Nebius Group N.V. ($NBIS)

Perfect. So I have to state my name at the beginning for the record. My name is Tal Liani. And I am very happy to host Roman Chernin, who is Chief Business Officer Nebius, which is basically as far as I understand your position, it's a product manager. It's a head of product. It's a product guy.

It's different things, but I like to say that I do what needed now for the growth. And sometimes, it's more on the product side. Sometimes it's more on like go-to-market side.

Yes, yes. Got it. Okay.

It's a previous position.

And I just mentioned it, so we're not going deep into the numbers. I want to talk to you about strategy. I want to talk to you about basically where your product is heading, what's your advantage in the market. And let me frame the discussion first as I normally do in kind of this meeting. So selling data center capacity now is selling water on a hot day in the desert in middle of the day. So if you want to make it kind of even more dramatic. And the question is, what is the value that Nebius brings to the market, meaning is the growth today sustainable longer term? The question is, do you -- is there a differentiation between Nebius and [ Cowi ] and maybe the data center of Oracle and Microsoft and all the hyperscalers. So I want to ask you about basically it's the sustainability of the growth, but it's coming more from the product side. I want to ask you about your differentiating factors. So maybe with that, this is kind of an open statement just to frame the discussion. Maybe with that, you can talk about -- what -- how do you craft your product strategy? Meaning what are you trying to be in this market? And we'll take the discussion from there.

Yes, it's a very broad question, I think, but let me start from the customers because I think that -- you said like selling data center capacity. Like in reality, we don't sell data center capacity. It's our product that is built on top. And like we build product for the customers. And to discuss it like first, we probably need to structure how the market looks like. And we look at the market and the product that we built in the following way. So there are new customers, like everybody talk about, hyperscalers or super labs that need a lot of compute, but they need only compute. They literally don't consume almost any additional services and its scale deployments. It's not easy to deliver those scale deployments, but at the end of the day, it's most like basic, not differentiated service. Then on top of that -- and this is like what people call [indiscernible] compute. On top of that, there are much more -- much bigger population of the customers, let's call them EI native labs or neo labs, hundreds, maybe thousands of customers that need infrastructure, but they prefer to consume it in a managed manner because they don't have all the full stack of their own software, and they want to just focus on their research, they want to focus on their training tasks mostly. And for them, we built -- that was actually the first -- probably the first category of the customer that we serve, and the build out of what we call [ multi-tenant ] cloud. So it's -- again, it's managed infrastructure. But then there is the next layer of the customers. They probably -- it's much bigger population again. People who don't want to deal with the clusters. They don't want to talk in terms of Kubernetes clusters, GPU hours. They consume models as a service. And these are the people who build product. you can call them vertical AI products like Cursor of the [indiscernible] like in coding, I don't know, [ Harvey Ligor ] in legal, [indiscernible] in content, [indiscernible] in CRM and so on so forth. So those are the customers that even don't think in terms of the cloud, they think in terms of the models, they consume and most of them started with the closed models, then for many reasons that we can discuss, they diversify their consumption towards open source or specialized models. And they need the next level of the product, and we build a managed entrance platform called [ tuck-in factory ] for them. But probably, this is not the final stage of the market. Now we see that a lot of people are building agents and the agent is a new type of the application, if you want, if developers who came for talking, they take all the orchestration, all the burden of building kind of end product on them. People who build agents, they don't want to consume tuck-ins. They want to get the final results of like agent execution, the outcome of the edge. And probably like -- it's a little bit like speculative, but probably they will consume the AI computing their way. They will not choose the model. They will not compare tuck-ins from this model and tuck-ins from that model, they will have some new level of kind of obstruction or new primitives how they consume. And so answering your question, how we build the product? The essence of what we built is the AI infrastructure. If you want, the simple words is the AI compute. Obviously, it's not only compute, it's not only GPU compute, you have storage, you have like CPU compute and so on. But the essence is infrastructure. But we think about the product in a way to flow the customers' segments and customers' workflows and be always relevant for the next wave of consumption. So we could just have [ Beneta ] compute. But then we -- our addressable market would be limited to a handful of big customers. We could stay on cloud level and provide managed infrastructure, but then we could serve hundreds of whatever, first thousands of customers. It will stay on the level of inference. But then maybe it's like whatever, tens of thousands of customers. We believe that along the way, along the adoption to AI, they will be the market of tens of thousands, hundred of thousands, and we don't know developers, builders who will build on their level of abstraction and our product strategy is to meet them there, right?

And you talk about the full stack here, basically from compute to software and then deployment and integration and is it what's driving revenue growth today? Or is it more in the future? Meaning what you're seeing today, did the market start with just pure compute capacity and then everything else you talked about will come in the future? Or does it start already from [indiscernible]. Yes, it's

Yes. Absolutely. It's absolutely Happening. And talking about full stack, it's important to talk about full stack in upstream and full stack downstream. Downstream, it's about how you build the infrastructure and how you actually control your supply chain and cost structure and upstream is how you evolve your product offering, right? And definitely, we already see that, for example, inference is the fastest or fast-growing segment. And we definitely see that agentic workloads like starting, and we can expect that they will continue to grow. So I think that even though that the big part of the market is still like sitting is raining, that opportunity to serve inference workloads the opportunity to follow like the new growing customers gives you much more flexibility and gives you much, give us as a provider and the platform give us much more flexibility and optionality on how we build our customer portfolio, how we build our contract portfolio, and how we benefit from the motion on the market, like obviously, we know that the prices are growing, and obviously, more flexible, more flexible workloads like inference let us benefit from that. So yes, it's already a significant impact on the business, positive impact on the business.

Is there a difference between what you're offering and what the other neoclouds or hyperscalers are offering?

Yes, I think so. And again, it's important to define the categories because people call Neoclouds quite different animals, let's call it. There are almost pure data center operators. There are people who have data -- don't have data center and just aggregate [indiscernible] on infrastructure and just aggregate compute. There are people that provide some software layer on top and people who do nothing and just like do the better sales -- wholesales of [indiscernible]. I think like if you look at the market of the neo clouds, it's like -- we don't like to speak about others. We like to speak about us. But I think it would be fair to say that from this full stack approach, both downstream and upstream, we're probably one of those who are the most sophisticated. And that gives, again, downstream, it gives you a lot of advantage on how the economics work and upstream, it gives you a lot of optionality how you can work with the customers and what customers we can serve. And eventually, it actually gives you the same economical advantage because I think it's like told even publicly that we have 3, 4 customers for each GPU competing at it. And you can think about it that the more competition from demand side you have, the more lucrative kind of business you can do because you can pick the customer that actually value what we deliver more and have the better economics and create more value for the customer, which is also important. There are customers, again, there are customers that just need data centers. Okay, it's not just not our business. But there are those who want to focus on what they built in the product, and they want much more value from this -- from the provider. And this is probably the customer that will value what we do.

So if I generalize and I'll say -- the larger customers probably that's the hyperscalers probably want the lowest value that you can offer. And enterprises probably want to have the highest value. How do you balance between the 2? Because at the end of the day, you're selling to both. You're not selling to only 1 group.

So first of all, I think we told many times that in different occasions that we believe that our long-term business is in a diversified portfolio of the customers. I met these enterprises start-ups and grow up and more established companies. Even though we appreciate the chance and we cannot really learn a lot and gained a lot of working with the biggest customers of the market, which are hyperscalers. And from the business perspective, those customers drive the growth for us because making the -- making business with customers like Microsoft or Meta actually open up like much more opportunities to finance the rest of the business for us. And if you think what are the drivers for our business, this is obviously demand and demand is there. This is the capacity, like how fast we can build and bring online compute. This is the product that we speak a lot. And this is the capital. And the capital is a very important component, and we by growing through the large contracts and large engagement that we have with hyperscalers, we have ability to grow faster and finance more aggressively in the rest of the business. But eventually, our goal is to have as much of the business in the like diversified kind of real cloud business and not wholesale business like of the large blocks. And then within that kind of part of the business, we try to build a very, again, diversified portfolio. I say it many times, the more diversified. But I think this is like a real philosophy of what we do. because we want to have like a lot of optionality. And we diversify the customers from their architypes. We diversify the customers from their type of workloads, and we diversify the customers like from their terms, we have long-term deals, we have short-term deals. We have some spot capacity that we can sell them a premium because it's available right now, we can sell something in advance. So like -- and again, it's not only driven by our willingness to have diversified portfolio, but the fact that we have different offerings for the market, let us have this different customers and different contracts.

Right. As the workloads move from training to eventually inferencing, how does it change the economics of your company.

First of all, I think like no doubt that this motion is happening because training is one-off investments to build the product and inferences is that -- it's a recurring part. What is really exciting is when you're engaging with the customer, with the partner and you support their inference needs, you align with them because the better business of our customer, the more they grow, the more business we have with that. So it's like much more aligned, a much more aligned business model. Then our changes economics of us is it's a good question. Actually, again, if we come back to what I said about different layers of the product, when people come for the training infrastructure, it's infrastructure sale. So even though we provided in the cloud in a managed manner, most of the customers know what they want. They come for the particular GPUs for the particular time of the time spend and so on. In inference, it's much more flexibility on our side that we can extract value through the software. For example, we can optimize different workloads for different types of the chips. And we are -- not necessarily commit customers for a particular cluster or for particular even type of the hardware, and you can abstract it through the software. And on the practical side, for example, it led you actually extend the lifetime -- the valuable lifetime of the CapEx investments because the new chips come, you can sign the first contract, for example for the most frontier training jobs. And then when those chips going out of this like we first contract, you can still utilize them for the inference workload. This famous Anthropics basic contract is 1 of the illustration that space move their training workloads, the more advanced cluster. But the chip they procured like in the first [indiscernible], they are still useful for the inference. And this is just the illustration of the life cycle, if you [indiscernible]. I

Right. I understand. When you go to Agentic AI, how do you monetize agentic AI, when you go to agentic AI, explain how your business evolves with agentic A deployment?

Again, it's -- you can think about it as yet another layer of obstruction on top of compute. So again, [ bare metal ], you sell megawatts. Managed infrastructure, sell GPU hours, tuck-in factory inference, you sell tuck-ins, agentic you sell agentic [indiscernible], you want a task be solved. Then what's happening under the hood, you consume the ton of compute through inference, through the sand boxes through other calculations. But you abstracted from the end customer and developer. And so how you monetize it, still selling infrastructure, but it gives you as a platform, another layer of optimization capabilities because now, for example, like I told about inference when customers not choosing what GPU to use, I can optimize for the GPU. At this layer, not customer choose what model to use, I can optimize what -- from each model, how many tuck-ins to extract. And so it gives you the flexibility. And this is actually people ask, why do you build software? Do you monetize software? You not necessarily monetize software directly but you deal the software to better -- like to unlock the new use cases and give you more more lever of optimization for the customer as a result for yourself.

Got it. You made 2 acquisitions recently made acquisitions, Eigan, hopefully, I pronounce it properly, Eigen and Clarifai. Can you take us through them? What was the rationale behind the acquisitions?

Yes. The rationale is quite simple. So these 2 are great teams, quite rare talent in the market, both actually in the area of inference optimization and model post training. And they both from different angles, will accelerate our journey with our inference platform, first of all. So [indiscernible] is the team here in San cisco very research kind of driven the founders MIT, PhDs, and they did a lot of work around how you take model and extract more value from one GPU, in a simple word. Then Clarifai is another team, the core team on the East Coast, and they more focus on inference as a system. So you can think about it as like 1 capabilities, you have a model, you have on GPU and how much tokens you can generate. But I imagine you run scaled system, you run on thousands of GPUs for millions of users. And then you're not only optimizing how 1 computer works, but you have entire system optimization, how your cash works, how you orchestrate compute, how you efficiently after scale, like when the [indiscernible] or cloud comes, how quick you can get new nodes up. And then when this spike goes down, how quick you can scale down and so. So to Clarifai is more team that works on inference job on an inference problem, but as a system. And so combining it together and actually combining with our in-house engineering capabilities, we believe that we now have a pretty strong maybe one of the best teams to build inference as a big system. And again, it converts directly to economics because the customer needs tuck-ins with the best quality, best price and best like performance. And these teams will just improve the product and economics and accelerate the growth for us.

None of you, meaning you, the neo cloud companies, none of you speak about unit economics, but -- so I'm not going to ask you directly about unit economics, but I want to ask you differently. Do you think your focus on software? Do you think that focus on Token factory and all the other services, do you think that it translates in reality to higher margins for you versus competition?

Yes, absolutely. And again, it's very -- again, like a very simple kind of thinking. Let's even put aside, which is not true, that for different workloads, people are ready to pay different cost because of like different value that we can deliver. Let's think very simple. We are in the world of supply/demand balance. I spent quite a lot of time in advertisement. And advertising business is very simple. The the more hot options you have, the more prices you have. So imagine you have a product that can serve 10 customers in the world, and you have a product that can serve 10,000 customers in the world, probably, that will drive your margins up just because you will have better supply-demand more like for us as a supplier, the supply-demand situation. Even we've put aside that when people buy inference, we can do much more optimizations under the hood and sell the same GPU with the best economics for customer and for us, by the way, even forget about it. Even forget that like on large deals like hyperscalers like probably have more buying power than some -- just if you target much bigger population of the customers. Just only that gives you, you can assume, gives you a very significant push of your market margins.

Got it.

And that doesn't mean that we don't have all the other factors.

Right. I understand. How do you manage supply constraints? How do you manage -- even from a contracting point of view, you signed some -- your duration of contracts is much shorter than the competition. You go for short durations. But still, when you sign a deal and you need to bring capacity on board, how do you manage the risk of commodity pricing going up?

First of all, I'm happy not to be the supply person in our company. Being on demand side, it's the easiest job in the world comparing to being on supply. I think there are -- like there are a few questions in what you're asking. So the main constraint is still the capacity, like not clusters built, but data centers and connected and ready to bring GPU's online. And I think the most important move that we have there is we announced few -- we have capabilities to build. We are not only renting. Of course, we're renting temporarily because it accelerates time to value for us. But we have the team with the expertise of building sufficient data centers from the ground up from the grid field. And starting from late this year, next year, very significant portion of the new capacity that we will bring online will be in the data centers that we built ourselves. And building ourselves give us so many advantages, both on the cost structure, but also on the control and flexibility on when we bring capacity online. So this is the first factor. The second factor, if you think about us, we don't -- we are not dependent on 1 data centers. We are managing like our customers are living in the cloud environment, most of our customers. And we think about our data centers as a portfolio, diversification of the game. So we run like a dozen of projects in parallel. And maybe some of them will be delayed, maybe some of them will be even like not successful. It will not impact our ability to deliver to the market because we kind of oversubscribed in a way. This is the second factor. The third factor, diversification of the workloads gives you more flexibility. For example, inference workloads, training workloads and the biggest contracts even if we put aside the hyperscalers, the biggest training workloads require a lot of compute in one place. You need to large clusters. Inference is distributed workload. So you can manage your -- again, you can manage your portfolio of data centers in different regions, in different like timing, very flexible. So that's what's helping us on the key kind of key bottleneck, key driver of being able to supply. And then on the commodity side and like chip side, I think that we pretty much well said. So our biggest contracts are locked and supply for them is locked. And I also wanted to say that, in the current market, actually the -- I think the effect of like on the prices, from demand-supply situation versus the cost structure and like raising some of the components like the supply-demand situation is much more strong factor. They're like, again, the demand responding to the supply constraint.

Got it. We have 3 seconds left, unfortunately. I didn't leave enough time for questions, but we can take probably 1 or 2 minutes from the break. Anyone has any question, please raise your hand. No, Very good. Roman, thank you so much.

Thank you for questions.

Thanks for enlightening us.