Fantastic. Well, hi everyone. I'm Alex Duval. I head up the Europe Tech Hardware and Semis research team based in London at Goldman Sachs, and I'm delighted to be joined by Arkady Volozh, Chief Executive Officer at Nebius, and Marc Boroditsky, Chief Revenue Officer. Thank you very much both of you for being with us. Thank you. Thank you. Great. Before we get started, just like to state that this conversation is not intended for the media and is off the record. Fantastic. L et's get straight into it and perhaps start, Arkady, with a bit of a discussion on AI compute demand. One of the key investor debates has been around how durable the current AI infrastructure cycle will ultimately prove to be. Can you talk about what you're seeing in the market today, and what does that tell us about the duration of underlying AI compute demand? To what degree has your view beyond 2026 strengthened? Well, I doubt I will say something new here. Everybody sees what's going on with the prices. It's obvious that demand today is much higher than the real world can build for this demand, and the prices always reflect this unbalance. What's going on is that AI is creating a lot of added value in some way in real life and in the industries, be it coding or recently security or office work and many other areas which are to come. There is a lot of value created there, which goes down to people who provide tools to serve this demand, who provide models, which goes down from there to people like us who provide compute and the cloud to serve these models, and then going down to people who build the infrastructure, which we also do. Demand is huge. It grows, let's say, times a year, if not tens of times a year. The physical world is limited. It cannot grow 10 times a year. That's why we have this imbalance. We probably have said that a couple of months ago, that if we wanted, we could sell all the 2027 demand today. We have much more ask than we can physically serve. We do not pre-sell much. We are focusing on free capacity, which we will be selling later. We're building capacity which will be sold later. S ince we last publicly spoke, actually this sign of this demand went even further. Now we receive orders for first, second quarter of 2028. People are demanding tens of thousands of vGPUs and GPUs now. W e see demand today as unlimited. Nothing is unlimited in this world, but for now, the demand in all the visible perspective is much higher than anybody can serve. For how long it will go, we were talking about 18 months. Now I think we can easily talk about 24 months, maybe more. T oday is the situation that AI is working. AI creates real value in real life. It needs to be served. We just, as an industry, we can't build as fast. That's why this pricing grows. Fantastic. It's amazing how fast the industry is evolving, and it's also fantastic to see how Nebius is evolving. I think you announced today a partnership with Palantir, where Nebius was named as a sovereign partner. Could you explain what that sort of sovereign partnership entails, and which customers you'll be serving through that? Particularly, it'd be great to get a sense of why Palantir chose Nebius. Well, what Palantir calls sovereign is sovereign for the companies. The companies, they use AI. By using AI, they produce a lot of data. These data go back to the models, improve the models, produce even more data, and this loop continues. The companies have actually today a choice. They can go to commercial models and start feeding their data back to them and improve their models, or they can actually use open- source, open- weights model under their own control, feed data to these models, keep improving these models for themselves, not sharing that data with the whole world. The practice shows that if you do this with generally open- weight models are lower quality than commercial models. C ommercial models are very universal. They are expensive. Open- weight models could be trained with this data in repetitive loops, so that in this specific narrow domain of this enterprise, could be trained better to a higher level of intelligence than general models. There is a lot of cases in our practice which it works. Just recently, one of our clients, Shopify, there was a famous tweet of CEO of Shopify which showed that they used open-weight models, trained it with their own data repetitively, and they achieved the quality which is higher than they had with GPT-5 and 6. O pen model, your own data, repeated many times in the loop, better quality in your domain. That is what enterprises should be doing, and by this, they do not share their data. They keep their data. They do not train the whole world, potentially their competitors. This is what Palantir, actually, and NVIDIA as well, and us see as the vision as the main mechanism how enterprises should be using their AI. To implement this, Palantir have instruments to create this loop, open models data back and forth, but they need infrastructure on which it could be run. They used to use hyperscalers for that, sending data there. Their clients do not like it because, again, their data is shared with everyone. They needed a controlled stack which could run open-weight models, which is controlled down through the models to cloud, to racks, and the data centers. Palantir can control the whole stack. We provide this, and the synergy is that we provide this whole stack up to our token factory, where we run a variety of different open models. They take it, add their tools on top of it, their customer base. For us, it is an excellent channel to enterprises. Together, hopefully it will be a nice tool to serve the market. Fascinating. Perhaps, a question to follow up on that with Marc. Palantir's commercial install base is obviously quite different from the sort of AI- native customer base that you've garnered hitherto. Perhaps you could talk a bit about how this fits into the enterprise customer base that you've been building. Well, as Arkady just shared, there's a very powerful value proposition that we share with Palantir with regards to how we believe enterprise adoption is likely to take place. W orking together, combining the full stack capabilities we have and the tooling that they have, and more importantly, the credibility and momentum that Palantir has in the enterprise, we're confident that we're going to be able to jointly develop opportunity that goes well beyond the existing market adoption in the enterprise, and actually will accelerate our overall enterprise business and contribute to the customer diversification that we have at Nebius. Fantastic. I guess another one for you, Marc. Obviously, NVIDIA talked about 70% growth in fiscal year 2028, and they talked about that being supported by demand across Blackwell and Rubin. I think it'd be really interesting to get your sense about what you're seeing on the customer demand side, both the newer generations of chips, but also the older chips. Well, like what Arkady said with your very first question about the overall demand environment, that holds, that demand continues to outstrip supply. It's actually growing faster than supply. T hat might lead you to believe that older chips are being used because people have no choice. There's a little bit of that, but the dominant experience that we're having across older chips is that the workloads that most of our customers have on those chips are actually properly aligned to those chips. They're simpler workloads. They're things like RAG infrastructure or text prediction or image creation, and the applications that those customers have built or the services that they're running, they understand the TCO. They have confidence in the hardware that they're using. It's reliable. It's vetted. They're satisfied with the kinds of reliability and performance that they're getting. As a result, those businesses, they grow. They're growing with us to the extent that we can supply them the older chips. Likewise, we actually have a list of new customers that are looking for any of the older generation chips that come available. V ery robust demand for the older ranges. On top of that, obviously, requirements continue to expand, and with them, the interest in next-generation chips are expanding. We're seeing that across the Blackwells. We have a very robust level of interest and adoption that continues to be at a multiple of customers to chips that we have available that actually projects out as well to Vera Rubin interest. While I can't confirm or deny the growth expectation that was shared by Jensen, I can say that it's extraordinarily robust in terms of the demand we're seeing across the full range of chips. Excellent. If we perhaps pivot to talk a bit about the landscape in terms of hyperscalers and neo- clouds. One thing that was interesting in the last few months, we saw Meta talking about exploring ways to monetize AI capacity, and talking about a cloud offering. Arkady, I'd be really interested in your perspective on what that means for Nebius and how you think about potential for large AI customers potentially to become competitors. H ow do you think about Nebius' positioning? What do we learn from this Meta announcement? How should we think about the market developing as more supply potentially comes online? You said cloud offering. What's. In inverted commas. What xAI and Meta. Yeah. xAI started doing, Meta didn't start doing yet. They might. They see the demand. They see that whatever they build for themselves, for their ecosystems, if they do not use it right now, there's such huge demand for this capacity, they can offer it to the market. When they offer it to the market, they actually offer it on bare metal, basic terms. It's on the cloud. That's why they sell it. We heard that xAI was selling to Anthropic, to Google, recently to Azure. People who actually take this basic level, bare metal, and then put their cloud and services on it, and then resell it to end users. This is actually what we are doing, so potentially we could be helping them to sell their unused capacity as well. We would definitely take whatever Meta does not want to use, because we're stranded on capacity. There is not enough capacity. If they could put more capacity on the market, it's good for the industry, it's good for us. We could take it and repackage and resell it to the end users. If we talk about total volumes, just let's put it in context. The total amount of capacity that came to the market a year ago was in the range of, I don't know, 5 GW and 10 GW. It will be probably 15 GW, 20 GW next year and 2028. What this xAI or Meta can add, it's several gigawatts on top of it. xAI talks about 6 GW, 8 GW additional. It doesn't change the market radically. It doesn't help the market grow 10 times. It's another 20% here, 30% there. It's good for the market to have more capacity, and if we put it in scale, it's not so much. It doesn't change, it doesn't kill the market, it doesn't change the market dramatically. It's just good to have. First of all, it's useful capacity, and we could actually help utilize it better through our cloud offering. Thanks so much for putting it in context, Arkady. Maybe one for you, Marc. Since we are talking about the market structure and how that works, I believe you have talked about three different buckets of customer types and the different kind of pricing you could garner there. It would be great if you could just recap that a bit for the audience. Why each of those buckets exists, and how do you think about calibrating between those different opportunities? Just to be clear, the three different types that Alex is alluding to include short-term. These are agreements that are typically three to six month in duration. They are opportunities where we are potentially able to secure premium pricing around specific types of workloads that I will share a little more detail about in a moment. There is what we call medium-term agreements. This is the one to three year agreements, and actually represent, in our point of view, the lion's share of opportunity in the market and some of the most important and diverse set of customer opportunities. Then thirdly, there is the longer-term agreements, which are five years and longer in duration. These tend to be the agreements that the industry, and we have done historically with the hyperscalers. When we have done them, it has been with the explicit intent that we are looking for the capitalization benefit in terms of being able to fund the expansion of our core business, which takes us back to our primary objective, which is the medium-term agreements. Medium-term agreements and the kinds of customers that we are talking about here run the range from the AI natives that we have a right to win with. This is the heritage of the company. We are AI natives ourselves, and we have built an experience and delivered the type of technologists that these customers are most likely to adopt from. These are the Higgsfield in media and entertainment, or a HeyGen in media production as well. These are companies that have built their apps on us, and they are scaling with us. There is also the larger- scale AI natives that are already producing revenue. They are into scaling out. They are looking for longer-term visibility around the capacity that they need. They are signing longer agreements, and they are looking for greater cluster sizes. Also in the same range of contracts, the one to three years, we are seeing digital natives like what Arkady mentioned earlier, like Shopify. Companies that are, again, looking for a reasonable duration of commitment to be able to fuel their requirements. We are also seeing enterprises adopt in this range. It is just the beginning of enterprise adoption, but it is a very important dynamic, and an important part of our market. We are actually prioritizing a portion of our execution, like we talked about earlier, around the Palantir agreement, to make sure that we are able to start building towards the enterprise. Because we believe, ultimately, enterprise will represent the majority of the opportunity in the market. Fantastic. Maybe one for you, Arkady, on pricing. I think it was fascinating to see at the last results you talked about this capacity auction. It would be great to get a sense of what you learned from that. What does it mean for how you are going to price and sell capacity for 2027? How should we think about the durability of the pricing in future years as more capacity presumably will come online? Well, when you have this imbalance between demand and supply, the best price discovery mechanism is something like an auction in one way or another. We tried to do it manually first to raise the prices. It did not help. The demand is still there. What is the best mechanism to discover what the real price balance should be? We ran our first experimental auction for the Blackwell processors a month ago, several weeks ago. We achieved results when we got 15%-20% higher price than we ever sold before, after all our price raises. We liked the experience. More importantly, our clients liked the experience. The guys who won this auction, they tweeted that they are happy that they were able to get guaranteed capacity through this mechanism. This is the only way when there is a deficit of resources and you need it, what is the way to get it guaranteed? Go in price competition and win. They won the auction, and they were happy about it, and we are happy about it, and probably it is the right way to find this right balance. We will do more of this. We are told that we will put more capacity on the auction. We now will start experimenting with our longer tail clients, self-serve clients. We have hundreds of thousands of clients which are self-served. Small client with preemptive mechanisms. We will go introduce something like the auction there. Probably this is the right mechanism to actually work with more regular and bigger clients. At least we should use some mechanism like the auction to discover the price, and then based on this price discovery, we should offer them discounts or upsell from that prices depending on different circumstances for these clients. Auctioning is the best way to find the right balance in this supply-demand situation, which we are now in. You touched on competition, so maybe one for Marc, just more from a Nebius perspective. Obviously, at results, you talked about those four landmark wins. It would be great if you could just remind us a bit about what you won, what enabled you to win those, and then how we should think about the pipeline from here. Is there more to come? Certainly. Actually, I realized when I answered the last question, I forgot one category. I am going to rewind just for a minute. I did not mention what is going on with the short-term agreements, which I prefaced I was going to come back to later. We are setting aside a small amount of capacity for short-term scale training runs. This is a very unique and finite part of the market, and we are going to continue to make that available. These tend to go out at a multiple of the ARR per megawatt that we see with our core or medium-term agreements. These four wins that we talked about last quarter fit in that core segment of our contract base. These were averaging $1 billion a piece. These were significant wins by new customers to Nebius that were looking for scale capacity. They obviously had incumbent suppliers, was highly competitive. Those wins were ultimately won in the POC. It was not about price. It was actually about the reliability and performance that we delivered in the experience. As a matter of fact, one of the customers actually even called out the fact that the POC, according to their technical team, was the best technical experience they have ever had with any supplier in the industry. G arnering significant confidence in the technology that we are delivering, then more importantly, also the humans that were part of that delivery. The expertise that we are able to provide them gave them confidence about the kind of experience that they are ultimately going to have. These wins have led to, obviously, the scaled relationships that we have. M ore importantly, we are partnering with these customers. They all have additional requirements. We are already talking to them about the next cluster that they have. In some cases, they are looking for additional GB200s. In all cases, they are all talking to us now about their Vera Rubin platform requirements. A lready set up for additional expansion with them. We also have other similar customers in our pipeline that we are pursuing with similar types of opportunities that could ultimately contribute in the same way that these four wins did. Fantastic. Staying with you, Marc, you mentioned there about new types of customers. Clearly over the last 12 months, 24 months, you've had this period of diversification. It's really not just about hyperscale at all. Obviously, there's a place for that, but you calibrate all these opportunities. Can you talk a little bit about how the mix has been evolving, and how you think that customer mix will evolve as we go forward? Certainly. Today, as I mentioned earlier, we have the right to win with AI natives. It's a very important part of our business. What we are experiencing is that we're starting to move up to larger and larger- scale AI natives, like the four deals that I mentioned a moment ago. That experience that we're gaining with those types of customers is actually building the profile and reputation with digital natives. Like the Shopify deal that I mentioned earlier, we're getting more and more access to critically leading digital-native customers where they're embedding AI capabilities in their offering, utilizing AI in order to be able to deliver a better version of their service. It's a very exciting portion of our business because those digital natives are not only scale, but they're also oftentimes supplying the enterprise, which also now is leading to our opportunity to supply the enterprise. There's been a significant uptick in our pipeline with customers that are platform providers to the enterprise. Some of the most important platform players are looking at us as the first non-hyperscaler supplier, and they're looking for two reasons. They're looking for the obvious ability to gain access to capacity when they need it, but they're also looking for a partner that can innovate with them to be able to support their requirements across their full needs, not just training, but also being able to support them with inferencing and post-training. Recognizing what we're building allows them to have the kind of capabilities that, as a matter of fact, we described a little while ago for Palantir. This is opening up a market opportunity for us, and then projecting out to the enterprise, where we're seeing significant traction in key verticals. We have a very focused set of vertical initiatives around physical AI, around healthcare, life sciences, around media and entertainment, around retail and financial services. We're just starting to get traction there, and we're excited about the progress that we're making. Aside from the different use cases and the different types of customers, one thing I think Nebius has talked about, Marc, is how over time you're going to have more and more of these value-added services. I think you've talked about how that can help with margins. Can you help us think about how large that can be over time? The evolution in that sense. It's a really important part of our business, Alex. What Arkady mentioned a moment ago in terms of the full stack, a lot of folks look at the industry, and they stop at the hardware. They think of the market that we're in as all about just compute. The reality is when you talk to customers, and especially as you talk to customers that are more and more, let's call it large- scale, their desires is actually for a much more full- stack experience and even more abstracted services. Today, when we go sell a customer that might start with us on compute, looking for model training capabilities, we are oftentimes also able to position our inferencing. The combination gives them significant efficiencies and advantages because they're going to be using the same platform where we've built the solutions together, and they can be used to compute in the future, between the different capabilities. Likewise, we've added to our platform the beginnings of agentic capabilities through our acquisition of Tavily, which is an agentic search solution. The conversations that we're having with AI builders, agentic search is a cornerstone technology necessary for them to build the agentic capabilities they have. We're going to be adding to our agentic capabilities. Right now we're delivering blueprints, but we'll be building out more and more solutions to support agentic requirements. As we project out into the future, the vision that Arkady has is us becoming a hyperscaler. We have the unique front- row seat of being able to watch what's happening with the customers that are building on platform in terms of things that they're doing that could potentially be turned into products, and the things that aren't being built, that they're turning to us to help them to build in order to bring them to market more efficiently. It's a unique situation where we're confident that we're going to be able to project into adjacencies that have more value creation and obviously, opportunity for us to do value capture and margin, and add margin to our results. Talking of value capture, perhaps another question for Arkady. It was really fascinating to see the announcement about your new asset-light model, where I understand you can actually partner with other infrastructure players. Can you talk us through how that works? How does that expand the opportunity and what do you see as the benefits when we think about growth and profitability? With this demand, the main impediment in supply is how much physically we can build, how fast, then how much we can finance all this build. Because one thing is to build a data center, another is to fill it with racks full of GPUs, and tens and hundreds of billions of dollars in our scale. W e develop our own network of capacity of data centers, which we build, and this is our natural pace of growth. Recently we thought that probably we could grow even faster if we partnered with people. There's a whole new class of people who came from electricity companies, from data center companies who are trying to go up the stack, understanding that there is margins up the stack, and they actually need help. What they have, they have physical capacity, plots, electricity, and many of them have access to cheap financing. We realized that if we partner with them, we provide them with our knowledge. They don't know what to do with it. We can bring them the knowledge, how to build the data center. We can build them our racks. We give them our racks. We can give them our software stack, and most of all, we can give them the supply. All of our go-to-market organization, we have so many customers for that. I t could be potentially a very good partnership with companies who have access to capital and who have their own physical plots. We announced this line of business, this partnership program, a couple of months ago, I think. We've got a long list of companies wishing to deploy, to try to do something here. All of this, each line in that huge spreadsheet is a separate big project. We're talking about capacity coming online in 2027 and later. We're working with several projects like this, and this will be a non-linear way for us to grow. One thing is we are building ourselves, and in addition to that, we'll probably partner with companies who have access to capital and physical resources to build it, which will be helping them to build and use our client base. Great. Maybe, I think we've got time for one last question. I think you talked about the physical limitations, and obviously power is really interesting as well as land. I'm just curious how you think about where you effectively deploy your incremental capacity, and within that, how you think about behind-the-meter opportunities versus perhaps other types of power. We work with all kinds of plots and power we can get. Anything which moves, we will utilize. It's never enough. Again, demand is so huge. Sometimes we build on the plots with electricity, which is ready. Sometimes we need to get connected to the grid, and sometimes we can just use our own generation. We buy generators. Recently we started partnering with a company called Bloom, which is hydrogen cogeneration, much more healthy. We have just like companies going up the stack from the ground, we have to go down the stack, from software to racks to data centers, and now sometimes to building our own generation, including our recent thing with Bloom, but not only. Whatever we can build, as fast as we can build, everything will be sold. This is the limit for our growth. That's why the more we build, the more we partner with somebody to build, the better for us. Brilliant. Well, on that note, Arkady, Marc, thank you so much. A fascinating discussion, and thank you for joining. Thank you, Alex. Thank you. Thank you.