Nebius posts 841% year-over-year AI cloud revenue growth, acquires two inference optimization teams to accelerate GPU efficiency

Speaker 2: And Nebius is only up 15% today, so we'll ask him

Speaker 1: The AI Cloud division posted 841 year over year revenue growth. Earnings per share beat expectations significantly. Shares are up, as you mentioned. The business has been on an absolute tear. And we're excited to be joined by Roman to break it down for us. How are you doing?

Speaker 2: I'm great.

Speaker 4: Thank you for inviting. I must say that I think three meetings ago

Speaker 1: Yes.

Speaker 4: Someone said that I need to take a nap. So don't be surprised. I would just fall

Speaker 7: Okay.

Speaker 4: Fall apart here.

Speaker 1: I can imagine you're working nonstop right now. How long have you been working on this? I feel like Nebius is in the news every day, but can you take us back a little bit on your journey with this company?

Speaker 4: Yeah. So I I I'm with Nebius from the very beginning. Yeah. And Nebius officially, in the current shape of the company, exists from summer twenty twenty four, so less than two years. Yeah. But we as a team have experience longer than that. Yeah. You you maybe know that core team came out from Yandex, which was a large Russian Internet company. Yeah. And, yeah, we started with quite a unique mix of talent that helped us to start, like, really fast.

Speaker 8: Yeah.

Speaker 1: So, I mean, people know Nebius Neo Cloud, but that can mean a lot of things. How are you describing the shape of the business? How are you describing the

Speaker 2: before we get into that, what's the other what's the other spin out that's also on a tear? From Yandex? Yandex.

Speaker 4: Yeah. You maybe know Clickhouse.

Speaker 1: Clickhouse. That's right. Yeah. The Yandex.

Speaker 2: The the talent talent at Yandex seems to have been just, like, genuinely insane. So Yeah. Yeah.

Speaker 4: You you you can find a lot of other startups that or companies that created with Xiamlex people. Yeah. But Yeah. Uh-huh. Not really. About how we describe our business, I don't really like the word new cloud because, I think it's too broad definition of too many companies that do different things. Like, someone is doing just really data center business. Someone is doing data center plus renting out hardware. Someone don't own hardware and do kind of marketplace play. We like to say about us that we built vertically integrated, full stack AI specialized cloud AI specialized cloud. So that's that's what we think we do. And, I think that the difference of, like, neo cloud or bare metal compute and cloud is quite significant because it's completely different developer experience. Right? So, if you need to come and rent a large cluster and wait, like, six months it to be delivered and then sign it for three years and run it, Probably it's not really real cloud experience. Right? In the cloud, you expect more flexibility, faster time to value, more tools that you can use as a developer, more layers of the value. So yeah.

Speaker 1: How do you think about lagging edge model inference, like, load over time? Because a new model comes out. There's an incredible amount of hype. Everyone's looking at benchmarks and testing it. And then at a certain point, open source catches up, different models sort of commoditized. But if you're an enterprise and you found that GPT-four is great for scanning your receipts or something, you might leave that capability, that functionality running forever. And you might never need to deploy a smarter model against that problem. And I'm wondering as we see skyrocketing frontier revenues, lots of debate over token maxing versus incredible results on the profitability side. How are you trying to understand what's happening with those workloads that have been sort of baked into the economy already? Like the AI diffusion story is is done because that capability has has been adopted fully.

Speaker 4: Yeah. So I think there are as we see a few brands, let's say so first, as you said, open source models are catching up Mhmm. Different tier models quite quite fast, probably.

Speaker 7: Yeah.

Speaker 4: I mean, I think that, like, there is three, six months gap Yeah. Like, that we see for the most of the tasks. And we see that a lot of people, a lot of customers, when they come to the scale, they so the journey is you start with the frontier model. You you want to have the best capabilities to unlock the use case, see to that things working to start growing. But then when you grow, you may meet the limitations. And first of all, in economics because at scale, you the the economics, the economics matter. And then if you have other capabilities from open source models that can deliver in a particular use case that you already understand, like, same or close to same performance and quality for times lower cost, you need you you're actually switching. And it's not easy to switch because you need to do those models. Nobody uses, obviously, vanilla models. You need to RL, like, you need to figure out your data and so on and so forth. But, again, if you scaled, you you you have a lot of value there. And then what you need from infrastructure where where we believe that, like, our value is created is to help you with that, to lower the barrier to a, tune those models and then b, run those models reliable and cost efficient. And, that's, like, the layer, the next layers of the offering that we are building.

Speaker 1: As you look out over the next couple years or even a decade into the future, which sounds like an infinite amount of time, but but on on on the near and midterm, how are you thinking about semiconductor bottlenecks versus energy bottlenecks? So you've been going back and forth on this. Both are in short supply. What's keeping you up at night?

Speaker 4: I'm a lucky person. I'm more on the product and demand side, and it's a lot of fun. So I I really I Yeah.

Speaker 2: Being the being the demand guy during the explosion of demand has gotta be great. It's like The best work. Yeah.

Speaker 4: The best work in the world. Yeah. So people hug you and

Speaker 7: but but

Speaker 4: no. I I think that now, obviously, the physical kind of the physical infrastructure, the the even not the energy itself, but, like, getting from greenfield to having data center that works and full of GPUs is actually the big, the big the big challenge. And I I like, the the bottlenecks are, like, along the all the supply chain, starting from whatever connecting to the grid or bringing local generation and all the way to, like, even human power and all the all the complexity. So I really kind of take my head off from, like, in front of the people who who who are bringing those, like, gigawatts capacity online, and that's challenging work.

Speaker 1: Can you So

Speaker 2: how do you there there's such overwhelming overwhelming demand and just energy to bring compute online. How do you measure yourself and the team's performance? How do you know whether or not you're doing a truly great job? Right? Because people are so desperate that that Yeah. They'll they'll work with like a third tier supplier just because it's their only option. You guys have shown, you know, tremendous capability. And so but but but then again, it's like how do you know if you're actually doing an exceptional job?

Speaker 4: It's a good question. I I think the race is so that you never know. Are you good enough or not? So you look like, oh, you're growing, like, whatever. You're growing, like, times a year or so. Is it good enough? Like, in any normal business, yes. Probably, it's excellent growth. And you meet your customer, and they say, oh, you know what? I grew seven times in the last four months. And you're like, okay. Probably

Speaker 2: probably we need to move harder. Yeah. Yeah. I and

Speaker 4: and but you you there is another component that I I think is people speak less about, which is how you finance all that. So it you ask about, like, the bottlenecks in the chips. You ask about, like like, the bottleneck to bring physical infrastructure online. But then there is another part. You need to finance all that. So I think that now if you would have unlimited capital, we would be able to on technical execution, on operational execution, we could even move faster. So there are, like, there are so many components in this business of AI infrastructure, that you need to do efficient, like, and fund like, at like, finance enough and finance smart to not eat all your margins. It's it's a part of the, of the science here.

Speaker 2: Mhmm. What is a job at Nebius that is, sort of the unsung hero of the business? Besides besides you. But we're singing for you

Speaker 8: No.

Speaker 1: He gets sung for all the customers are singing. They're hugging you, you

Speaker 3: Yeah. Yeah.

Speaker 2: Like But but I but is it like somebody that's like perfect. Yeah. Boots boots on the ground.

Speaker 4: I I yeah. No. I I think that our business is very much like like, as any cloud business, to be honest, it's post sales. It's post sales business. So, like, when you sell, you give a promise. People hug you that you allocated capacity to them and, like, they have to start working with you and so on. But then how do make sure that the customers are happy and things are working? And, again, everything is moving so fast. We we've got new chips every three months. We got new physical data centers every month. We got new customers that of the scale that we never saw before every month, and the workloads are changing so fast. And I think that it's not like one role, I think, that underappreciated. But, in general, delivery in this and I started with that. It's execution business. It's it's a it's it's not not that it's not it's it's it's it's it's less magic. It's a lot of a lot of boring execution in each layer starting from finance,

Speaker 3: like

Speaker 4: physical infrastructure, software, customer kind of facing engineering support, like, every piece. And then only magically, everything comes together from supply chain to customer facing, person that, make sure they don't waste their money with us, then that that then things start working. So, yeah, I think delivery is the key. Delivery chain is the is the key in this business.

Speaker 2: Post sales.

Speaker 1: Last question. Can you talk about some of the recent acquisitions and talent moves at Nebius? I'd love to know the philosophy and and strategy there. Yeah.

Speaker 4: This the philosophy is very simple. We need to build so many things that we need to move, and we need to move so fast that we're always looking for people who can accelerate us. And it's it should be exceptional talent, and it should and or it should be something that had a great adoption. Our two recent acquisitions that we announced just like with the two two weeks post, Two teams that so two teams that work on inference optimization. So you can think Mhmm. Again, our part of big part of our business is how efficiently can your GPUs into tokens and, like, a value for for the customers. Yeah. And these are two teams. One of them, Aigen AI, another Clarify One is very much focused on model optimization, like the the engine of inference, how how you run specific model and all the techniques around spec decoding, quantization, and so. And another is more like system optimization, all the routing, k b caching, orchestration across the big cluster of compute and so on. And we had a very strong internal team also working on inference, but we felt that in we need to move fast. Like, we need to move faster. We need to bring more capabilities there because the market is so fast that every every three months, you can lose the the pay, like, so so so big pie so big piece of the pie. Yeah. And it's gonna show up

Speaker 1: in your in your margins, in your earnings, obviously, which are important now. If you if you don't if you unoptimized chips, you're not gonna get the yields that you want. The the results and the tokenomics are gonna flip.

Speaker 4: Yeah. That's true. And and also what's important that, it unlocks the new types of the customers, the new types of the workloads that we can serve. It's like not just against our raw compute, but serve all these fast growing, vertical AI companies and enterprise adoption.

Speaker 1: Yeah. That makes a ton of sense. Well, thank you so much for taking the time to join us on a busy day.

Speaker 2: Yeah. Are you gonna get some sleep or are you, you'll be asleep, in a few years?

Speaker 4: I will take enough now.

Speaker 2: Fantastic. Just a nap. It's I know you're you're you're across the pond, right? So you should be just going to sleep for eight hours but he said it's

Speaker 1: A nap. I love it. Well, for coming on the show, Roman. Talk to you Yeah. Have a good one. Yeah. Goodbye. The last post I want to talk about today is about video games of course. There's a individual who beat. Is this a Dark Souls boss or but they built a controller that uses their full body. It's not a VR game. It's not a VR simulation. But if you want to attack, you strike the dummy with a physical stick and that presses the button in the game to trigger your character to attack in the game. And then it appears that he's playing on a green screen or something because you can see the game actually playing behind him. And he also is wearing glasses of some sort. You don't like this. What's not to like? You need to get a workout while you're gaming. The future of gaming

Speaker 2: I just don't like

Speaker 1: very well might very well produce some of the fittest people. Yeah.

Speaker 2: I don't like the aesthetics of of of this.

Speaker 1: You don't? No. You think you should you're more in the But

Speaker 2: I get I shouldn't knock it till we try it.

Speaker 1: You're more you're more a fan

Speaker 2: of seen it.

Speaker 1: Let's get You're more fan of of unironic LARPing. Like live action role playing where you get the like the SD kid concert. You're a fan of that. Put on the actual armor.

Speaker 2: Yeah.

Speaker 1: Do the real battle.

Speaker 2: Yes.

Speaker 1: Don't play the video game version of that. That's a poor summer locker room.

Speaker 2: Yeah. Hit the park. Yeah. Hit the park.

Speaker 1: Or find a multiplayer version of that. Just get another human to stand next to you and you can learn defense.

Speaker 2: Perhaps. Before we go

Speaker 1: What?

Speaker 2: There's some there's more of a scuffle on the timeline Meet the Meet GC campaign.

Speaker 1: Okay.

Speaker 2: What's I realized why Mark is taking shots at it and it's because I think they use an actor to try to look like him. Let's watch it again Oh, with that

Speaker 1: really interesting. I did notice that the actor that plays the general catalyst character is like over the top handsome in the sense that like the original Apple Mac and PC ad Justin Long, he's sort of like an everyman and and the the actor who's playing the general catalyst guy looks like a male model to me. I don't know. You can be the judge of it, but

Speaker 2: Yeah. All all I'm all I'm saying is I think I think we got a Drake Kendrick situation Okay. On our hands. I think the new media team at Andreessen needs to respond.

Speaker 1: I think

Speaker 2: They need to try to dunk on

Speaker 1: General catalyst? GC. Okay.

Speaker 2: I wanna see a war. Yeah. I wanna see a it should be a bloodbath.

Speaker 1: Anyways Demanding acrimony in the tech industry. Sowing descent.

Speaker 2: Yeah. No. No. They they need to get

Speaker 1: peace? What about peace? Need to

Speaker 2: get revenge. Mark says this is a clear inversion, clever inversion of the original Mac versus PC commercials. This time it clearly intended to make the sponsor appear smarmy and judgmental. Fascinating. So

Speaker 1: You got rage baited. They rage baited you. You could have just muted this. You could have blocked them. You could have not amplified it. Now it's on everyone's timeline. Now people

Speaker 2: It's war now.

Speaker 1: Because of this

Speaker 2: We're gonna get a VC content wrapped

Speaker 3: up.

Speaker 1: No one would know what General Catalyst is until now. Now there's a fight and it's working. No. It is entertaining though.

Speaker 2: Very fun. But Duke

Speaker 9: it out.

Speaker 1: It's a good campaign.

Speaker 2: It's time.

Speaker 1: I I do I do Eric Torrenberg. Reggie James. The first oh yeah. That's a matchup there. It is just I I I mean, to to, you know, put aside the the war, it is just interesting seeing a produced thirty second, sixty second ad for a venture capital firm. It is sort of a first. I can't see

Speaker 2: a a non Vibrio Yeah.

Speaker 1: Video. The Vibrio meta, that was sort of played again.

Speaker 8: At this point. So

Speaker 1: we gotta we gotta go a different direction.

Speaker 10: And we're going to TV ads. We're going to television full thirty seconds, highly produced ads, running on the Super Bowl on Saturday and live. I don't know. When did NBC advertise?

Speaker 1: Bloomberg, CNBC?

Speaker 2: I don't know. We gotta get on with London. Yes. So we will see you tomorrow.

Speaker 1: Leave us five stars on Apple Podcast and Spotify. Sign up for the newsletter at tbpn.com, and we will see you tomorrow at 11:11AM Pacific sharp.

Speaker 2: We love you. Goodbye. Goodbye.