Nebius AI Infrastructure: From H100s to Blackwell | The AI Rabbit Hole 2025 | DVC

Well, thank you everybody and it's our second year sponsoring this conference. So, we're super excited to be back here and we'll get started about 20 minutes of content for you guys. Get going. So, the hopeful takeaways from this presentation today are first that we built something pretty insane in the last 18 months. We've been a private company that went public. We have tens of thousands of GPUs globally today in our data centers. we're working with you know angel to seed to series A BC even in enterprises today scaling their foundational models to AI agents that are out there and so if you are sitting in the audience today whether you are a startup or someone who's got an idea we have an official NVIDIA inception program for startups so we have credit available so as as mentioned if you do need GPU credit we have an application process there and we we can service you there if you have a current contract obviously with another GPU provider we can figure out how to add value to your operations today.

and then so what we've you know where we come from is a pretty wild story too and I'll get into that as well. not many companies have kind of had our launch in the last 18 months which is pretty wild. And then as I mentioned I didn't know there wouldn't be Q&A here today but if you want to ask some questions right afterwards you'll find our booth in the main hall here. So if you have questions about our data centers, GPUs or any model questions you might have, please reach out to our team outside in the hall. So first obviously AI is nuts, especially within this region of the world. It's what everyone's talking about. you know, you could just type in AI funding into Google and you'll see how many of these companies are getting funded. Seems like there's some every few minutes at the moment.

And so I think what we'll we'll do is kind of explain how we fit into that picture and how we can add value to what you're doing. so first we came from Yandex and if anyone knows what that is we were the Google of Russia. So think about Google the same business as search Uber maps all that. and it was a pretty thriving you know $30 billion market cap business. And then all of a sudden a war started and we a group of our employees about 2,000 very brave souls decided that that was not something that they were going to support. Our founders didn't want to support that and we started the process to kind of move that business away and there was a settlement and that's how Nebius was born. But it's it tells a lot about the people who are at our company who in some cases haven't seen their families in over a year and a half and decided to take this great journey to have a world of innovation a world of freedom that they could have.

And so it's a it's a real story and it's very unique that's out there u today. And then that brings us into obviously who is Nebus today. So we are a global GPU infrastructure provider with tens of thousands and growing of GPUs that are out there. we are publicly traded as I mentioned. I won't give away any secret information today that won't be coming out in a few months. but Nvidia is a lead investor in Nebus. So that's you know there's about a hundred NCPs today which are cloud hosting providers for GPUs. And to be one of the ones that Nvidia invest in is a is a really cool hat tip to the industry. And but that's you know where we're going in the future. We're going to have Blackwells and other iterations. We're a full stack platform. So we have all the software. We're not bare metal.

So if you're used to that AWS like Azure GCP like virtualized platform, that's Nebus but with only NVIDIA GPUs and we'll get it started. and that has mentioned some of the hottest foundational models to enterprises were the kind of cloud they're plumbing so to speak the infrastructure which we think is sexy behind the foundational models that you guys are using today on your phones or at home on your laptops to get it started you know why did we create this company you know before the war happened in Russia Yandex was actually the third highest GPU using company for Nvidia of GPUs and so we had really base knowledge of how to at scale be able to integrate software hardware into our our platform. And so and then the data centers and knowing how to build and operate these things today. And we'll get into our data center map in a minute, but bringing all these things together is how we really birthed Nebius over the last really 16 months.

And so again, what we put together is you know, not normal within the space. A lot of times for example someone will get you know 500 nodes of of GPUs we're launching clusters that are in the thousands and so that is the expertise and the drive that we're we're we're scaling at and operating at our GPU fleet today is about 10,000 H100s that are in production. there are still people out there that want A100s I know but you know this is pretty much the lineup that we have today. So, if you want if you don't believe that there's that that much memory difference between an H100 and an H200, we're we're we're there for you on the H100s. Our H200 clusters arrived this year in our both Finland, Paris, Iceland, and this week, Kansas, USA. So, it was excited for us to actually launch in the US.

We're not just a European GPU provider, but we actually have GPUs now in Kansas, which is pretty amazing. And then the Blackwells will arrive this summer. That'll be the B200 and then later in the fall the GB200 will be in production at our new New Jersey data center. But this is our line card of GPUs that we have both for startups and enterprises down to anyone who's looking to scale an AI idea that they have. This is a map a little bit of kind of our data centers around the globe today. As I mentioned, we started in Finland. This was a Yandex data center that's one of the most sustainable 16th largest supercomputer in the world that is in Finland. all the towns that surround that data center actually get their power from the data center because it's so efficient and we'll talk about that in a few slides.

we opened up our Amsterdam HQ. Where did all the our our amazing colleagues go when they they left? It was a pretty wild story, but they ended up in Amsterdam. And Amsterdam was pretty smart. there was over 500 software developers that came out of Yandex and they're like we want them and that's where our headquarters is today in Amsterdam and then we opened up Paris Iceland and then the two US data centers the Kansas is operational this week and then New Jersey will be again for the GB200 that will be there and so we're going to start to obviously look in the western US eventually to get there but today if you're training a model you don't really need to be within a geographic footprint inference is very different obviously you want to monetize your AI idea, you have to get it closer to the user.

And so that's where the reasons why we've opened up these data centers and the places that we have we have plans for Asia down the line and other things. But this is our our core map. And as I mentioned earlier, between the software developers, the data center expertise, it really pulled all this together this company within months. And so in fact we had to rewrite our entire platform to get reertified to be a publicly traded company again. So when you go through the process with NASDAQ it could be like a a black light on your soul a little bit but we were able to confirm that very clean operations everything is sound and rewrite our entire platform for clients to be able to use. So today for anyone in the audience today if you want some promo code we can give that to you. So if you want to spin up a an H100 VM, please reach out to us and we'd love to be able to make that happen for you.

and again, we are of course publicly traded and an Nvidia reference part partner which matters within this space. And these are just a few of the clients that we have today. We've we have Nvidia themselves which is a client today that we don't publicly usually talk about or maybe I just did but it's exciting again to have them on the platform helping us be better every single day and then we have all the way down to Stanford lab graduates that are out there wanting to train their idea that they have and we love to be the innovation and the brains and the and the compute behind those ideas that are there. and so whether you're, again, it doesn't matter how small of a startup. One of our clients that signed this year are actually two existing Stamford University students. So like they're like still there in classes, yet they raised money and the GPUs that that funds their business is a company called TILD AI.

And so it's a it's a pretty cool to see the wide variety. We have the startup program that I mentioned, the university program we have if you need to get credits. So please reach out. Go to nebius.com. We have forums out there if that is of interest to you guys. So, first it it's worth noting in the space buyers of GPUs are pretty savvy, right? There's like used to be two asurances in life death and taxes. And the third is that everyone's out there shopping price for GPUs and I don't blame them for doing that. It is incredibly expensive. they say for a AI startup the GPUs are either one or payroll is two in terms of expenses for your business that you have out there so they're going to negotiate on price and so in this industry that we're talking about which is really only three years old I often will interview people and they be like I have 10 years of AI experience and I'm like well I don't think that was a thing 10 years ago but so what we're trying to do is move on from just the pricing capacity and when you're selling the same thing, the same bare metal GPU server, it becomes a race to the bottom.

It's a commodity at that point unless you have managed services and software above the GPUs, which is what we're most proud of within our business. and as I mentioned, all the pricing is on our websites. we have on demand, so if you're somebody who says, I don't want to put down a down payment at the moment for GPUs, we have the ability for you to start out with no commitment with us. And so you can do true hyperscaler alike on demand and we can raise your quota that you as you need it there. And obviously that's going to be a little bit more expensive per hour versus a reserve model where you're committing to a certain amount of time that's out there. And so reach out to the team you'll see in the hallway today and we will hook you guys up with either promo credit or an ability to do a PC on the platform.

Moving into platform flexibility which I'll get into right now. It used to be in this world where if you were only selling bare metal, somebody would hand you your bare metal through SSH keys and IP addresses and you would have to go out and buy all the different software solutions that are out there. And so what we did is we have put that on top of our GPUs. The same solutions you would have to go out and buy like five or six vendors and they're all software as a service. You now have that included on top of Nvidia GPUs. Oh, and here's the kicker. We're not charging for it. So that's all free. There's probably $4 an hour worth of software or more that is included on top of what we're about to show you. and the most important thing is we build it. We own it.

We support it. The only thing that is Nvidia reference architecture is the actual motherboard of GPUs that that belong in the server. The A cards we build everything else around it. And that makes us more efficient. We're not buying from the likes of the large amazing also hardware providers like Dell, HP, E and those guys. We do it ourselves. We're our own VAR in a way. And that means that we can have more margin. We can pass that back to clients through pricing breaks. And so that's how we we go to model. And as I mentioned, this is just shots of our data centers that are out there. It used to be that I remember the two Neo clouds that I worked at before. We would wait four or five months to be able to get 256 GPUs 32 servers to to come to our data center.

And now we can roll out thousands within months. And so it talks about the scale and about how efficient we are at deploying GPUs. And this is our platform real quick which I'll get into. Obviously there's the actual Infiniban network the speed the GPUs at the bottom and then everything else is products that we offer today storage. for with the foundational model you cannot have a conversation about training a model without having storage conversation. So that is included as well as a price that we have our own storage. And then we have industryleading partnerships with the likes of vast wet DDN and other other providers that are out there. We have Kubernetes. We have a slurm operator that goes on top of our our GPUs. So if you're used to having these tools that you have to go out and buy in a bare metal environment, let's talk because we could probably reduce your your costs that you have at the moment and operations that are there.

And as I mentioned, you can go out there and promo this at our website. you can create a virtual machine quickly add a region to it, a VPC, network storage, what have you, and then again just operate within that environment. this really isn't possible in the bare metal world. When when you buy GPU bare metal, they kind of hand you that bare metal and then you have to operate at that point with a with about seven or eight nine clicks of a mouse, you can be able to be training something or doing inference or fine-tuning or whatever your use case is. at that point and as I mentioned you can actually dive into the actual you know a lot of times what happens is these supercomputers run really hot and they break all the time. Our ability to have a virtualized fault tolerance platform allows us to be able to replace that card very quickly that node that has busted and get you back to monetizing and training your idea or thing.

And as I mentioned earlier, this gives you an idea of kind of the a little bit of the software that we have here within the industry right now in terms of schedulers. People are primarily using Kubernetes in terms of containerizing their workloads and calling them up or they're using slurm and the bare metal environment. We combined those two things together into our own software product which is free by the way out of out of the platform and it it just literally every single thing that if you ask anybody who runs a GPU network today these are the very very difficult things about running it and orchestrating it and making that happen and that's what we did with one of our products which you'll be able to use that's out there as efficiencies I mentioned earlier this is the Finland data center that's out there this is what came over from Yandex and it is one of the most efficient we're building a twin version of this data center across the street and it literally heats 25 homes, 2500 homes all around the local area there and there aren't many data centers that can do that especially within we're having kind of a a data center issue within the US here of space and power and what we have here at the moment is pretty special and some model we'll be using at our other data centers and as I mentioned earlier the software we have an AI inference studio product if you're looking to get out to the models like deepseek lama it could be mistrol out there we have connections to go from so once you've trained your idea and and if you need model connections that software is already built into our platform

as well you don't have to have two separate platforms the models are on top of our managed services and GPUs that are available for you to be able to use we go and so the three things I I would just leave with you guys again obviously we have a ton of GPUs that's that's already been said. We have a full stack platform that's out there that's waiting to help you train your idea. that's there. And if you have any questions you wanted to ask during the session that because we didn't have a Q&A session, come outside to the the the the booth that's out there in the hall, ask your questions. I'm Shanebius.com. It's like the easiest email address that I just gave out ever at a show. And reach out and talk to us. and we're we'll talk anything AI, training your idea, startup credits, and we're here to help you, the community, be successful in the space.

Cool. All right. Thank you guys. Thank you. And I finished early, so that that's the way to Thank you, man. So much.