Roman Chernin: Cofounder at Nebius AI on how they've built Europe's $25bn AI Infrastructure giant

0:00Hello and welcome to the Scaling Europe show. I'm Seb Johnson. Thank you so much for joining me. If you like what you see, please like, comment, subscribe. The more engagement, the more visibility, the better guests I can get and the better coverage I can provide of the European tech ecosystem. I would also like to say a huge thank you to my sponsors, Checkout.com, one of Europe's leading fintechs, and it is where the world checks out. And SurrealDB, the multimodal database for AI agent, which is partnered with thousands of the leading organizations, including the likes of NVIDIA and Samsung. So if you like both of those, check them out in the links below.

0:32Thank you very much. Hello and welcome back to the Scaling Europe show. I'm here with Roman, one of the co-founders and chief business officer of Nebius. Nebius is such an interesting company. It kind of was born out of Yandex. Yandex was obviously the big Russian company. I think it was often called Google for Russia or the Russian Google. The Ukraine war came into play and as a result, Yandex had to sell off and offload all of its Russian assets, property, infrastructure, got delisted and then came back as Nebius. Nebius is now kind of like an AI infrastructure company building the groundworks, the infrastructure to enable the enterprise companies and AI companies to succeed at scale.

1:14So yeah, I'm here with Roman. How was that, Roman? Was that an accurate reflection of the journey so far? Yeah.

1:22Roman Chernin:Hi. Thank you for having me here. I think it was quite accurate. I have one small correction, actually, that we never were delisted. Oh, okay. our yeah our trades on nasdaq were uh frozen for two and a half years a little bit more and then we just like overnight continue to trade with a new business new name like uh some of the team inherited yes but yeah i think it's quite unique situation on the public market i mean it's so interesting i mean i mean yeah let's get into all of that but first is for a touch on like you've just announced some big news can you touch on what you've announced who it's for and why it's important sure yeah so what we launched today is uh the product called nebula stocking factory which is essentially the next layer of the um cloud uh full stack cloud we built and it's built to serve the inference needs of every vertical AI company and cloud-native companies that apply AI and in the future enterprises, classical enterprises.

2:31Roman Chernin:And we provide the full experience to use open-weight models in an optimized way, fine-tune them, and actually build a flywheel of improvement AI applications and AI products on scale. I think it's the very natural step for us. We're always told that how we see the company is not just the large infrastructure company, not just scaling horizontally and building more and more gigawatts and deploying larger and larger clusters, but actually building the full stack offering for everyone who built on AI and we started to do it with our multi-tenancy cloud, which I believe is one of the best products now out there.

3:25Roman Chernin:And now the next layer is actually for the next wave of customers that we see who, as I said, vertical AI companies and cloud natives and enterprises, who actually start from a part of the previous kind of most of demand. And they not start from building foundational technologies, like foundational models, but they come actually to build the use case for the customers. And they need kind of new capabilities that we provide. Yeah, absolutely. And people, you know, you've already got some early adopters kind of using the token factory, right? Can you touch a bit more on who those people are? What are they using it for?

4:08Yeah.

4:09Roman Chernin:So just a few examples. For example, we have a video gen company, Hickswild, who are building B2C product in video generation. And what we help them with is to build a very, very efficient, scalable infrastructure to serve inference and meet their quite aggressive needs in unit economics. so you can think about it like you cannot build it in a very naive mode because when you come to B2C market the unit economics of acquisition becomes super important on scale and you need to do all the optimizations that you can to not over provision resources and serve scale up only infrastructure that you really need and you also want to do all the tweaks like I don't know, tiering the customers and serve more premium customers with the better SLAs and the tail customers with some other relaxed SLA, but for the cheaper price and so on and so forth.

5:21Roman Chernin:So we help them to build this infrastructure that actually enables them to do the business they do. Another example is one of the e-commerce companies in the process, which is like large international holding. And this is a different type of the customer coming from this pre-AI era, like cloud native, digital company, but pre-AI era. We helped them with a few use cases. One interesting use case, they moved all their user notification processes to AI. And they had the task to reduce the cost comparing to the closed source models that they started with dramatically and we could reduce the cost 26 times for them, which sounded like quite insane.

6:12Roman Chernin:But this is doable with the right combination of the smaller models and optimization that we did for them. Another example is like large e-commerce company that that needs to process a lot of data in, we call it batch inference. So again, you are not optimizing for the latency, but you optimize for the throughput and the price. And here we leverage the fact that we are full stack. And we actually can use the fact that we control the underlying infrastructure and we can schedule, we can orchestrate their workloads in a spare capacity that we have from other customers and make sure that we actually provide them the cheapest price possible.

7:04Roman Chernin:Because for us, it's like, you know, everything in our business, everything better than idle capacity. So if we can utilize it with this kind of high throughput, high latency type of workloads, it's good for us. And for them, it's also super, super win because they reduce the price and have very flexible throughput. So this is just an example of the type of the customers, but we see everything from conversational AI, like all the types of the chat and conversational applications. We see this kind of data processing workloads. We see e-commerces that do recommendations and search. and surge so like variety of cases yeah and 26 times cheaper is a phenomenal result do you see you know is that like a best case scenario do you see that going even further as time goes on yeah it's interesting because actually i think the market let's say the non-professionals on the market underestimate how good you can optimize for the specific use case.

8:22Roman Chernin:So we see a lot of leaderboards that show what model and what provider performs better. And it's important to show up there. But in the real use cases that we see, when the specific customer comes, they always come with their specific SLAs that they need because everyone has their own limitations. And what is important is that everyone has their own stream of queries, type of customers and type of data they deal with. And then you have a variety of optimization techniques that you can apply for the specific workload. And actually, the platform we build is the platform that simplifies for them to find the ideal setup that actually serves their specific workload on their scale, on their limitations in the best efficient way.

9:11Roman Chernin:to meet the economics, to have the flexibility to scale with the customers. And actually, which we didn't mention, but the important thing is to control their destiny and control the data to make sure that their main asset, which is in our AI era, is data, is handled right and applied to improve their use case in the best way. yeah got it super interesting and i guess like so much of this has been built around enabling product product inference using open source models are you seeing an increase in demand for open source models and you think that's a trend that's going to continue yeah absolutely i think that we are in a very interesting time when And I like to tell it that the AI demand has few waves.

10:11Roman Chernin:The first wave of demand came from foundational model builders, not necessarily the largest ones, but we have the largest one, the super labs, and we have dozens or hundreds of specialized model builders. And they were the beginners of the market. Then the second wave is actually these vertical AI companies. And there was the word, not very pleasant, GPT wrappers at the beginning of this AI race. But now we see that those GPT wrappers are building like DecaCorns and the fastest growing in the history, new type of SaaS company. Like you can think about Cursor, Lovable, Cresta, you name it. So they start from the user.

11:05Roman Chernin:And they start from actually using the most powerful models to unlock the user scenarios to show the real value. And it really works because the models are very powerful. But then when they scale, they actually start seeing the limitations of sitting only on top of the frontier models. The limitations that we discussed, actually. The price, the unit economics, the technological mode, the data they need to apply safely, like sovereignty requirements, a lot of things. And this is the moment when they start shifting to open-weight models. and they need infrastructure to serve it and they need the tools to like actually lower the barrier for them to achieve the same results because when they come to google open ai they actually not only getting the best models but they also got a lot of like tooling around that and i think what we do is actually helping them to achieve the same results or better results with the ecosystem of on a certain model.

12:15Roman Chernin:And we definitely believe that this will be the pattern when a lot of use case, a lot of companies, when they unlock the use case, they will either switch or combine different models. Because for some use cases, you want the most powerful reasoning model that thinks long time, kind of has a higher cost, but provides like the best result. But for some other cases, You can think about low latency scenarios or low cost scenarios. You may be satisfied with another model. And you can take it from open source. You can fine tune it and control. So it's not necessarily the full shift. It's more like building more comprehensive system in a less naive way to serve the real SLAs.

13:12Roman Chernin:and that's that that we see it as a trend we see a lot of demand coming and we see that open source the ecosystem of open source models are continue to be quite developed we see nvidia pushes a lot on having enough of open source choices in the world we see even the big labs are open sourcing some of their models so it feels like a good trend yeah absolutely definitely definitely i feel that more people are getting used to relying on those open source models like more everyday builders are finding ways to access those open source models and build on top of them and not just default to the big frontier labs um i want to talk also about the year that you've had i mean i think you've raised billions the share price is up several hundred percent um you know you signed i think a deal with microsoft that's just under 20 billion dollars this has been a phenomenal year for for the team like what's it been like working at the company during the course of 2025 uh it was a lot of fun a lot of non-sleeping nights a lot of efforts you can imagine no it's definitely one of the maybe the fastest industry we ever saw in the in the world in the history of the economics everything is moving so fast the cycles are so fast like we just we just you know it's crazy we just three years from GPT-3 was released this aha moment happened and now we have like maybe the most impactful area in the world economics like all this news about GDP is driven by infrastructure deployments and so on so like everything is moving so fast and we are obviously super happy to be part of this race and And we see our role again as enabler.

15:07Roman Chernin:And we call us tech-to-tech companies. We're really excited to work with all the great builders, foundational model builders, frontier labs, smaller labs, and obviously this new wave of vertical AI companies that are doing such an amazing stuff that changes the businesses and changes the human's experience every day. Yeah, it's just, you know, it's just very, very fast run. And we see a lot of opportunities to create the value by our motto is this vertical integration when you build all the layers from physical infrastructure through the robust cloud platform and then to value-added services to help people build.

16:06Roman Chernin:And, yeah, that's a lot of fun and a lot of excitement from people use what you do and grow with you. It's great. Yeah, no, and, you know, things are moving, like you say, so quickly. Things are changing so quickly. The whole industry is growing at an unprecedented rate. Given the speed at which things are changing and growing, are you able to plan effectively for 2026? I mean, the scale and the demand. No, we obviously plan a lot. Obviously, our plans will be revised and year is the century in this market. But I think we have a very clear understanding like the North Stars goals that we're going to.

16:59Roman Chernin:And I think that in a simple word, it's two dimensions of growth. Horizontical, like we just need to build more infrastructure. We have a very aggressive roadmap for new facilities for the next year. And the following years, we invest a lot in A, physical infrastructure, and B, make sure that the platform is scalable and we can serve larger and larger customers. And then the second dimension is actually this building the new layers of the product offering and unlocking the new possibilities for our customers through software. And this is a token factory. And we obviously have more plans for the next year.

17:45Roman Chernin:What new capabilities to the platform to bring. Yeah, no, it's super interesting. I also wanted to touch on the, you listed on the NASDAQ, right? Like being listed gives you access to kind of raise money in a different way to private companies, to fund that expansion and to grow the infrastructure. But obviously you had such an interesting journey where your shares weren't being traded and then they were being traded again. You know, doing my research, I was like, what must that day have been like? You know, what was that day like where you switched it back on, where your shares were being traded again?

18:19Did you have any idea about, you know?

18:22Roman Chernin:No, it was absolutely crazy. And it's really happened like almost overnight. So we just got a call from Nasdaq. Okay, you can trade again. And we're like, okay, but nobody knows us. Like no coverage. Like we have some legacy investors that invested back in Yandex and they just stuck in the stock and they had no idea what to do. So it was just like full uncertainty. I think what helped us a lot is we've got some strategic investors that actually believed in us. And we just after the going public again or trades again, we had this pipe led by Axel Ventures and NVIDIA that actually set the trust stamp on us.

19:15Roman Chernin:And I think then in the next like six, nine months, it was like every month, like somebody new like learned about us and we worked hard to get this awareness, both in the product and in the market. And then Microsoft deal was obviously like the important milestone because finally everybody read about us in the news, even the people who didn't know. like i had a funny stories like my whatever from people from my past 20 years ago they just messaged me like oh i didn't know you do it so uh it was quite a five minutes of the of the honor and um yeah i think now we finally in the right in the right place from awareness perspective and Nebius is in the kind of our industry of AI clouds, AI infrastructure is a well-known name.

20:18Roman Chernin:What we need to do, the next step is actually position us as much modern infrastructure. We know, like we built a lot internally, but I would say market yet don't know how much we can bring more than just GPU hours. And this is the next kind of milestone that we are targeting. Got it. Okay. So greater awareness of the other side of your business. Also, you know, when you're going through that whole transition from kind of Yandex to Nebius and, you know, shares being traded to shares not being traded, did you ever consider actually maybe we should just entirely delist and become a private company again?

20:58Or were you always thinking actually being public and having that option is going to be a huge advantage to us when we get things up and running again.

21:05Roman Chernin:Yeah, it's a great question. Actually, I don't think we had a choice because we had a responsibility to our shareholders and we definitely had to proceed this way. I think if you would have a choice, we would maybe prefer to be private for some period of time. but in reality that was a big luck for us to be on the public market that early kind of like depending relatively to our level of much development for the company because you you're absolutely right this is the market of combination of tech and capital and being kind of in the public market opens a lot of opportunities to finance our growth and i think this is quite a quite unique thing about nebios that a we started with some capital that we've got from like this painful but and then successful kind of corporate restructuring after the yandex but then being so early relative to company journey to open markets we got access to to the capital to finance growth and it let us do things even not standard for this ai cloud new ai cloud or world when we can actually deploy capacity in advance finance it from our balance sheet and do this more like harper scalar game on smaller scale obviously but not just deploying clusters for large customers yeah it's such an interesting business that you're building and it's it's going so well you know you i think you're still one of the most unknown tech companies of your size in europe and you know i think it's always i think you want the phrase that you gave is like tech to tech it's such an interesting one because you you are powering so much of the the growth and the tech that we're seeing across europe whether it's from you know the smaller to the enterprise ai companies um but look roman we're out of time but you know i feel like we could chat i could chat ages about this um but thank you so much for joining me and i'm we should we should do it again we have so many topics to cover including european tech for sure and i actually had a bunch of questions on that and your views on uh european tech stack versus uh us and infrastructure layer versus application layer um so look if you if you've got time it's definitely scheduled a longer a longer form interview at some point i'd love that but yeah thank you so much thank you for having me my pleasure that's great to talk thank you