Welcome to Nebius Group's Q3 2026 earnings conference call. The presentation will be followed by a Q&A session. If you'd like to ask a question, you can click the Ask a Question tab in the top right of the live stream player. Then just type in your question and click Submit. You can submit questions at any time during the presentation, and the Nebius management team will try and answer them during the Q&A portion of the call. I will now hand it over to Gili Naftalovich, Head of Investor Relations, to start the call.
Hi, everyone, and welcome to Nebius' third quarter 2026 earnings conference call. Joining us on the call today, we have our CEO, Arkady, our CFO, Dado, and the broader Nebius management team. Before we get started, I'll quickly cover the safe harbor. Some of the statements that we make today regarding our business operations and financial performance may be considered forward-looking. Such statements are based on current expectations and assumptions that are subject to a number of risks and uncertainties. Actual results could differ materially. Please refer to our Form 20-F for a list of our risk factors. We undertake no obligation to update any forward-looking statements. During this call, we will present both GAAP and non-GAAP financial measures. A reconciliation of non-GAAP to GAAP measures is included in today's earnings press release.
All earnings-related documents are distributed and available to the public through our Investor Relations website, which can be found at nebius.com. And now, I'd like to turn the call over to Arkady.
Thanks, Gili, and welcome everyone to our call today. Last quarter, I gave you a better understanding of our business model. This quarter, I want to show you how it's working. First, just to say that Q3 was a record quarter. In a market growing like this, it should be. Capital markets have been unsettled lately. Two questions are being asked our whole industry: How fast can AI build-out happen, and who will pay for it? These are important questions. The way I think about it is simple. The demand for AI is real and substantial. Everything else, like building the capacity and financing it, is just a task. Tasks can be hard, but they all can be solved. Demand is strong. Our customers are committing to us for longer, at higher prices, and they're paying more upfront. Prices kept rising.
Q3 agreements closed above $25 million per megawatt, another step up from the $20 million-$25 million per megawatt we achieved in Q2. Our customer agreements are getting larger in size. Three months ago, we reported our first billion-dollar-plus contracts. In Q4, we are already signing multi-billion dollar agreements for our AI cloud. The base of that demand is broadening too. AI natives are scaling with us. Enterprise AI leaders like Databricks, Datadog, and Snowflake signed multi-year agreements with us for the first time during this quarter. Everyone knows that demand is exceptional right now, but I also want to talk about how we are preparing for the future. As you remember, we sell on three horizons, and we choose the best mix between the three. First is on spot and short-term at a premium price. Another is one to three year contracts, and this is our core business.
A limited number of five-year agreements diversify our contract tenures and further help us to finance our build-out. The second and third are covered by prepayments. Two things are important here: how much of our CapEx they cover and the payback period for our new capacity. Let me explain this in more detail. A customer pays today for capacity that comes online in the next 12 months. Why do they do it? First of all, they want to be sure they reserve their capacity, and they want to reserve today's price. The most important change is that the prepayments on new contracts now cover more than 100% of the capital needed to bring this contracted capacity online. Just to remind you, last quarter, it was 50%-60%.
As a result, payback in Q3 shortened to one year and seven months from one year and 10 months a quarter earlier. Our customers are paying more, they pay earlier, and they are committing for longer, and we are getting our capital back faster. Now, let me turn to how we achieve this and how it allows us to grow profitably. First, how we sell. We are creating new tools to find the real price of what we are selling. These include reserve capacity auctions, public spot pricing, and short-term capacity. The market tells us what our capacity is worth, and we sell at that price. The second thing that helps our margins is, of course, our low cost. We build the full stack, both the compute infrastructure and the software on top of it. We have built Nebius layer by layer since the start.
First, for large-scale training, then for inference, and now for post-training and agentic workloads. Customers start where it suits them and then scale with us as their needs evolve. Our managed inference platform, Token Factory, shows this is working. Token Factory is on track to reach multi-hundred million dollar ARR at better margins because we use our own GPU, and GPU is not rented capacity. While our prices reflect the premium value to our products, our cost structure is one of the lowest on the market. We build the infrastructure ourselves. We design and build our racks, we design and build our storage, and these costs will also fall further as more of our capacity comes from our own sites. The premium prices for our products and the low cost makes our margins.
This is what we have always focused on, to build a healthy and profitable business, and having such margins is fundamental part of that. Our margins are also supported by lower cost of funding. On top of prepayments, we have the financing that we can raise on our investment-grade contracts. Together, all this reduces our costs, expands the market we can serve, and lets us go after higher value customers. That is what takes us to profitability, and we expect to become pre-tax profitable in the fourth quarter of 2027. This is important. At some point, supply and demand in our industry will be more balanced. When that happens, we will be ready. We invest in three things at once, the capacity, the software platform, and go-to-market that meets customers where they are. All three matter equally, but capacity is the foundation of this business. It defines our scale.
Building at this scale is hard, and delays are part of our industry right now. We are not immune. For example, some of our colocation sites, including New Jersey, are behind schedule. Customers have been kept informed, and none of that capacity is canceled. Some of it has shifted to 2027. This means that a portion of our this year's planned revenue moves into the start of next year. With our own sites, we remain on track and even ahead of schedule. That's important because going forward, most of the next gigawatt we're bringing online will be on our own sites, where our own team does everything, land, power, shell, hardware, software, and operations. This gives us confidence for next year. We're building the next gigawatt. This makes us a different scale of a company, and we are well-funded to make it happen.
The payments will be over $10 billion in 2026 and should increase significantly next year. We will have access to financing on attractive terms against our investment-grade contracts, and we also have our strong balance sheet and potential external financing. We have strong confidence in 2027 and are building beyond that. There will be ups and downs in the industry, as in any industry. We are building a business and a financial model that can thrive in good markets and be resilient in challenging times. We expect revenue to quadruple next year on a conservative view of the pace of our build-out, and we expect to reach pre-tax profitability in the fourth quarter. To close, the prospects for our business have never been better. Our confidence in what we're building has not changed. We will keep going. With that, let me hand it over to our CFO, Dado.
Thank you, Arkady. We delivered a strong third quarter. Revenue was up 571% year-over-year, with the majority driven by capacity we added in the last 12 months. EBITDA margin expanded both quarter-over-quarter and year-over-year. We also strengthened our funding position. Prepayments became the standard structure of our agreements, and we expect $10 billion in committed customer prepayments in 2026. Our long-term agreements with Microsoft and Meta can be used for further asset-backed financing, and our balance sheet held $7.4 billion of cash and cash equivalents at the end of the quarter. In addition, cash generated from operation grows with every tranche of capacity that comes online. I will cover our third quarter results, balance sheet, and funding needs for CapEx, then our updated 2026 guidance and our initial outlook for 2027. All comparisons are year-over-year unless noted otherwise.
Let's start with revenue. Group revenue was $980.8 million, up 571% year-over-year and 68% from the second quarter. Nebius AI Cloud revenue was $973.8 million, up 643%, and accounted for approximately 99% of group revenue. In the third quarter, we recognized approximately $325 million of non-recurring revenue as part of our asset-light model. This reflects an initial phase of a rollout as we engage partners to deploy Nebius architecture and the Nebius AI Cloud platform on capacity that they fund. Under these partnerships, we earn revenue both when we enable our architecture and platform on a partner's capacity and when the capacity comes online. As our asset-light footprint grows, we expect revenue from these activities to become a more regular part of our business. Turning to EBITDA.
Group adjusted EBITDA was $465.8 million, compared with $-5.2 million a year ago, and $236.2 million in the second quarter. Group adjusted EBITDA margin was 47.5%, up from 40.6% in the second quarter of 2026. Our Nebius AI Cloud business delivered an adjusted EBITDA margin of 53.2%, up from 50% in the second quarter. The difference between the group and the AI Cloud adjusted EBITDA margin reflects our investments in Avride and TripleTen. Now, to balance sheet and funding. Prepayments became the standard structure of our agreements in the quarter. More than 90% of the contract value we signed in Q3 carried a prepayment, and customers prepaid a larger share of every agreement. The weighted average share of prepayments to total contract value rose by approximately 20 percentage points from the second quarter.
Coverage is, as Arkady described, more than 100% of the CapEx behind our Q3 wins, up from 50%-60% in the second quarter. This has continued to rise on agreements signed so far in Q4. These prepayments cover the full CapEx, which includes costs such as GPU, networking. Prepayments now come with agreements on capacity we have already deployed, as well as on new builds, so they fund the build-out as a whole, rather than one project at a time. Because we choose the mix of contracts we sign, by customer, by duration, and by the prepayment attached, we also choose how much external capital we need and when we go for it. That is what lets us be opportunistic rather than dependent. We raise where the cost and the timing are right.
We weight every source against cost of capital, dilution, and the health of our balance sheet, and we are not reliant on any single route to fund 2027. In July, we closed a $775 million facility, the first under our asset-backed financing framework. That facility was secured against long-term customer agreements with investment-grade customer at SOFR plus 250 points. These terms and the flexibility we have on when to tap into such facilities represents why we sign such deals. We have over $40 billion in such similar commitment revenue and see this as a repeatable framework. In August, we also completed a $5.75 billion convertible senior notes offering. Also in the quarter, under our at-the-market program, we issued 4.6 million Class A shares at a weighted average price of $229.1 per share, for net proceeds of approximately $1.0 billion. 7.6 million shares remain available.
The ATM remains a flexible tool, not a commitment. Turning next to CapEx. Capital expenditures were approximately $7.3 billion in the quarter, driven by server and network equipment and by our own data center expansion. We expect to leverage the range of financing vehicles available to us. Our outlook. We are updating our 2026 guidance and introducing our initial view of 2027 on the same basis, what we see today and a more prudent view on timing across all of our sites. For 2026, as Arkady mentioned, we are seeing some third-party capacity shifts into the second quarter of 2027. We now expect revenue of $3 billion-$3.2 billion, compared with our previous range of $3 billion-$3.4 billion. Annualized run rate revenue reflects the same shifts in timing.
We now expect ARR to exceed $7 billion for the month of February, even before New Jersey ramps up to full capacity. After that update, we will no longer guide to ARR, as our business is scaling into the next phase. This metric does not reflect the full scope of our base. It is based on one month of the quarter, and we don't build our business around any one month. As we scale new partnerships, this revenue can add choppiness to ARR until such revenue ramps up. We are also shifting to the next phase, where our business is based on revenue and profitability. As we are growing and getting mature, we believe revenues and pre-tax earnings are the most appropriate way of evaluating us. We continue to expect a group adjusted EBITDA margin of approximately 40%, unchanged from our prior guidance.
CapEx for 2026 is now expected to be around $27 billion, up from $20 billion-$25 billion previously. Contracted power remains unchanged. We continue to expect 5 GW of contracted power by year-end, up from more than 4 GW today. Demand has never been stronger, and the changes in our 2026 guidance are just about timing, not demand. Turning to our 2027 outlook. We expect revenue to increase by around 4x to $12 billion-$13 billion. We also are excited to have a path to profitability. We now expect to reach break-even on a pre-tax earning basis in the fourth quarter of 2027, on five-year depreciation of our servers and network equipment, as our own sites ramp through the year.
We expect CapEx of $40 billion-$45 billion, sized for more than a gigawatt of new capacity, a majority of which will become online from owned sites that hold a stronger margin profile than colocation capacity. To fund this ramp, I want to highlight that we expect that 75%-85% of our CapEx range will be funded by prepayments and asset-backed financing of our current investment-grade contracts. We expect that in 2027, customer prepayments will account for $17 billion-$20 billion. This number is not a ceiling. Prepayments depend on how we allocate capacity among contract types, and we set that mix for the best balance of price, duration, and prepayment. We also have a very healthy balance sheet, giving us the flexibility to be opportunistic with external capital when it is cost-effective to do so.
We enter 2027 from a position of strength, a clear path to break-even in the fourth quarter, a funding plan built on our customers, our existing contracts, and our balance sheet, and an outlook we are confident we can deliver. We will keep scaling with the same discipline we started with. With that, I will turn the call over to Gili for Q&A.
To ask a question, you can click the Ask a Question tab in the top right of the live stream player. Then just type in your question and click Submit. You can submit questions at any time during the presentation, and the Nebius management team will try to answer during the Q&A portion. I will pass it back to the Nebius management team for a Q&A session.
Thanks, operator. Our first question is from Alex Duval at Goldman Sachs, who asks, if you need more cash on top of current cash, to what extent could that come from prepayments, given conditions in some of the debt markets? Marc?
Thank you, Alex. To answer the question directly and simply, yes, we could tap prepayments to shore up any additional cash requirements that we have. Let me give a little bit more of a detailed answer so that you can get a sense of how prepays are playing a role overall in our execution. As you may recall, back last year, we landed a couple of landmark deals, Microsoft and Meta, you may remember. Both of those came with significant prepay. We have known that prepayments are a fundamental part of our business ever since. For our AI cloud business, we started to actually secure prepayments back in Q1 of this year, and now they are a standard structure of our agreements, with more than 90% of the contract value we signed in Q3 carrying a prepay. So, how do we get such prepayments? There are three components.
First, we are recognized as a market-leading platform, sought by AI builders who are willing to pay a premium for our reliability and service. Second, we said last quarter that we have the demand and could sell all our capacity at current terms, but instead, we were holding back to secure better terms, and we were able to do so. That is what we are experiencing now as we have raised prices, expanded prepayments simultaneously. Third, our absolute dollar amount of the prepayment translates to higher percentages of CapEx due to our overall approach to managing our cost structure to a lower absolute value. Higher prepayments and lower costs has resulted this quarter in greater than 100% coverage of the full CapEx, including data center costs. By the way, an important factor that may not be obvious is that we are receiving prepayments long before delivery.
I should also mention that prepays also apply to capacity we already own, so they help fund our overall build-out. I hope you recognize that our overall prepayment program provides us great flexibility in responsibly financing our business going forward. As you called out, Alex, it gives us the lever that we may need at times in order to be able to shore up cash requirements. I think Ophir might be able to provide some additional insights in the overall financing utilizing prepayments.
Thank you, Marc. I think that you covered it quite well, but at the end of the day, when your prepayment on your contracts covers more than 100% of the CapEx, you are in a great position. When you are making the allocations between the various type of contracts, you have the flexibility to take into account also the capital market conditions. In periods where the interest rates, for example, for corporate level debt will be high, we may decide to allocate our capacity differently, provide more for longer term contracts, receiving a lot of prepayments covering our CapEx needs. If market conditions will change, we may be very much opportunistic, raise corporate level debt or other instruments, and then keep more of the capacity for a shorter period of time.
We have ton of flexibility, and we feel that this quarter result, both in Q3 and what we see in Q4 about the prepayments, give us the ability to finance our business in a very attractive way.
We have another question from Alex Duval. Can you contextualize how Nebius works with local governments and communities in places like Texas? Can your practices help mitigate deployment issues? Tom?
Yep. Hi. Thanks, Gili, and thanks, Alex, for the question. Yes, so specifically with regards to Texas, we don't currently have anything there, but I can definitely speak to the half dozen states or so where we have existing operations or where we're building. As of today, we don't see or anticipate any meaningful impact on timelines coming from the shift in the political landscape. I guess just to give some also some broader context, because obviously this is a topic that we pay very close attention to and it's been the subject of a lot of discussion. I think that the questions around AI and how we see it, the questions around AI and towards data centers in particular, they haven't just been coming from the Democrats, actually. Questions have been coming very much from both sides.
By the way, we think many of the questions are very legitimate and deserve answers, and we try to give those answers. We don't see that the overall trajectory of sentiment has changed since the elections. Questions will continue to be asked and it's on us to ensure we're showing up. We engage very actively with state leaders, with communities, and so on, where we try to be very transparent, explain what we're doing, and to explain the benefits that we can bring to communities. We think it's something that we do well, and I'm sure we'll continue to do so.
Thanks, Tom. Next, we have one from Rob Oliver at Baird. What is the financing strategy for 2027? How much additional funding is required to support the current backlog and data center build-out? Dado?
Thank you, Rob. Actually, Marc and Ophir have already shared quite a lot of details about our financing approach. We have worked really, really hard on our product and GTM offering to reach prepayments that enable us to cover over 100% of the CapEx. We are getting this well in advance of the capacity. We have also worked hard to have available tens of billions of dollars of investment-grade contracts that will enable us to receive significant capital in asset-backed financing based on these contracts at very attractive terms. At the end of the day, these two factors will give us a lot of flexibility. You should remember that we can allocate our capacity in the core business between the long-term contracts and the short-term contracts based on our capital needs.
For example, if the terms in the capital market will not be attractive, we can allocate more capacity to the long-term contracts. If, on the other hand, we will have the availability to raise capital at attractive terms, we will be very opportunistic about it, and we will shift more of those allocations. At the end of the day, prepayments are covering 100% of the CapEx and puts us in a very good place with respect to our financing needs. Look, if markets and financing markets are not attractive, we can do more core and long-term contracts, less on spot and short-term deliveries. If we can raise more, we may allocate more to the short-term and spot market. We have the flexibility to optimize the mix.
To be in a position where you can cover 100% of the CapEx, including data center, this is a very good place to be.
Thank you, Dado. Daniel Wang from BNP Paribas asks us, can you provide more color on the underlying causes of the delivery delays at Vineland, and what other sites are impacted as the shareholder letter cited several delays across your colocation sites? Andrey?
Yep. Thank you, Gili. Thanks, Daniel, for the question. Regarding the Vineland, the main reason for the delays are the power story there. When we pivoted to the Bloom generation in summer, the plan was that we will use for some time the temporary generation provided by the DataOne while the Bloom generation will be ramping up. In late Q3, DataOne decided to wind down the temporary generation faster than it was expected, so we are expediting the Bloom deployment as much as possible, but we still will have a gap due to this situation. Having said that, I just wanted to mention that we continue to deliver the new tranches to Microsoft throughout the Q3, and we will continue to do so in November and December, together with the power delivery.
The second point is that the Bloom is already providing basically electron. The generation is live and providing electrons on site. It's just a question of ramping that up as soon as possible.
Yeah.
Right.
I would just probably add, as I said in my part before, this is one of the sites which we build, not ourselves directly, but with help of third-party constructors. As you can expect, many more things may go wrong in such sites. In this case, again, in August, we were told that the old plan of diesel generators will not work. We came up with this new solution with Bloom, which we now implemented. The power is on the site. Of course, it took a couple of months of delay, but we're back on track and electricity is there. The project is still going. All the engineering teams of all the companies, of the contractor, Microsoft, ours, work together. There is a big respect between the teams. So, there will be actually, we plan to use more Bloom generation on our other sites.
By the way, this new solution probably will bring some additional power even on this site. We'll discuss what to do with it. Last but not least, even with this delay, you should know that we stayed in the range for our revenue guidance.
Thanks, Andrey and Arkady. Next, we have one from Ryan Lountzis from Morgan Stanley. How should we think about the level of demand in overall TAM for these shorter duration contracts relative to longer term dedicated capacity deals? Just trying to understand how we should think about where the mix between longer term deals and shorter duration ones are. Marc?
Thank you, Ryan. Great question. I think this does tie together a couple of things that we're talking about today, and I'm going to try to zoom out a little bit. I think you get the sense that we're trying to be market-driven. In other words, meeting the customer where they are. We're not making markets. We are the beneficiary of the demand that Arkady shared earlier. As we've shared in the past, we maintain an opportunistic approach, and we have since been developing the means to get a stronger set of pricing and term signals. Between our auctions, spot pricing, and the longer-range discussions taking place with our customers and prospects regarding their plans, we are able to adjust our terms and offers in the market to optimize our outcomes. Ultimately, we are looking to drive the most beneficial offer relative to opportunity in the market.
Today, we now have systems in place so that we can adjust our execution to optimize for these outcomes. What we're seeking is the ability to optimize between short-term consumption and long-term reserve. As Dado tied together earlier between my and Ophir's responses regarding prepays, the same flexibility allows us to switch between the offers in order to optimize short-term revenue versus securing CapEx for future build. In the fullness of time, based on Arkady's vision, we want to deliver this entire capability on a standard trading platform that allows us to algorithmically adjust our offers for all types of customers, from small self-service customers all the way to scale sales-assisted enterprises.
Thanks, Marc. The next question is from Tyler Radke from Citi. What are the key assumptions behind the preliminary 2027 revenue guidance? Dado?
Thank you, Tyler. Well, revenue in the guidance for 2027, $12 billion-$13 billion. We are quite confident regarding this guidance. In order to reach this revenue guidance, we obviously will need two parts. We will need to have the capacity on time, and this capacity should be ready to be sold. On the capacity part, we took a very, very conservative approach about the timing, which give us a lot of comfort. As a matter of fact, a significant part of this 2027 capacity will be owned by us and not by third parties, which give us control over the delivery. We believe that the capacity will come really on time, which may even increase the potential rate of the revenues that we have given as a guidance. If so, the revenue, we expect that it could be higher.
On the demand side, we have no doubt that this capacity will be sold. Clients are looking for capacity already for 2028, and prices for capacity in 2027 are going up, so we do not see any risk on this front. Really, overall, we are really confident with our guidance.
Thanks, Dado. We have a follow-up from Tyler. Customers committed more than $12 billion of TCV since the start of Q3, but RPO only increased 3%. How much of that TCV remains unrecognized in RPO, and what are the primary factors limiting reported backlog growth? Marc, do you want to take that one?
Certainly. Tyler, let me start by giving you some broader context in terms of how this quarter unfolded and how that relates to the momentum that we're experiencing. We've been saying for multiple quarters that we can sell all our current capacity at current rates. That includes, by the way, as well, the ability to be able to sell future capacity at current rates. What we didn't know was what's the right current terms. As we shared, and I answered earlier, we've been conducting programs to get a better understanding on pricing, including the auctions I mentioned and, more recently, our spot pricing programs. We've also been talking to customers about their longer-range needs and have a strong sense of the terms that are possible. At the start of September, we increased prices.
We added higher prepay terms and contract duration to a new set of capacity that we were releasing. As a result, since the start of Q3, customers committed more than $12 billion of contract value, including multi-billion dollar five-year agreements. In Q2, we shared that we landed four landmark deals. What followed was expansions by the same kinds of customers on longer terms, higher rates, and with more paid upfront. Several of the agreements signed after the quarter end and show up on the next quarter's reported commitments. Importantly, the momentum that we're experiencing has continued into Q4.
Thanks, Marc. We now have a question from James Kisner from Water Tower Research. Token Factory was a Day 0 launch partner for Kimi K3 in July, and OpenRouter reported in September that open-weight models accounted for 72% of its weekly token volume. How do you see the long-term split between open-weight and closed models, and is a shift toward open-weight a net positive for Nebius? Roman?
Yeah. Thank you, James. Just a reminder, we were Day 0 launch partner for much more models in Q3 and the beginning of Q4, including Meta, Reflection AI, NVIDIA, and others. We absolutely see the adoption of open models accelerated even in the last few months, and that happens to bring AI at scale with the reasonable economics and quality improvements driven by the unique data and business knowledge that the companies have. We build the full suite of the services to support our customers in this transition. We were always saying that the bigger diversification on the market would be, the better for us. This diversification comes from open models or other ways to build specialized models for the particular use cases. Again, why it's good for Nebius?
Because it extends the universe of customers and partners we can work with and the value we can create. More new products are built, more enterprises bring AI into the production, and we will continue supporting this adoption of the diversified models across the ecosystem.
Thanks, Roman. James Fish from Piper Sandler is asking, you plan to add more than 1 GW of computing capacity annually while taking a cautious view on commissioning timing. What are the main constraints between contracted power and customer-ready compute? What milestones should investors track to assess whether deployment is on schedule, and are you changing how you're thinking about timing and messaging around them? Andrey?
Yep. I'll take it. Thank you. First of all, I would just to say that the power and the timing connected to the power is just the main challenge for the industry and for us specifically. The big change for 2027 for us is that most of the capacity that will be delivered to our cloud will be from our own sites, and I believe that was already mentioned in a couple of answers. That gives us much higher confidence compared to the colocation projects because we develop the projects from the very start. We have the full control over the pace of the project in general. In 2027, we have Alabama, Missouri, two projects in Finland already being up and contributing to the revenues in 2027, and Pennsylvania that will be going up and probably contributing to the future revenues as well.
We are very confident in our executions on those sites. The other probably most commonly regarded bottleneck is the supply chain for the chips. I mean, GPUs, CPUs, network, and storage. We are in a very good position there. There was never an issue before with us, and I wouldn't expect that this will be an issue in 2027. We started working on that in early 2026 for 2027, basically on the R&D and on the securing the supply. We expected this to happen. If we talk about the new generations and new chips that will come online in 2027, Vera CPUs, Vera Rubin GPUs, and Groq. We already have all the R&D integration and tests and benchmarks already done. We are very good to go from the technical standpoint. From the supply chain standpoint, we have confirmed all the supply for 2027.
We are, again, here, we are also very confident in our execution.
Thank you, Andrey. Our next question is from Arsenije Matovic at Wolfe Research. Walk us through the $325 million of non-recurring revenue recognized in Q3. What drives that large amount of revenue recognition to be non-recurring in the asset-light model and executing that in the quarter going forward? Ophir?
I can take that. Actually, we are very excited about this achievement. We think that it was one of the nicer achievements in this quarter. We are all the time exploring and adopting various mechanisms and partnerships to bring more capacity online. One of them, as we said before, is the asset-light model. By that, we are bringing capacity online with the Nebius architecture. In this case, some revenues are recognized when we lock this capacity, the supply chain, and the architecture, and some of the revenues are recognized when the capacity actually comes online and being sold to customers. This is actually our first transaction, but as the transactions will grow, and they will become more regular and part of our business, then I assume that we will add them to our base model. Now, I think that's an important point to put about our revenue guidance for 2027.
As was mentioned before, we were very conservative about this guidance. For example, we took an assumption, a very significant assumption about potential delays in the provision of capacity, even though most of the capacity will be owned by us and we have much more control over the delivery times. So if we will deliver on time, then obviously things will look much better. The same goes actually also for these partnerships and asset-light model. We were conservative on that. We didn't add any potential revenues from asset-light model and other partnership for 2027. If these opportunities will come alive and will grow, it will obviously have effect on our guidance as well.
We have a few questions about the breakdown of AI natives versus enterprise, including this one from Rob from Baird. Looking at your new customer growth, can you walk through a breakdown of those customers, enterprise, AI native, and others? Marc?
Certainly. Thank you, Rob. We do not actually break out our results by segment, but I am happy to give the color that you are looking for. Enterprise momentum is picking up, but it is still well behind the overall potential that we all expect in the market. A great leading indicator that we are seeing in our results is our adoption by the leading data and infrastructure platforms that enterprises build on. As Arkady mentioned earlier, he shared a couple of examples where we have signed multi-year agreements with the likes of Databricks, Datadog, Snowflake, and Palantir. And with these, we are actually experiencing what they are experiencing, which is the enterprise adoption that is taking place around AI. Palantir might be a great example.
They selected Nebius specifically as their preferred AI cloud to go to market with our Token Factory solution and AI cloud capabilities in combination with their solution called Foundry, that is an enterprise data and operational platform that allows them to combine our solutions with open-weight models in order to serve sovereign AI requirements by major enterprises. I was able to actually join a boot camp of theirs recently, and it was really exciting to see the enterprise interest in the room around supporting the opportunity for enabling AI to automate and transform businesses. My confidence is very high about this type of opportunity, and we are looking for the same type of opportunities with the other platform companies that I mentioned earlier.
Simultaneously, we are actually seeing direct adoption as well with leading significant enterprise logos, as enterprises are starting to bring on the first non-hyperscaler relationships, oftentimes for the innovation that they are seeking to implement in their organization. My confidence is in future quarters, we will be talking a lot more about the buildup in enterprise in our business.
Thank you, Marc. Michael Donovan at Compass Point Research is asking, adjusted EBITDA margin stepped up to 47.5% from 41% in Q2. What drove it, and how should we think about Q4, given a full-year margin of roughly 40% implies a step back to the mid-30s%?
Thank you, Michael. Our guidance for 2026 is to be around 40% EBITDA margin. Look, quarter- to- quarter, the margin moves with timing. New sites carry rent, power, and people before they bill. The third quarter had revenue recognized at a point in time that does not recur in the same form. I would not analyze any single quarter up or down. We do not guide quarterly, actually, but for 2026, as I said, we expect circa 40% EBITDA margin. What I can tell you about Q4 is that it has the full quarter of capacity that we added in Q3, further colocation phases coming online, and the new ERC tranches delivered throughout November, and that it carries the costs of sites arriving ahead of the revenue. For 2027, actually, the direction is up.
Capacity on our own sites come online in the second half with no third-party rent, and the agreements signed in Q3, and since, at higher prices and larger prepayments turn into revenue. We are quite confident about our guidance for the 40%, around 40% EBITDA margin at the end of 2026 for the year.
Thanks, Dado. Nehal Chokshi from Northland is asking, what percent of the $12 billion of the TCV signed in Q3 comes from AI natives and large enterprises? Marc?
Certainly. Thank you, Nehal, for the question. The high majority of the $12 billion is coming from AI natives, but I want to speak a little bit more granularly about it. These are customers that, I think we mentioned this, but I want to reinforce it. These are mostly customers that started somewhere else on our platform, either coming from an auction or having already had deployed capacity for other workloads or requirements. What's common to them, though, is that they are all scaling. What's also really interesting is the revenue momentum that they are experiencing. These are customers that are making significant commitments against scaling revenue-producing opportunities that we are very privileged to be a part of and excited about their prospects and growth.
Thanks, Marc. We have a question that's coming in. What are the updates to the software stack given that's driving margins and acting as a moat versus competitors? Roman?
Yeah. Thank you for the question. Let me talk a little bit about Token Factory. Token Factory is priced per token and sold as a managed service with different SLAs from GPU hours infrastructure. It provides Day 0 support for the new model that we already discussed, advanced routing and caching capabilities, variety of inference optimization techniques like custom kernels, quantization, and more. It also provides post-training capabilities and inference optimizations around it. This is our relatively new service that is growing very fast from the low base now. We are not looking in the margin product by product, but as we run in our own infrastructure, we have much better cost structure than any inference provider renting GPUs. Also, as Arkady told at the beginning, that is not isolated business. Nebius is serving customers with entire platform in the way the customer want.
Q3 and the beginning of Q4 were the period of significant acceleration for our Nebius Token Factory. We activated our go-to-market organization around both large-scale training and inference workloads. Customers, a lot of customers, including Revolut, Higgsfield, Applied Compute, are combining different use cases on Nebius and benefiting from the flexibility of integrated platform. Token Factory is a Day 0 launch partner for leading open models that we discussed, and consistently delivers leading performance at independent benchmarks such as Artificial Analysis. What is even more importantly, we see the same in real customers' workloads. We help customers train and run specialized models at scale across different use cases, including personal AI assistant, code generation, long horizon enterprise agents. Token Factory benefits from running on Nebius infrastructure, with direct access to the latest and greatest GPUs.
As an example, we'll be the first to support NVIDIA Groq 3 LPX's new product. This progress is supported by aggressive investments in the team, including acquisitions we made, Eigen AI and Clarifai earlier, which are now fully integrated, and the most recent, Inferize. Inferize brings leading model loading capabilities, which solves the cold start problem, and it's especially important for improving GPU utilization with spiky on-demand workloads. So, this is the vertically integrated platform we are talking about. Everything runs on one platform. Customers can expand from one product to the next, and they can choose the optimal way to consume each workload with us.
Thank you, Roman. This concludes our call. Thank you all for joining us this quarter.
This concludes today's call. Thank you, everyone, for joining. You may now disconnect.