BONUS: Nebius’s Arkady Volozh | On building in the agentic era | Spotlight On | AccelVC

0:00 The demand is so high, coming from so many directions. Whatever is built is consumed, and demand is totally objective. Everybody sees how much new value this thing generates. You can do either 10 times more things or 10 times cheaper, and it's a huge huge industrial uh next step. Welcome to Spotlight [music] on, a podcast about how companies are built from the people doing the building. One exhilarating decision at [music] a time.

0:33 Arty, I am thrilled to be back together discussing Nebius with you. You and I actually set out to have this conversation 1 year after our first podcast, but with everything going on inside of Nebius, uh I think this was the first window of time we could find. Just in the last 2 months, you guys have announced large-scale customer commitments, rapid capacity expansion, a deepening partnership with Nvidia, major financings, and continued build-out of your AI cloud platform. I promise you we will hit all of those today, uh but I do think that there's a a bigger story brewing here around what exactly it is that you're building. A story that illustrates why, when put together, all of these announcements are far more strategic than they may appear at first glance. To get at that story, I want to take us back to where we left off.

1:23 It's mid-2024. Nebius has emerged on the scene as the newest entrant in the AI infrastructure race. What is it that the market maybe got wrong or didn't understand about what it was you were building? I think back then, uh we had a clear vision. We started with a tiny 25 MW data center, and we had a clear vision that this huge This is a huge build ahead of us, and we will need to build a lot both in terms of scale and in terms of new functionality, new features, new tools which needs to be built for this new era. We saw our niche from the very beginning as helping people build AI. We have this experience. We were building AI in the previous 20 years and we built a big infrastructure for that both again physical and tooling.

2:18 And we realized that this in this new world there will be needed hundreds if not thousands times more than we have built before. And it will require a completely new set of tools. With our experience both in hardware and software and building scale scale systems, we left behind hundreds of megawatts of built data centers. At the same time, we are totally new startup on the market and we understand very well the life of startups the lives of enterprises trying to go into this new un- unchartered territory of AI. By understanding them them deeply, we understand what tools they will need. When we started almost 2 years ago, there was this vision that we need to build something very very big.

3:07 We should own the full stack. We should build through the whole stack. We need to optimize it with me which means that we need to own the data centers. We need to build our own racks and we need to build all the software layers on top of it from basic cloud to this all new services for today we can say inferencing or agentic, whatever it comes next. We felt that we are by our qualification, we are one of the few teams in the world who can build it both in scale and functionality. We understand what needs to be built.

3:40 And I think when we started, people just didn't look at us this way, but again, there was different times. People probably didn't understand well what's what kind of things we're facing. Now it's much 3 years later, people understand that whatever we we will build will be consumed. And we need to go to space for that, maybe even. And people also understand that it should be a new thing. It's not just old classical cloud and tooling reshaped for this new era.

4:13 It's completely new new set of tools for completely new set of users. It's not just software engineers who are using it now. And they need different tools. And that's what we're building. And big scale full stack optimized cost effective. Yeah, I think you know, unpacking that a bit, I think people may underestimate just how much depth is behind the team inside of Nebius. I mean, what you just articulated requires domain expertise from land and power acquisition through to hardware development of racks through to a large corpus of engineers going and scaling the software primitives on top of that infrastructure. Your team didn't start from zero. Could you tell us a bit about the team that you've built inside of Nebius and and why it's unique for such a young company? A year and a half ago when almost 2 years ago as we started, we were quite an unusual startup because we came with a lot of experience with 20 30 years of experience of building large scale infrastructure for machine learning before it was called AI. We worked with Nvidia many years before this company was created. Again, we started from a very small installation uh which we inherited just 25 megawatts. 18 years later, we already were running 10 times of this. And this year we're building we're growing again. Maybe this will be this uh hopefully big huge platform with hundreds of thousands of GPU of capacity and layers and layers of services tailored built from scratch for this new AI era uh for developers serving the workloads which are demanded by them now, not what was demanded a decade ago.

6:11 You mentioned the long-standing relationship with Nvidia. So, I want to transition us to some of these big announcements that we've seen from Nebius over the past couple weeks. Help us connect all of them and why don't we start with Nvidia? It's not just investment, just money for us. It's actually uh it's much deeper work which going on behind this investment with Nvidia uh where we together uh develop and optimize we help them optimize uh their systems. It's a lot of engineering work between the teams uh in the background. It's again, it's a sign of a deeper collaboration with Nvidia and we are extremely proud of course that they become uh our partner and shareholder.

6:55 But again, it's not about money, just money. It's good to have $2 billion but what we're building requires tens and hundreds of billions of dollars. Let's refocus to another big announcement that you guys have made which is a large commitment from Meta. The headline of that commitment is $27 billion but really I think what's most interesting is how you architected that relationship. Can you talk a bit about what that $27 billion really means to Nebius and what it means to the the overall platform that you're building for the cloud? When you see the news about this 27 or so billion dollars, you should actually understand uh that in this particular context there is actually two two contracts inside. One is 20 12 billion dollars when we really provide bare metal capacity to Meta for themselves to use. And for this, we're getting the second part of the contract, which is 15 billion dollars of backstop. What it means for us, we will be building we will we will be use uh actually Meta's readiness to off-take it, go to the bank, and finance this capacity for our own cloud.

8:09 much trial levels of the cloud, we will be selling on spot to thousands, tens of thousands of smaller uh clients. We're all watching those charts showing GPU per hour rates go through the roof. And I guess what I'm hearing from you is that 15 billion dollar contract is locked in at today's long-term rates, should you not be able to sell it. But what you're saying is you're going to take that 15 billion dollars of capacity and bring it to your core customer, that retail customer that leans on your software, services, tooling on top of the infrastructure, and pays what you call spot rates in the market, much higher rates. Effectively meaning that 15 billion could mean 50%, 100% more to Nebius over the life of that contract. Isn't this a similar thing to what we saw happen with Microsoft as well? Different but similar thing happened with our Microsoft contract last year. It was uh 17 plus billion dollar contract we where Microsoft actually got bare metal capacity from us, and we're building bare metal capacity for Microsoft AI lab.

9:15 But it allows us to go to the banks and use this contract to finance construction of our own cloud cloud. So, we use these big bare metal contracts not as our main line of business, but as the one of the ways to finance building our own cloud for for the rest of the market. That's interesting. You you have said that before that these large deals with some of the big hyperscalers are a great way to finance what we do. I I think getting off the blocks to be able to deliver capacity to small startups, startups that frankly don't have a contract that a bank cares to lend against has been a big kind of chicken and problem I think for the ecosystem. I know you have a number of thoughts around how many GPUs are actually serving that end of the market. You know, the core end of the market that Nebius is targeted. You mentioned how we got in there and how we're trying to find ways in there to finance our way through it, but talk a bit more about that core customer to us and and how we serve them. So we see the main market for us, the our main business model in serving the AI native startups and enterprises going online. And these are again thousands of different clients.

10:31 They don't have behind them this huge uh files of capital to build their own capacity, their own uh thousands of GPUs. We're building a cloud for them. This should be a big cloud. What we're building is actually even this year will be measured in hundreds of thousands of GPUs. So it's basically the same scale as the other big hyperscalers. It's a shorter term smaller contracts, but in large much larger amount of clients who actually need much more services from us, which is important. We just not just provide them with GPUs, bare metal. The big clients like Microsoft and Meta, they don't need our software stack, they have their own.

11:18 These guys, they don't have their own stack. They need a lot of help from us from us. When we're building tools for them, this is our main business. And this is what we're building and we're building it at scale. To be there, you need to provide a scalable, very efficient, very performant, cost-effective systems. That's what we're building. Two thoughts on the capacity side of things. One, Nebius thinks about scaling capacity maybe slightly different from some of the other clouds or neo clouds you might be compared against for this exact reason. Delivering a multi-tenant hosted cloud experience to the torso of the market requires a different type of capacity planning and scaling than you may see out of a large hyperscale deal on bare metal. So, I'd be curious to hear a bit about capacity planning when you think about your core end customer and how you think about whether we need a 5 GW facility or we need a 30 MW facility or both. Bring us inside the capacity planning of Nebius. To build what we want to build, you need to be very efficient and you need to start building from >> [snorts] >> scratch from land. You need to understand everything from land permits and electricity permits up to how to build data centers, how to build racks, what software to run. If you outsource it, first of all, it will never be efficient. You you pay all the margins to data center constructors, to rack builders, to other software companies. We build it through the whole stack. It's a full stack build, which gives you economical efficiency, but more more importantly, it gives you uh technical engineering efficiency. It is optimized to the last bolt. It's uh optimized on all the layers. It works together as a system. This is what we built. Again, it it should be built uh cheaply and optimally performantly, and it should be built in in scale.

13:19 Should be big and very efficient. >> You know, looking at the capacity today for Nebius, we have both owned locations as well as collocated locations. What I'm hearing from you sounds like the direction of travel is to to own more of our infrastructure from the land all the way through to the software. If you extrapolate this over the next few years, what is the footprint of Nabiax between owned and co-located? And And what is the purpose of co-located for us today? Well, we started when we started, we we took everything. We took even small colocation sites.

13:51 Although they're more expensive because we needed to start quickly. As we build, we build more and more of our own capacity. It's usually larger sites. It's hundreds of megawatts, sometimes gigawatt plus sites. What we're building, but not only again, as I said, we will be working with any capacity we could find. So, if could be several gigawatt scale sites, but it also will be a range of smaller sites. 10 megawatt here, 50 megawatt there, 100 megawatt there. It's a network of of sites. We will be working over this network.

14:28 Majority of this capacity will I think even this year will be owned by us. But again, we will be ready to work with everything which moves. So, let's pull the financing thread a bit because I know people love to talk about how you come up with tens of billions of dollars as a 30 to 40 billion dollar company. You know, 2 years in, you've got 47 billion dollars of commitments from some of the most creditworthy end customers in the industry. You found very strategic ways to fund the business to this point. I believe the the CapEx plan for this year on the upper end is about 20 billion dollars. Bring us inside the financing of Nabiax. Tell us where you are today and what are some of the levers you can pull to continue to stay ahead of the pack the way you have. All capacity we're building this year is financed, fully financed.

15:17 And we're now talking about And again, if a year ago it we were we were talking about billions of dollars for this year, we're talking about tens of billions of dollars. And we to to fulfill our programs for three and more gigawatts, we're talking about over 100 billion dollars already. The second question is how expensive is it? Because for a startup like us with all of our public Nasdaq public status and size, we're still a young startup for the banks working in a very risky new part of economy. If you want to get your financing for, I don't know, 10 12% probably it's you can get it. Again, depends on the size, but it kills all of all the economy and actually huge difference. People maybe don't understand the difference between 6% annual interest and, I don't know, 9% annual interest, 10% annual interest.

16:23 They think see it as a 4%, but on a 5-year contract, it's like 20% of of extra margins. It's a huge difference. So, we're looking for affordable financing in pretty big volumes, tens of billions, probably hundreds of billions of dollars. Very helpful. Very helpful. So, if we just take a step back, these announcements, embedded in these announcements, you have engineering partnerships to help with the software that's on top of our infrastructure. You have financing available to us to be able to build for the vertically integrated platform we hope to build over time. You talk about the layers of infrastructure. You talk about everything from land through to software. Can you illustrate those layers for us? Because I think a lot of these announcements give us the capability to compete at every layer.

17:11 And then I want to take us to the platform on top, but But I first want to understand the layers, because there's very few, maybe only the hyperscalers that have actually stated a vision to own all of those layers the way you have. So, walk us through what those layers are. It starts from LPS, land, power, shell. Means land with industrial pyramid, power reserved. Is it from the grid or self-generated? And the shell the the data center itself. And it's not just the shell, not the walls, it's all the electrical equipment in it. And this first layer actually costs, let's say 20% of total capex.

17:55 And then you need to put uh the system aesthetic gear the racks. Uh we build our own racks. Then there is a layer when you just uh let's put call it a bare metal service when you put on, I don't know, Kubernetes or layer layer on top, and you provide it usually in big chunks for long term to people who have their own virtual systems. It's big tech companies. Bare metal services, usually wholesale volumes, big volumes, to several customers. That's the business of new typical new cloud. We are not there. We we are building layers on top. Then on top of this bare basic uh cluster software, there is a platform layer or infrastructure as a layer. Just basic services which you would find in any classical hyperscaler, AWS, Azure, GCP, Oracle. On top of this basic IAS level level, there is platform with higher level services.

19:01 And again, you can choose to to to sell bare metal at wholesale prices, pretty pretty low prices, but big volumes, or you can go in kind of retail on spot market and start offering services which those companies need. They don't have They don't develop themselves. Be it basic services, like cloud services, and higher in the levels could be inference, like our talking factory, or agentic services, like search agentic search with Stratio, which we recently bought. These services are completely different from what a classical cloud would provide 10-15 years ago. It's not just different services. They are What is really different is the end user of the services. What we are building, we're building it from scratch and we're not building it for the previous generation of customer. We're building it for this new generation of customers who actually have completely different requirements. They don't need a hundred of different tools to stitch together in their cloud. They don't think in these terms.

20:14 They think agentic. They are usually multi-cloud. They don't care actually where compute is happening, where different services from where different services are consumed, because they don't do it manually. It's the agents who do it. And the agents, they are optimizing efficiency. These new users are choosing the platform which is measurably better, more reliable, more efficient, cheaper. Yeah, I notice you shy away from the term neo cloud.

20:46 And in many ways you do it because of the five layers we talked about. Uh when you actually own the land, the powered shell, the rack, the software on top of that, and not just delivering Kubernetes in a bare metal experience, but delivering inference post-training, fine-tuning on top of that cloud, now you look like an >> agentic services. Yeah, additional services. That's That's the AI native hyperscaler that you're talking about. But, I think what's under appreciated today is how you actually compete with hyperscalers because neo clouds aren't your competitors.

21:19 You're well beyond the scope of neo clouds. Hyperscalers are your competitors. Hyperscalers have been building a platform for over 10 years. And, they have many many software tools primitives that sit on top of their infrastructure already. What you're articulating with this agentic world and today's AI native developer is potentially a window into how Nebius becomes a true competitor to today's hyperscalers. Maybe double click on that experience for a developer today as far as why they would choose Nebius over one of these embedded large proven hyperscalers. Why Why work with Nebius? When we're talking about competition here, it's not a red ocean. It's totally a blue ocean. It's a totally new area.

22:07 They need to move there as well. Also, when we talk about competition, it's not like we're we're we're going to compete and kill any of them. They will continue to exist. They will be dominant. There is an opening now. We see it as an opening. There is something where they don't have any protective moat. These new requirements, these new users, it's all new to the market. And, we're talking not about taking their business. We're we're talking about building new business together with them and talking about the market shares will be changing. And, it's again, it's not a consumer market when the winner takes takes all. It's a B2B market. Prices are important.

22:48 Efficiency is important. Everything is important. And, there is an opening now, which I think a market share available for players like us. I don't see too many players it of as if now who are building A the scale, B the depth or heights of the uh of the platform. It's interesting how many how many people in the market talk about delivering a cloud and what that is actually started to mean to the market.

23:19 And I think your definition of a cloud is about as broad as anyone's in the market in terms of what you want to cover. We actually do compete, I would say, at every layer of that cake. And there are very large businesses being built at every layer of those five layers of infrastructure. And as you build higher and higher, you eat the margin of each layer. So, although there are very large businesses built there, you aspire to take all of that margin all the way up to the top. Let's stop at the top for a moment because you mentioned something interesting around what has emerged as potentially one of the largest revenue opportunities in the world of AI, which is delivering inference. This is actually getting value out of all these models we're training on these huge clusters. You have Token Factory.

24:03 What is Token Factory? And what what is your right to win in the world of inference? There's a range of companies who provide inference services. People are training models for then these models to be used for something for to generate value in other industries. And using models is is inference. And theoretically, in the beginning, most of capacity was spent on training new and new models, but eventually there will be more and more usage of these models. You more and more inference. There is Frontier Labs.

24:38 We have now what? Five five of them in the western ecosystem. Uh which provide inference services, be it Anthropic, OpenAI, or others. Uh and there is a range of newcomers who provide inference on uh open weights uh models. People like Together AI, Fireworks, uh Base 10. There's a range of them. Frontier Labs, of course, they're much bigger and their limiting factor is capacity.

25:13 Uh they told many times that they their revenues would grow even faster if they had more capacity. And they're getting and buying capacity from everyone. These other players working with open models, their main main challenge is again to uh to get capacity. They don't uh build their own capacity. They're buying it on the market from hyperscalers and new clouds. Since uh the capacity becomes more and more scarce resource today, there is not enough capacity. Everybody talks about it. We cannot build as much and as fast as needed.

25:49 Uh their business is limited. They cannot grow fast as fast as they could have uh grown if uh they had this capacity. They don't have it. We build our our own capacity for this. And we have our own capacity for inference. It's not just having this capacity. We also optimize the full stack down to to the ground, to the to the racks. Uh and provide more efficient uh inference. Again, this recent acquisition of Agent is to provide even more efficient uh faster, cheaper, low latency, uh high throughput uh models. I think this is why we're growing in inference so fast and it quickly becomes the next layer of our business. Again, it's not selling bare metal or or cloud, not selling GPU by hours.

26:43 It's a different measurement. You You sell tokens, and people don't care what you run it on. They're buying tokens, so tokens are different quality tokens, uh cheap price. This is what they they they interested in. This is I would say that it's a business of 2025, 2026. The next coming business, which is happening now, is agentic. Agentic workloads require different services, different environment. Our recent acquisition of Saville is trying to provide services in for the first layer, and there will be more and more services coming, which we build ourselves, and we will be partnering with specialists in all those fields. This year we will build this next layer, agentic layer, to our cloud. Yeah, I I like talking about these services because I think it highlights a lot of what we've been talking about today. So, first, as you mentioned it, you control capacity. You also control engineering from hardware through to software. But, the other point which kind of encompasses your vertical vertically integrated AI hyperscaler narrative is margin. So, if you're an independent inference provider in the market today, hypothetically, you could be procuring a GPU from a provider who doesn't own the data center or the racks. So, they're leasing the data center and paying someone for the racks.

28:04 And that data center operator may not own the land. They may be leasing the land. And so, as you look at every layer of that cake, again, there's margin to be optimized. And where it shows up most visibly is when you can actually deliver the service on top, and then you control that whole chain. And that's a a in a lot of ways the summation of the uniqueness of that vertically integrated product that you guys are building.

28:25 Vertical in integration brings you margins. It brings you also efficiency and speed because you control the layers, and it brings you scale, size. You cannot buy a lot uh trying to buy it here is uh you need to have a pipeline of construction and know that the next gigawatts are coming. And if people stay take a step back from this conversation, uh the deals, the capacity, the platform, what do you think people are still missing? What do you want people to leave today understanding about Nebius and where we go from here? It's a play in two di- dimensions. It's a scale in a depth or a product game. In scale, we're building uh again, there is just big tech companies and just couple of new clouds who who build in this scale and we're talking again about hundreds of thousands of GPUs and potential millions of GPUs available available for public use. And then the second dimension is what you're building. And what we're building is again, the cloud for this new Today, I would say agentic world world world, uh what it will be next year, we'll see. But definitely, there's every year there's something new happening here which you need to build. Uh which opens up completely new, much wider range of clients. And a year ago, you and I were asking ourselves, who's going to buy this infrastructure? Who are going to buy these GPUs? Now, we're demonstrating execution at at quite a big scale and we're asking ourselves, how do we bring this capacity online as fast as possible? What are we going to talk about a year from now? I think we will still be talking about capacity. I don't know what will be in 3-5 years.

30:14 For now, demand is so high coming from so many directions. Whatever is built is consumed. You can sell everything on bare metal or you can sell on the higher level of uh production somewhere. And I think in a year from now, it's still be will be a constraint, capacity, how much the humanity can build to serve this new demand. And demand is totally objective. Everybody sees how much new value this thing generates.

30:46 You can do either 10 times more things or 10 times cheaper. And it's a huge huge industrial uh next step. Also, I think we will be talking about some new applications which emerged which we don't know today. >> [snorts] >> Uh which will require new tools and the new audiences of users will be coming to the field. Arkady, it never ceases to amaze me when you and I get to sit down and you articulate how grand the vision is for where Nebius is going.

31:19 But, you and I have also talked about how hard it is to convey that to the market and to people that follow our story. And so, I hope today played a small part in being able to help you to do that. So, thank you as always for everything you and the Nebius team are doing and we're excited to watch the progress from here. Thank you.