Models are getting more and more complex. Efficiency gains are fueling incremental usage as Jevons Paradox holds.

Moving forward, I think the most likely bottleneck is energy. The regulatory environment could easily get more restrictive, our archaic grid is already running on overdrive and companies like Bloom Energy can only do so much to alleviate that issue. If we get to a point where we don't have the ability to power the chips and racks that we're already building, that would lead to significant chip deflation in 2028 and beyond. This outcome would hurt many AI infrastructure players… but Nebius would be in much better shape. Their plugged-in compute would become even more scarce and even more valuable in that scenario... meaning the most likely risk I see for the sector actually helps this name.

Why this one?

In short, it comes down to two things. A much cleaner balance sheet that I think can be stretched much further than CoreWeave’s from here (more later). And? A much more developed software ecosystem to drive compute performance gains than IREN. They blend these two things – software utility and relative balance sheet health – better than any other pure-play in the space… so I pick them. When cybersecurity leaders like Palo Alto CEO Nikesh Arora talk about neoclouds being a massive bubble, they don't include Nebius in that equation. Why? Because this company doesn't exclusively rely on a purely commoditized product (bare metal). They take this product and they enrich it with obsessive, constant optimizations to compute efficiency and agent/model workflows and data processing. They create more integrated compute services that can maintain pricing power much better than selling a customer GPU-based compute and leaving the rest to them. Nebius creates more dependence.

Next, I love their relationship with Meta. They aren't as reliant on OpenAI and Anthropic as most of the others are. Meta has agreed to purchase up to $15B more in compute capacity through 2031 (on top of $12B already committed) if Nebius doesn't sell it to others. I think there is real financial risk for OpenAI/Anthropic. There isn't for Meta. That is one of the highest-quality customers anyone can possibly have and they're agreeing to backstop a lot of financial risk for Nebius.

As always, there are more risks. Expect more dilution and debt raising even as customer prepayments make up a larger portion of its overall book. They’ve explicitly said this will happen. Something we need to be ok with if we’re going to own this. The annual revenue they receive per megawatt (now $20M+ and sometimes up to $50M for spot market bidding wars) and the ongoing fixed cost leverage we’re seeing in this model make spending the correct decision. At the same time, it means they’ll need capital markets to cooperate to fund this buildout (or just sell to Meta). 

For another potential future risk, Nebius is making a ton of extra dollars on red-hot spot market pricing. This spot market creates financial upside today, but it offers real-time price discovery that could more quickly lead to pricing pressure across new contracts if the backdrop worsens. 

And generally speaking, their financial results look amazing because the mismatch between compute demand and supply is enormous. If that goes away, NBIS’s results would quickly look very ugly. I just don't think that's an overly pressing concern at this stage of the cycle. I'll keep obsessively searching for signs of cycle cracks forming... but none of the signs are alarming right now, in my opinion. Next, some think depreciation schedules are unfair and will shrink. Considering depreciation is by far the biggest operating expense here, that would be an issue for margins and returns. All I’d say there is that 5-6 year depreciation schedules look entirely fair when 6-year-old GPUs are renting for healthy prices (like they are). Yes, largely that's because of shortages, but there’s a very long way to go between what we’re seeing now and economically worthless.

Finally, there's execution risk here. That's always true, but it's especially true when revenue is set to go from $500M to $50B+ from 2025 to 2030 and it was on display during this Q3 report (more later). That ramp depends on a ton of capacity coming online as expected. It entails exponentially boosting their footprint and moving through giant construction projects with complex supply chains on schedule. That is far from certain. But? If they pull this off? We could be 12 months away from this trading for 15x forward EBIT with EBIT growing at an (expected) triple-digit multi-year clip. That leaves enough potential financial upside to make me comfortable with the significant downside risks. It makes me compelled enough to own a very small position. I will not be adding to this from here regardless of how far it falls. It's either going to work extremely well or horrendously. I would NOT recommend owning this name if you need the money at any point in the next 5 years or have a risk tolerance that is any lower than mine. If you have an especially high risk tolerance beyond mine and use riskier things like options and/or margin... I'd just say be careful. Err on the side of longer expirations... and just don't use margin here.

Nebius beat revenue estimates by 8%. Revenue included a $325M boost from non-recurring revenue tied to asset-light rollouts. This wasn’t a shock, but without this help, Q/Q growth would have been 13% and Y/Y growth 350%. At the same time, as they keep shifting to the asset-light deployment model, these “non-recurring” pieces are going to get more and more frequent and shouldn’t be discounted as accounting gimmicks. It’s real revenue from reaching milestones of onboarding customers and bringing capacity online. For context, this was NBIS’s first-ever asset-light deal. Revenue from these comes in two chunks: one when it locks in the design and supply chain for a partner’s capacity… and another when that capacity goes live and gets sold.

Annualized run-rate revenue (ARR) more than doubled Q/Q to $6.4B, but this is one of the last times we’ll talk about it. Leadership is moving away from the metric, as those same asset-light related revenue recognition events cited above can lead to ARR (again annualized run rate not annual recurring revenue) getting quite noisy. 

Total input and operating expenses were 95% of revenue vs. 189% Y/Y. Roughly 7 points of that improvement came from input costs, 19 from R&D, 37 from SG&A and 32 from depreciation and amortization (D&A). Revenue growing this quickly explains most of that leverage and the $325M high-margin revenue bump helped too. And finally, the decision earlier this year to move from 4- to 5-year useful lives for their servers and networking gear is easing depreciation intensity meaningfully. That’s a huge expense for them, and this accounting move cut it by up to roughly 9 points as a percent of revenue this quarter. For most companies I cover, EBITDA is a decent proxy for operating profit. Not here. Chips are the product, and D&A (36% of revenue) is how the cost of those chips shows up. That makes Nebius’s 47.5% EBITDA margin feel a lot more like a gross margin than a typical EBITDA margin. For context, when moving from adjusted EBITDA to GAAP pre-tax income (EBT), that 47.5% margin moves to negative.

It has been a busy year for capital raising. Nebius issued a little over $10B in convertible notes between its March and August raises and added another $1.0B through its at-the-market (ATM) stock program in Q3 alongside a short-term Citi facility. Nvidia also invested $2B in March. All in, that’s roughly $20B raised in 2026. 

Interestingly, leadership confidently told investors today that future capital raises are up to them. It’s their choice. How could a company say that when it has years of deep cash burn ahead of them beyond what their current balance sheet can support? Amazing question. In a word: Prepayments. While prepayments covered about 55% of the CapEx behind new 1+ year deals last quarter, they covered more than 100% this quarter – a 45+ point spike. Across all new contracts, the average prepayment share of contract value jumped about 20 points Q/Q. When Nebius sees favorable debt or equity raise opportunities, it can tap those markets with more assets left to back secured debt than its competitors. When those conditions aren’t favorable… like right now… it can require that customers front over 100% of newly committed CapEx. That’s the luxury of its demand environment. While investor sentiment swings like a roller coaster, that reality has not changed.

While this is nice, I will say I’d prefer them leaning fully into prepayments over seeing much more debt and equity raising. I realize that would inevitably cap some pricing upside as customers demand concessions for bearing the financial risk… but that’s a great trade for this company. They are expected to compound at a 180%+ clip from 2026-2029 to reach $36B in revenue and $12 in GAAP EPS. That is insanely impressive and also quite uncertain. I think sacrificing a few points of growth while eliminating a great deal of liquidity and financial risk would be something investors (including myself) would applaud. I would take a much higher degree of certainty in $10 in GAAP EPS (12x 2029 EPS) over the existing uncertainty and $12 in GAAP EPS (10x 2029 EPS). Just me.

Between prepayments and asset-backed financing for mainly its Meta and Microsoft contracts, leadership expects to cover about $34B of its expected $42.5B in 2027 CapEx. That nearly $40B backlog is ripe with lucrative assets to access more capital. From the $8.5B gap, about $5B should be sitting on the balance sheet after October raises and estimated Q4 cash burn. That leaves a $3.5B funding gap to plug with debt, equity or more prepayments. It will inevitably want to add more than that to make sure it has some level of cash on the balance sheet.

The table below compares Nebius to IREN and CoreWeave after all known raises. It’s the best positioned of the three on nearly every leverage metric, and it isn’t close vs. CoreWeave.

For 2026:

For 2027:

From 2027 on, it will guide to revenue and pre-tax earnings. Leadership thinks those say more about how its business is really doing. I like that. Pre-tax profit includes depreciation and interest, the two largest costs of this model. Furthermore, ARR can unfairly reward or punish NBIS depending on how much asset-light non-recurring revenue it pocketed in a given month. It’s supposed to provide a smoother view of demand, and it’s now doing the opposite by amplifying the monthly peaks/declines.

NBIS trades for 12x forward EBITDA. EBITDA is expected to grow by 20x this year, 370% next year and 107% the year after. EBT is expected to turn positive by the end of 2027. Net income is expected to turn positive some time in 2028. FCF is expected to turn positive some time in 2029.

Dictating Deal Terms – Pricing, Duration & Structure:

To give an idea of how favorable the multi-year demand backdrop currently is, deal size and duration both continued to rise despite NBIS introducing new ways to add short-term capacity. They’re dictating terms and notching ongoing megawatt pricing power. For context, last quarter, NBIS disclosed pricing per megawatt jumped from a little over $10M at the time of the spin-off to $22.5M on new deals. This quarter, that again ticked higher to $25M+.

On one end, as briefly mentioned, NBIS is introducing various market models to offer short-term compute. This helps them take full advantage of the vast supply/demand mismatch, often fetching over $40M per megawatt for these more temporary deals. They’re now letting customers bid for blocks of guaranteed capacity and renting out idle chips at spot rates without large commitments or long-term lock-ins. NBIS has enticing flexibility to make this capacity available when it’s able to, and quickly pull it when reserved demand requires them to do so.

Along with on-demand and short-term clusters, these give customers more ways in the door and give Nebius a live read on demand. Early results are encouraging. Reserve auctions closed 15%+ above the minimum price. Both winners were brand-new customers, and the first has already grown its contract value 10x.

Nebius is also enjoying stronger 5-year contract demand for Nebius AI Cloud (not just dedicated deals with mega-caps). These deals are long, large and typically prepaid all at once. Customers lock in today’s rates during a time when GPU prices continue to soar. Nebius gets better visibility and a sub-2-year payback period.

Capacity – Footprint Updates:

Q3 was Nebius’s biggest quarter ever for adding active capacity. Four new sites across the U.S. and Europe started producing revenue, and New Jersey got more tranches in Q3 and into November. More is coming in Q4, including new colocation phases in the U.S., Europe and Israel. It’s at 4+ GW of contracted power today and still expects 5 GW by year-end. As a reminder, contracted power is not operating capacity. It’s secured access to electricity. The buildings, cooling, networking and chips still have to show up before a customer can use any of it. 

Capacity – Geographic Expansion:

In terms of geographic footprint expansion, Asia-Pacific is going well about 6 months after launching. Customer contracts have been signed across Korea, Japan, India, Australia and Singapore and they’ve hired managers in Korea and Japan to spur more growth. Its first two facilities in the region will be colocation projects in India. Both are expected to start operating in 2027, which is when regional revenue should start to scale.

Moving into more parts of the globe is vital for NBIS’s value proposition. On one hand, governments want to run their AI infrastructure locally to help with tight data controls and easier security. NBIS can provide that, with localized go-to-market across Europe (and now APAC) that lawmakers and core suppliers like Nvidia find valuable. Speaking of Nvidia, they also get to use NBIS to cut through project red tape and expedite regulatory procedures with their more local knowledge. Every country handles data residency, chip export rules and permitting differently. It’s annoying, hectic and inconsistent. Nvidia obviously wants its chips everywhere, and it’s much easier to get there through local, independent cloud partners that already have the sites, power and approvals than to do it alone. That’s part of why Nvidia invests in companies like Nebius and hands them new hardware early. NBIS is happy to be the soldier, considering how much money they expect to make from these AI factories.

Capacity – Mix Shift to Owned:

So demand is insatiable, pricing is rising… and Nebius still cut its revenue guide. Why? The capacity didn’t arrive on time. A number of colocation sites are now slated for delivery in the first half of 2027. According to leadership, this is purely timing-related. Nothing was canceled. They proactively communicated this with customers and expect the delays to be short-lived. In August, DataOne, a key New Jersey facility partner, told NBIS the original generator plan wouldn’t work, forcing a switch to cleaner, behind-the-meter Bloom Energy fuel cells. DataOne’s temporary generators were supposed to bridge the gap while Bloom ramped, but it wound them down faster than expected. That gap caused the delay. Again… high degree of execution risk here.

All of this is even more strongly motivating NBIS to prioritize owned facilities over colocations. This is how they pocket a larger piece of the profit pie. This is how they control delivery schedules, with all owned NBIS facilities currently on time. This format also grants more flexibility to scale up capacity or slow down expansion if the backdrop softens for whatever reason. It does mean a lot more CapEx to get projects online, but once that’s incurred, they get all of these benefits in addition to higher-margin business. Nebius also designs its own racks and storage, and leadership expects those costs to keep falling as more capacity comes from owned sites. As seen below, NBIS plans to rapidly shift to this format going forward, and it already has “all power and components secured” for its 2027 demand plan. And yes… that includes Nvidia’s next-gen hardware. Supply of Vera CPUs, Vera Rubin GPUs and Groq chips is confirmed for 2027, with testing and benchmarking now done.

More on Strong Customers:

Nebius has $40B+ in committed revenue from investment-grade customers, led by Microsoft and Meta. Microsoft signed a contract worth $17.4B to $19.4B in September 2025. Meta signed for $3B that November and added up to $27B more this March ($12B committed, plus up to $15B for capacity Nebius doesn’t sell to others first).

Why are these backstops so much more comforting than one from OpenAI or Anthropic? Because of who’s writing the check. Microsoft and Meta are two of the most profitable companies on the planet, with pristine credit ratings. The AI labs are growing at an incredible pace, but they’re still far more dependent on external capital to honor multi-year compute contracts. If funding markets tighten, that’s a problem for whoever built capacity for them. Lenders see the difference too. It’s why Nebius can borrow against these contracts at SOFR + 2.50%. That’s exactly what it did in July with a $775M facility, which is the first under what leadership calls a “repeatable framework.” Other hyperscaler-backed GPU loans price similarly (SOFR + 2.25% for both CoreWeave’s Meta-backed facility and IREN’s Microsoft-backed one), while CoreWeave’s 2026 loans backed by AI labs and smaller customers priced at SOFR + 4.50% to 5.50%. At the same time, this also invites concentration risk. Two customers make up most of the backlog. And as our Amazon review covered, hyperscalers earn much lower margins renting from neoclouds than using their own sites. They’re here because they’re short on capacity. When that shortage ends, renewals are no guarantee. The good news? These are 5-year contracts and new TCV is increasingly coming from AI natives and enterprise platforms like Databricks. The base is diversifying.

More on Depreciation & Payback Period:

Ignoring prepayments and considering all costs, expected payback on comparable 1-3 year deals shortened from 22 months to 19 months Q/Q. That’s so important. GPUs cost a fortune and depreciate far more quickly than other assets like the data center buildings. These being useful for 60 months means NBIS has 41 months of profits to collect. That 41 months is worth north of $85M of revenue per megawatt at current contract pricing… with most of that revenue flowing down to net income. Anything this company can do to shrink payback periods by even a single month is worth roughly $2M in added profit per megawatt, and stretching useful lives has the same impact.

In fairness, there’s a counterargument. Old chips are holding value because compute is scarce. If supply catches up with demand, the oldest and least power-efficient chips lose pricing power first. And if power becomes the bottleneck, newer chips that produce more tokens per watt can justify swapping old ones out early. I wouldn’t pencil in 7 or 8 years. But 5 looks safe.

Platform Innovation – Inference (Token Factory), Post-Training & Agent Execution:

As mentioned, run rate here is approaching multi-hundred-million-dollar levels while contract value rose 17x Q/Q with compelling profitability. Margins are notably favorable compared to other inference providers due to NBIS's direct chip ownership. To keep nurturing the momentum, it purchased EigenAI to increase tokens-per-chip throughput while also buying Inferize last month to fix the “cold start” problem (slow model loading when demand spikes). That will allow chip utilization rates to rise further and NBIS performance dynamics to improve too.

Next is post-training. Some quick reminders. Training is building a model from scratch. Inference is using that model to conduct work. Post-training is ongoing upgrades to existing models to make them better. Examples include fine-tuning it on a company’s own data, and teaching it through trial and error. To avoid model lock-in, customers can easily transfer that progress to a different model if they find a cheaper option. And importantly, on NBIS, customers can do all of this while running their inference engines and accessing a full suite of software in one single place. It’s a routine sale too, as the infrastructure (training clusters and Token Factory) is already there. Applied Compute and Cursive both post-trained custom models on Nebius AI Cloud, then picked Token Factory for inference.

Finally, like we’ve talked about in other AI infrastructure reviews, agent execution requires a lot of CPU-based compute too. GPUs are incredibly powerful, but they need clear instruction and orchestration, and that’s where CPUs come into play. So Nebius is adding CPU capacity next to its GPUs, including those early Vera CPU deployments. CPU-rich work generally means higher margins too, and it opens up a wide array of potential enterprise agent services to enrich margins and returns. This is why Tavily is so important. Giving agents the search tools needed to fetch timely information and insight is a big part of ensuring outputs are high-quality just like CPU-based presence is.

More Notes:

Messy quarter. Third-party issues held back their 2026 guidance and amplified the already loud execution risks surrounding this neocloud and every other. At the same time, it’s good to hear that all owned projects are on time, considering that’s the focus going forward. Furthermore, everything from a demand point of view looks excellent. Pricing… contract duration… prepayment mix… new types of customers… everything. That’s happening while they commit to positive pre-tax income by Q4 2027, showing everyone how real the operating leverage currently is in this business. That’s absolutely partially dependent on future pricing remaining favorable, but it’s also simply a byproduct of hefty upfront spend translating into revenue as part of multi-year contracts.

None of the looming debates were going to be settled this quarter. All this could have been was a positive progress report telling investors the highly difficult plan could actually work. And? Aside from unimportant near-term timing issues, it was. This, to me, is a moonshot. The potential upside associated with its expected revenue and profit ramps is tempting enough to warrant a small position. The downside associated with cycle length uncertainty, capital market uncertainty and supply uncertainty is why I want it to be small right now. I’m happy to let it grow to a larger part of the portfolio… but only on its own.