I'm super excited about this session. As you all know, the emerging cloud infrastructure players are in the spotlight these days. They're growing through some of the most amazing growth trajectories in all of tech. I'm sure many of you were, frankly, unaware of Nebius two years ago. Now they have a EUR 25 billion market cap. Those stories don't happen a ton. And so I've been looking forward to this one because it's, well, Neil and I have had a few occasions to chat. I haven't had a chance to meet Marc yet. So Neil, nice to see you. Marc, lovely to meet you.
Pleasure is mine, Carl. It's great to be here.
Thanks, Carl.
Just because this is a newish story for people, you don't need to, Marc, go into the origin story. By the way, for those of you that want to go deep on Nebius, Marc just did a podcast, which I was just joking I still haven't had time to listen to. But it's a very good hour-plus pod running through the Nebius story. But I guess what I wanted to focus on just as part of the origin story is, obviously, the execs, a lot of them, engineers came out of Yandex in Russia. What was it about that experience that enabled you to scale Nebius? I'm sure the answer is just partly that you had a lot of amazing software engineers who knew how to scale a business.
100%. It's not the normal startup story, as you pointed out, Carl. The reality is that Yandex was already at scale and successful internet company that was operating some of the largest infrastructure in the market, actually the biggest Nvidia customer in Europe at the time. And when Arkady Volozh, our CEO, who was also the founder of Yandex, started Nebius, he brought with him 1,000 engineers, so software and hardware engineers that had been working together for, in some cases, more than a decade. So while the company is a two-year-old, or less, actually, company, the reality is it's a team with deep, longer-term experience and that they had built scale infrastructure of the type that we're building today in order to be able to service some of the most important workloads.
So we're starting with a bit of a head start compared to a lot of other companies that are in the category, given the resource depth that we have in the team and the expertise that they bring.
Marc, do you want to hit on one essential question I get, which is, what is the enduring strength of Nebius? Because there may be a misconception that some of the emerging cloud infrastructure vendors are scaling like they are because we're in a compute-constrained environment. But a lot of these investors are looking out three to five years to a point that will inevitably come when those compute constraints aren't as severe. And at that point in time, Nebius and your peers will have to sink or swim based on your core strengths. So what are those?
Like I mentioned, we have this wonderful, extremely experienced team of engineers that span hardware and software. These are individuals that have built scale infrastructure in the past. An important part of our offering is delivering the software that AI engineers are going to need. This is AI engineers of all types and sizes, or from all types and sizes of organizations. Being able to service a startup that's coming in self-service, that wants to put in a credit card and get access to a small cluster, be able to burst up to their needs, all the way up to a large enterprise that's looking for tens of thousands of GPUs to be able to service their requirements. That software gives them the capabilities, the tools, the workload management necessary for them to successfully build on our platform. Second, it's scale.
Scale is critically important in order to be able to deliver against the increasing requirements that are on the market. Today, we're one of very few gigawatt-scale organizations, and that's pretty significant when you look at bare metal. It's even more significant when you recognize that there are even fewer scale organizations that are delivering cloud infrastructure. We're not building a GPU as a service platform. We're building a hyperscaler, and that requires the combination of scale infrastructure and scale CapEx necessary to continue to fuel the expansion of that infrastructure, and then thirdly, we are highly focused on engineering at every layer of the infrastructure. We build from scratch. We can literally start with brown dirt, greenfield projects, and be able to put in place the data center, the power, custom cooling, custom racks. We spec out our own servers.
Our team is looking for every opportunity for cost optimization. We're looking for ways to continually improve our efficiency in the way that we're delivering and operating so that we can drive up our margin, and then lastly, and this is one that's a little bit hidden, and since I joined the company, I've been able to watch it in action. It's pretty extraordinary. We have a support organization that's made up of AI engineers, so the humans that are actually taking the inquiries from customers and dealing with the technical escalations, they are just like the humans on the other side. I hear this all the time from customers. They're very impressed with the skills and capabilities of the engineers that are responding, and they're able to get 24/7 coverage of their needs, so we're building not just a hardware layer that can be sold to customers, bare metal.
We're building the AI equivalent of a hyperscaler.
So let's dig in on that, and in particular, the full-stack vision. So if Nebius's dream is not just to be a great per-hour GPU rental infrastructure provider, what does that full-stack look like, Marc? How high up are you going to go? Is there ambitions to get like the big three hyperscalers are deep into the database layer? What's full-stack mean to you?
Today, what our stack looks like is the fabric that is stretched across all of the hardware that is providing the ability to manage workloads. Think of this as virtualization and managed Kubernetes. On top of that is an administrative layer, a control plane that gives the AI engineer the ability to manage their security, the operational characteristics, and ensure that they're getting the performance they're looking for. We're constantly adding new functionality to the platform. We're getting the guidance on that new functionality all the time from customers. The way that we're thinking about that new functionality is as one contiguous, vertically integrated platform. You can go from training to optimization to inferencing without having to do any extra steps. You basically have the tooling at your fingertips. Our vision is to continually specialize our platform.
Looking to the special needs that, as an example, our customers in healthcare life sciences have, which are different than, say, our customers in physical AI. They're delivering very different requirements to be able to service the end requirements they have. We're trying to put at the fingertips of the AI engineer the tooling necessary. In the fullness of time, our expectation is that we will build out an entire set of functionality that services all the data, development, security, and operational requirements that a customer has. Where we don't feel like it's a top priority need for us to own it, we will be partnering with ecosystem suppliers so that there's a readily available ecosystem of already integrated capabilities so that the AI engineer doesn't need to leave our platform and instead can actually meet their full needs.
So there is no ceiling per se, especially this early in the game. We're looking to all opportunities to ensure that we can service the full requirement the AI engineer has.
OK. So let's flip from the stack to what the customer base will look like three to five years out. I think everybody is well aware that for younger companies entering this business that's very capital-intensive, scale-based, you often have to start out with one large customer, scale up with them, and then use that as a beachhead to diversify. So we will get to that large customer in a quick second. But I want to start with what the five-year-out vision looks like, where you've got a number of customers loosely. Maybe you look at it a different way, Marc. But you've got the big model providers. You've got AI natives that are very compute-hungry. And then you've got traditional enterprises like UBS. When you think about what your plan is for five years out, what do you want that customer base to look like?
Where do you think the bulk of compute spending will come from?
Well, first off, our intention is to build a platform that can be consumed by all segments. We're going to pursue the four corners of opportunity in the market. It's very, very difficult to say with certainty that this movie is the same one we saw the last go-around with CPUs, although there's a lot to learn from that version of experience. And we're taking as much as we can from that experience to deliver on our market development. And if you look back at the experiences of AWS and GCP, early going was all startups. And the inevitable enterprise opportunity came much later. And in some respects, they're still making their way through the enterprise opportunity. And if you look at the addressable market, and if anybody has a strong, well-vetted, detailed, data-driven model, I'd love to talk to you because we're looking for that.
But right now, as we do the basic analysis, two-thirds of the opportunity is likely to be in the enterprise. And the challenge that we're facing is that it's still really early going. And I'm not surprised. I mean, UBS, as an example, is a conservative organization. And other companies like UBS, there's no upside for taking too much risk. But organizations like UBS are actually doing very specific things with AI, as an example, high-frequency trading or model-driven trading. And we're seeing those green shoots, those very early use cases in key vertical markets. And we're pursuing those with very focused attention to be able to get those footholds in organizations.
And then as they scale and we have the opportunity to pan out, my confidence it's likely that two-thirds of the opportunity, three to five years out, is going to be servicing scale enterprises that are looking to ensure their relevance into the future. And it's to be seen. But if that doesn't occur, we're servicing the rest of the market with the same infrastructure.
OK. Let's now talk about that one anchor customer, where obviously Microsoft signed an enormous deal with Nebius. I think with options, it's $19 billion over obviously multiple years. You're probably somewhat limited in what you can say. But can you say anything about duration? What exactly Microsoft is using Nebius for? Anything that might help the audience understand the nature of that relationship?
It's a five-year agreement. It's a very scalable, single-instance, massive cluster that they can be utilizing for any aspect of their business, and their demands and needs are extreme. If you think about it for a moment, they've got an established business around Copilot, which, by the way, is doing extremely well. We're a Copilot customer. We have it activated in our environment. People are utilizing it all the time to enhance their experience, and you can see some latency in its performance. They need to put more compute behind it because they've done a good job of selling Copilot to the market. They've also got products that they're building, models that they're creating. They've publicly said they want to own their own models, and they want to diversify their capabilities, so they're investing in model capabilities.
And then they have their Azure business, which competes with us in being able to deliver compute in order to service end customers. So we could be across all of their business. Our confidence is it's probably in the former categories mostly.
OK. Got it. I want to go into a little detail around Nebius's effort to scale up that infrastructure. But before I do, I want to hit one last thing more on the demand environment. When we listen to you, we listen to CoreWeave, we listen to Lambda, even when we listen to Microsoft, what they're all expressing is that this year it's hard to pinpoint it. But it felt like six or nine months ago, everybody just started talking about a surge in demand for compute. We saw a ton of huge deals, long duration, something inflected. And I don't pretend to have my finger on exactly what happened. So I'd like to ask you, what on earth happened, call it the spring, when it feels like everybody's tone just pivoted, Marc? Where did that come from?
I wish I had a crystal ball. I was not at Nebius nine months ago. I joined Nebius six months ago. Correlation, not causality, but what I can't tell you is what the experience is, and maybe this is just more of the same, but I do think it's important to share. We are definitely seeing an acceleration in demand, and this is not isolated. In terms of thinking about revenue and pipeline production, we saw our pipeline, we reported this in the third quarter, accelerate by 70% quarter- on- quarter. We generated $4 billion worth of pipe in the quarter. Couldn't service it all, obviously, so we are seeing demand beyond our supply to be able to service customers, and we're seeing it in every category, so startups that are building the next generation of whatever category they're in.
And oddly, interestingly, there's a fair number of startups I've never seen this before that seed or A stage have a- billion dollar on their cap table. So they're actually looking to build the next frontier model or the next specialized model. And that's consuming a fair amount of supply in the market. We're also seeing the scaling AI companies do something interesting. Those that built on the foundational models are also now starting to specialize open source or even build their own models. Michael Truell at Cursor has been very public about the need to expand their capabilities to ensure that they remain relevant and able to deliver the best autonomous coding capabilities in the market. We're also seeing software vendors move from just adding minor tooling to rethinking their entire offering, going from an AI-enabled to an AI-first capability.
Some companies are even looking at a complete restart in the way that they're thinking about being able to remain relevant and compete effectively against the AI-native players that are out there. We're seeing enterprises, as I mentioned a moment ago, very, very focused. But we're also seeing that focus accelerate in terms of their willingness to take up AI as a solution to whatever area of expertise that they're in. So we're seeing an extraordinary expansion in demand. By the way, we're seeing use cases expand, people going from training to optimization to inferencing. We're seeing modalities expanding. What was largely LLM generative AI is now including voice, video, imagery. So with all of this is coming a greater dependence on infrastructure being available like the kind that we supply.
So maybe this is a good segue for one hot question for you, Marc. And that is, the equity markets have corrected in the last couple of months. And your stock has corrected on a view that we're in an AI bubble, that you, CoreWeave, Microsoft, Oracle are massively overbuilding ahead of demand that won't be there to justify that scale of investment. What's your retort to that?
It's going back to what I just said.
That the demand is there.
The demand is there. I mean, we're seeing it in our pipeline. We're experiencing it with our customers. We're watching customers doubling every six-eight weeks in some cases. We're watching the uptake take place. An interesting anecdotal side note is watching what's going on with inferencing. Just to make sure we're on the same page, you can create a model. You can optimize a model. You put it into production. And then when your customers are consuming it, that's inferencing. So for us, that's actually the realization of the commercial delivery of the capabilities the customer has. Inferencing is growing at an extraordinary rate. As a matter of fact, there's a guy at The Times that is tracking inferencing. Evan O'Donnell has a blog where he's tracking all the different inferencing suppliers and looking at their scaling compared to CapEx investment.
What he's showing is that inferencing is growing 40% faster on a monthly growth basis than CapEx investment is. I think we're actually still lagging the demand that's out there.
OK. Your point from that is that actually there are outstanding returns on that CapEx so far.
That's correct.
OK. Let's talk about your need to stand up capacity. None of this is going to happen. You're not going to hit your forecasts unless you get those gigawatt-plus facilities up. So maybe you could give us an update on go-live progress on getting capacity online and the extent to which you've run into any unexpected supply bottlenecks?
This is a complicated industry. I just want to say that right up front. If you go look at my LinkedIn, you'll see I'm a software guy. We don't worry about, in software, physical delivery, so there's quite a bit of complexity in being able to orchestrate and deliver the capacity that's ultimately being sold by our organization, and what we do actually, going back to a statement that I made earlier, the team that we have is extraordinary. They have a lot of experience at being able to look at a greenfield project, being able to diligence it all the way down to the details. They don't take a cookie-cutter approach. The market doesn't actually lend itself well to cookie-cutter right now.
You need to be prepared to go to unique locations with unique configurations to be able to actually look at whether or not you can actually build and deliver the capacity. And our team is able to look at those projects and understand with confidence which ones are going to actually make it through to full maturity. And they do that by managing every single variable, controlling every aspect of the project so that they can confidently deliver the capacity that's being promised so that we can ultimately sell it to the market.
OK. What do you think is the biggest bottleneck right now? It used to be chips. Then we heard power. And then CoreWeave's CEO said a couple of weeks back that it's actually not power. We've got enough of that for the next couple of years. It's actually just literally building the data center shell. What would your contribution to that bottleneck debate be?
It's a reality that the lead times to go from handshake or signed agreement on a given project to it being delivered, there's some physical limitations in terms of how fast you can do that, so the reality is, if you're planning your delivery as one sequence of serial events, you're actually potentially delivering at the most extreme and not necessarily the best, most optimized way, so it's managing a series of projects in parallel that have different likely scenarios in terms of their ultimate delivery and fruition, and being able to manage that span of projects is a special skill. It's not something that, again, there's 10 companies that have done this before. Agile infrastructure delivery, that's a new thing, and what we're trying to do is apply the same agile methodology that you use for software to the challenges and discipline of delivering the infrastructure.
Got it. There's probably a role for AI in that process.
There certainly is.
Neil, maybe we'll bring you in as well for a couple of questions. I think the other aspect of these bubble concerns that go beyond just will demand be there is, will the financing remain there? Where it feels like there's been plenty of funds that are very willing to fund these data center build-outs. But the concern from the street is, will that be the case tomorrow? So perhaps you could comment on, let's say, the health of the financing market right now. And by the way, for everybody, I'm keenly interested in this subject given that I'm not a financial guy. But it's become a very important ingredient to this subject. So we actually invited two of the bigger participants in the financing side, Blue Owl Capital and Magnetar, to come to our event. And I'll be up on stage with them this afternoon.
But Neil, do you want to take a shot at that?
Sure. Yeah. It's a great question, Carl. I think one of our strategic advantages has been our ability to raise capital and deploy that capital effectively. So maybe just helping level set what we've done so far. After that investment in the summer of 2024, we received EUR 2.5 billion of cash. A few months later, we raised another EUR 700 million of capital through a pipe transaction with Accel and NVIDIA. That capital allowed us to really start to accelerate and grow our platform aggressively. Over the course of this year, we've raised another $5.3 billion in the form of converts and an equity follow-on offering. And recently, we put into place an at-the-market program that will just give us future flexibility to raise capital. So all in, we've raised about $8.5 billion of capital to date.
That's enabled us to go from one data center location in Finland to multiple locations around the globe. We will have expanded our connected capacity from 2 MW at the end of last year by almost 9x by the end of this year. And it's also enabled us to bring on marquee customers like Microsoft and Meta. So we've been able to deploy that capital and do that all with a net positive cash position. As we think about the near to medium term, we'll continue to look at capital markets. We think that there's opportunity to raise money through debt. We're looking at asset-backed financing. We're looking at corporate-level debt. And we'll continue to look at equity.
I think what's also really interesting is we have these other businesses and equity stakes that we think can yield us multiple billions of dollars in additional funds that we can use to invest back in our core business, so Avride is our wholly owned autonomous vehicle platform. They just today announced they've rolled out a robotaxi service with Uber in Dallas. We own a large stake in ClickHouse, which is one of the fast-growing AI database companies, often compared in the same vein as Databricks or Snowflake, and then we own a significant majority stake in an AI data labeling business called Toloka. Think of it as a mini-Scale AI, but we recently brought in Jeff Bezos and Bezos Expeditions as investors and some other high-profile investors in that.
So we think over time, as these businesses and equity stakes grow in value, we can use these to actually fund our business without having to tap into excessive debt or raise money through equity. So we want to be mindful of shareholder dilution and be really mindful and disciplined on the capital structure side.
Neil, let's also talk for a moment about what the ultimate profitability of this business will be. Marc, you mentioned a little bit ago about the parallels to, call it the hyperscaler 1.0 era, the CPU era. That era created gross margins, even though Amazon and Microsoft don't disclose them. Let's loosely say mid-60s%. It doesn't feel like any of the GPU clouds are on a path to that. But maybe they are. Can you talk a little bit about what the five-year-out margin structure of this business should be?
Yeah, absolutely. As much as we're focused on growth, we are equally focused on margin expansion. We are not a growth-at-any-cost company. If you just think about where we were at the end of December last year, we were still at a group level losing money on an EBITDA basis. The core business turned EBITDA profitable in Q2. In Q3, the core business posted 19% EBITDA margins. And now we're on track for the entire group, all of the businesses, including the core business as a group, to be EBITDA break-even by the end of this year. Back in April, we also gave our medium-term outlook. We said we expect EBIT margins, so factoring into account depreciation and amortization, to be in that 20%-30% range. And we are definitely focused on delivering on that. I think there's a few structural reasons why we think we can get there.
First of all, there's scale efficiencies. We are not just building a GPU as a service bare metal product. We are building a cloud business on top of that. That cloud enables us to actually drive better customer retention. We can charge more of a premium price for a premium product. And ultimately, we can deliver higher customer LTV. We also, Marc talked about having 1,000 engineers. That was a huge benefit for us. That actually leads us with operating leverage. So as we scale our ARR by 7x-9x in 2026, we don't necessarily have to scale our headcount by that much. So there will be a lot of operating leverage, having started with a strong team of engineers and infrastructure guys. And then on the core infrastructure side, we actually build our own racks and servers. This provides us with 20% lower total cost of operations.
Then finally, we just have very strong financial management. We take a more conservative view on depreciation. For Hopper, we depreciate those chips at a four-year depreciation standard versus many of our peers that are doing it at six years. It's not because we think the chips have a four-year useful life. It's just more of a conservative approach to accounting standards. Then when we also look at these large deals that we are doing, we are focused that these deals will actually deliver in that 20%-30% EBIT margin range. We are very focused on the economics of these deals and making sure that they are helping us drive towards better economics and better margin expansion over time.
OK. Thanks, Neil. Thank you, Marc. We're out of time. Really appreciated both of you coming, and for those of you that want to keep this conversation going, walk over with me to Ballroom A, because I'm about to hop on stage with Chase Lochmiller, the CEO of Crusoe, and Mike McNamara, the CEO of Lancium. You may know that Crusoe and Lancium built the Stargate data center down in Texas, and they'll be building out future OpenAI data centers, so let's keep this data center conversation going. I'll see you in 10 minutes.
Thank you, Carl.