Good morning. Welcome to day two of Nebius Live, coming to you from Berlin. There has been no shortage of conversation about AI here this week — or SI this week, if you prefer — from Nvidia GTC just down the road to Nebius's house, where we are broadcasting live right now. And over the next 90 minutes, we're going to be bringing you some of those conversations from GTC with the people that are building, financing, and putting AI to work. Here's a quick look at who is joining us this morning.
Plenty to get into. You'll see we've got lots of great guests. We are starting, though, with someone who is right at the center of the AI infrastructure boom: Nebius's founder and CEO, Arkady Volozh, joins me now. Arkady, it's nice — I feel like I've seen you a few times this week. It's, you know, Berlin has been buzzing and infrastructure is obviously the theme. Maybe you could start by just outlining how Nebius is — how you see Nebius as being different from the pure GPU rental place.
Well, what we're building is, we're building an AI-specialized cloud service provider, which means actually two big things. We build scaled compute, and we build a set of tools for people who then build AI, who train their models, who build their applications. And compute means that you need to build a lot — scale matters — and we are one of just a few companies in the world who are building now. We will be building more than one gigawatt of compute a year, hopefully more, but this is at least. And on top of it, we have the full-stack system. We build everything from dirt — from land, power, shell. We build our own computers, racks, specialized computers, optimized. Then we have our software from the basic layer — multi-tenant cloud, billing, auction systems, and so on — up to the agentic layer: first inferencing, token factory, and up to the agentic layer.
Right, and that software layer is really important. That's what sort of pushes you beyond just a NeoCloud and into a cloud provider. So what kind of customers does that attract? What differentiates you to them?
So if you look at the marketplace generally, you see big tech companies who have full ecosystems of their own services. They have their own compute, their own tooling, and they build the models and they build their own apps. And there's the rest of the market: a lot of AI-native startups, the corporate world, the enterprise world — they don't have their own tooling ecosystem, they need to use something to rent, and they have a choice: either to go to the big tech companies, the three biggest cloud service providers, or here we are. We built just the basic layers — compute and tooling for them. We don't compete with them in model building and applications. And we are actually — our past life, actually, we came from this big tech world. We remember how great it is to live in these ecosystems. We bring this ecosystem feeling to these new independent players, to the longer tail.
Right. And you know, we heard from some of your customers and other executives yesterday, especially in the physical AI space, how that lets a company focus on what they're trying to build and you guys can take care of the back end. I wonder, does the software and the full-stack approach give you guys a real edge, or are you mostly facing the same demand and pricing risks as everyone else?
Well, our clients are — we're working with the longer tail of clients: the enterprise market and startups who build. It's kind of the retail part of the market. We sell them not just bare metal basic services on which they need to build their own platforms — built up to the higher levels — and we provide like more solid, just more services to them, not just bare metal basic wholesale operations.
And I wonder if that gives you an edge in some of the pricing risks associated with building out large infrastructure like this.
It means that we need to build everything in advance, pre-built. Of course, we started by working with big tech companies who helped us. We served them on this basic bare metal market, which allowed us to finance the key part of our cloud. And now we have our own capacity, and we're growing in this capacity, and this capacity goes directly to end users.
Right. You just mentioned a moment ago bringing on one gigawatt of compute per year. Your CRO, Mark Boroditzky, told me yesterday that that would be one gigawatt of Vera Rubin capacity. You already have significant demand for it. What is the realistic timeline for getting that new capacity online and generating revenue?
Well, it will come throughout the year. Of course, it's skewed to the end of the year as usual because everything grows and you build more and more each month. But we will start getting new capacity as soon as Q1, then much more in Q2, and then a lot more in Q3 and Q4.
And then did you say it's a gigawatt of capacity years after that as well? One year after?
Of course, yeah, it's just the beginning. Again, it depends. We have more than around five gigawatts reserved already, contracted. We will be building it throughout over these several years and filling it with GPUs, which requires a lot of capital, as you know.
Of course. And speaking of capital, you often talk about these four C's. Capital is one of them. There's also capacity, chips, customers. What is the biggest constraint right now?
Customers, meaning demand, is definitely not the constraint.
Capacity. We are okay now with capacity; we have enough reserved. Chips — there should be enough chips next year, but it's becoming more and more... this probably will be the next bottleneck for the market, because supply chains for the chip producers, they are constrained as well. You see what's going on with memory, TSMC — like, chip allocation is a concern for this, like we need to watch it.
Is that a concern? Not for next year, but a year?
For 2027, getting enough chips?
I hope. But again, this is a thing to watch.
Will it be a bigger problem for the rest of the industry than it is for Nebius?
Well, we hope we will get all the allocations we need, but in general, the industry is constrained.
Right. Right. So would it be chips, you think, is the main one of the four C's looking out?
Yeah. Yeah. Capacity will be there. Capital — capital is also, again, but capital is also, as I said, it's not a problem, it's a task. It's how you get there, right? And we have a pretty healthy balance sheet. We always thought about being more conservative here. We used all kinds of instruments. From last week, probably some of the capital sources are temporarily closed, like asset-backed or debt market, but we're working now with customer advances and several other instruments which will probably provide us enough finance to build this gigawatt at least.
Nebius has taken sort of a much more disciplined approach to that balance sheet. What's behind that? Is that your history in running Yandex? Is that sort of a philosophy you have?
Just we're building a real company. A real company is measured by earnings, how much value you generate, and you just need to — a healthy business is a business where we generate cash.
Yeah. Let's get to this sort of public spot pricing that you guys recently announced. Started running auctions for capacity. We talked a little bit about it yesterday. What are you seeing? I know it's early days on pricing and demand so far. And does this actually change how you think about allocating capacity between big contracts and more flexible self-service customers?
Big contracts are important for two reasons. First, it provides us forward-looking advances — money to build. And second, it de-risks us because it's usually long-term contracts, five-year term contracts. But we don't want to give everything there. Of course, these contracts are priced lower. We want at least half of our capacity, maybe more, to be allocated to today's needs of impatient customers. They don't plan five years in advance, right? They just build a startup, have an idea, build a startup, they need capacity today, and we want to be able to serve them.
So, it's a balance, right? How much of your capacity are you comfortable then putting through spot and auctions versus locking those up with longer-term contracts?
It's hard to say. Today is probably half-half. Probably the majority should be on the main business, on the short-term contracts.
And I know it's early days, but who's actually using it so far? Do you expect larger customers to become interested in this side of it?
It's not like steps. It's like a gradual — we go from very big customers down to smaller customers, and then mid-size and longer tail. On the bigger names, we announced many of them. We work with NeoLabs, almost all of them, and the corporate market, and then...
But in terms of sort of the spot pricing and the auctions, do you think that larger customers will be interested in this?
They are, actually. For now, we have just manual auctions for them. We've run several runs of the auctions for like mid-range capacity — hundreds and thousands of GPUs. The fully automated, we're now trying this from this week, we're trying to implement on the longer tail, on preemptables, where people buy from one GPU to hundreds, let's say. And we plan actually to go up and eventually to auction bigger plots, bigger sites.
Right, and I know you've talked about the history of the company and how being in ad auctioning with Yandex has sort of prepared you guys for that.
Yeah, we used to work on demand-supply constrained markets, be it internet advertising or delivery markets. Yeah.
Where are you seeing the strongest customer pull right now beyond raw GPUs?
We again — we launched our token factory inferencing layer, and we have great pickup there, and that grows nicely. And we're working on the higher levels of tooling, the agentic level. We're now improving our agentic search. We're building a pretty big, massive web index which will be available to all of our customers, and many other tools in that level.
I mean, I ask because, you know, Meta is a big customer, consumer agents — they've obviously put out Muse, which has had a lot of popularity. Consumer agents as a whole are starting to scale. How much of demand are you seeing is coming from that kind of usage, and does it change how you think about capacity going forward?
That's a good example of yet another sector coming to AI. I remember a couple of years ago we didn't see demand. We were scratching our heads, having, I don't know, 30, 40,000 GPUs, scratching our heads who will buy it. And then wave after wave, applications started coming. It all started with chatbots. Then kind of what we're using today, all this advanced search capabilities. Then coding came — huge market. Then security market. Everything goes AI as we speak. And yet another wave now, in the recent days, months, is this individual private agents, starting from Muse. But there's another two, three startups which are picking up — they are our clients — and we see how fast they grow. It's definitely the next wave, and there will be more.
Can you describe sort of what that wave looks like at this early moment? Is that large compared to some of the other ones you've seen?
Yeah, it means that you have a couple of customers who grow like doubling quarter on quarter is not a record now — doubling month on month, and continuously for several months, and growing 10 times a year and more. That defines a wave. There will be more of this. This is just yet another one.
I think Jensen singled that out during his presentation yesterday, too. Just this huge wave of personal agents coming online. The inference that'll require...
That's because it's happening now.
Exactly. Arkady, wonderful to sit down with you a few times this week. Thank you for taking the time. Thank you for having us at Nebius's house, too. There's a lot of energy. This is a great spot to be during GTC week in Berlin.
Thanks, Arkady Volozh. Throughout the live stream, we will be showcasing some companies benefiting from Nebius's infrastructure, beginning with one that's using AI to help find a cure for Alzheimer's. Have a look.
Dementia is the leading cause of death in the UK and the sixth highest in the US. We are building foundation models that uncover new biology about the brain, focusing on Alzheimer's.
Having Nebius as a partner has really made that fast iteration that we need in order to really make meaningful progress at pace possible. We were also able to scale our model from 1 billion parameters, which is what we were doing ablations on, to 10 billion parameters in the matter of weeks. We're really pushing super hard, and we look to work with partners that can allow us to move even quicker.
A very close friend of mine recently, her mom got diagnosed with Alzheimer's disease. Working at Prima to try and really push the boundaries of what we understand about the disease and whether it is possible to diagnose it earlier and change the course of it — it's empowering.
Welcome back to Nebius Live. We were just talking GPU pricing with Nebius's CEO, and I want to stay on that topic because the way that compute is bought and sold is changing. As we discussed, Nebius has introduced live spot pricing for GPU capacity, with prices moving according to supply and demand. And so we are really starting to see the emergence of a market that looks more like energy or other commodities. I also asked Jensen Huang about this yesterday, and he pointed to the growing durability and value of GPUs and what that could mean for financing AI infrastructure. To discuss, I'm pleased to be joined once again by Nebius CRO Mark Boroditzky. Good morning, and...
Deirdre, it's fantastic to be here again.
And Carmen, this is the first time we've sat down in person. So, we've talked plenty. So, it's wonderful to have you, founder of Silicon Data, I should say, which has actually been a big topic this week at GTC. We talked about it yesterday in the pregame, and it's — yeah. So, I'm excited to dig into this with you guys. Mark, let me start with you, because Arkady just walked through sort of these early days of the spot and auction model. We spoke about it a little bit yesterday, but I want to know from the customer side, what has the reaction been so far? And so I was asking Arkady this as well: do you think that larger customers will engage with it? Are they starting to engage with it on this level, or does that take place somewhere else?
Well, at this point in time, what we're hearing from them, there's a lot of curiosity, recognition that the market is transforming. When I'm actually speaking to bigger customers, they want to understand what's taking place. They want to recognize how it potentially could play a role in their own needs. As a matter of fact, an unnamed customer yesterday, literally on the heels of our discussion, said we need to take a closer look at this. There's a portion of our requirements that we should be getting preemptable, and it should be driven by the price availability.
That makes a lot of sense. So it could be a mix.
Absolutely. Actually, that's the reality — just like the world is multicloud, right? It's going to be multi-AI stack. It's going to be multi-requirement. They're going to be looking at workloads, and if they have an experimental, early-stage workload, these preemptables are a fantastic way to be able to get the capacity easily and quickly to maybe prove something out, and then make the determination about how big you want to go and how much you want to invest.
And Nebius is very early to this. But this does fit in with, of course, GPU pricing. And Carmen, that's what you do — you track GPU pricing across the market. How does what Nebius is doing with live spot pricing compare to the rest of the industry right now?
It is definitely, I would say, the front-leading indicator for a lot of the new finance, the pricing structures, and everything else. We see traditionally, we see even people call spot pricing doesn't really change throughout the days, right? It's pretty static. So maybe it's changed — I think it's up to 15-minute intervals, which I think is great, especially for utilization improvements. And then we also see, interestingly, a very different demand-supply curve based on the chips themselves. Some of the chips, the variance is pretty sharp. Some of the chips is actually pretty steady, right? Which is a good thing. You want transparency. You want to make sure, hey, it's not all chips all created equal. Everything has a sort of different customer demand, to your point, right? This is great for certain customer needs and doesn't mean it fits for everything, but I think it's a great first step.
It's amazing to have the opportunity to partner with Carmen and her organization because, at the end of the day, to be able to create this market, there needs to be a market. Yes. And so when you see the prices below maybe benchmarks that we want to be seeing, it's only because the participation levels are too low, which, by the way, is a call-out: if you're not paying attention, you're missing an opportunity. There may be a discount there. But as we actually scale the utilization, as we get more customers utilizing the preemptables, my confidence is it'll start to move towards benchmarks that we're expecting.
Right. And this is the new partnership just this week, right? I mean, you guys did it fast. We just had the spot pricing, and then now you're helping to create a market for it. So, you know, it shouldn't surprise anyone, but the space moves fast. Carmen, we were talking about this chart yesterday in the pregame. How it's so fascinating — you know, GPU rental prices, everyone thought they were going to like depreciate over time, and the exact opposite has happened. Even Jensen Huang was telling me that he thinks the lifespan of a Blackwell is like up to eight years, maybe the same for Vera Rubin. Is that in line with what you are seeing? And then also talk about the dynamics of like token prices coming way down.
Yeah. So it's interesting — last year when we talk about GPU pricing, you probably see a similar trend when it's like GPU pricing will always come down. Remember last year? Well, the term structure was sloping down last year. Everyone consensus-wise thinks price will come down. This year, all the curves shifted up, and then it became contango. People think the price will rise in the future. Again, market consensus — I don't think, no one knows the future, but the market consensus is the price will rise. So what's interesting, if you look at token prices — token prices, it's a combination of all different models which you guys support as well. The open source tends to be cheaper than the closed source models. However, all the open source and closed source models, the price has actually been going up recently. It's just the mixture — people shifting from the premium, more expensive models to cheaper models that really drives our indices down, right, versus all the prices came down. So that's really interesting. So we call it compute crunch spread, you know, between GPU prices and token prices.
Another way to describe it is Jevons paradox, right? It's like literally the chart form of that, and that's certainly what you're seeing in the Nebius business as well, right?
Absolutely. And actually, let's take it a layer down lower. And the reality is customers are actually maturing and recognizing that they have options, and maybe the most expensive, highest-premium model isn't necessary for everything you're doing. And to the extent that they have the ability, they want to look at alternatives, and being able to work with open source models and specialty models is becoming a reality. So it is shifting in real use cases, and that's actually a big part of what we're seeing in the adoption of Token Factory and the way that customers are taking control of their intelligence.
Right. And that is something that I think enterprises are getting on board with in a big way, right? Is sort of model routing, using a number of them. That's what we've heard all this...
Is that what it was? Yeah, completely out.
And we're in the age of value maxing. CFOs are looking for justification and ROI. And that's pushing the teams to move more to more rational sources of compute and rational sources of models that are leading them to a lower overall average cost by being more in control of the way that they're applying the capabilities.
And yet we're still so early in that shift, right? Because, you know, a lot of tech and AI companies who we talked to are doing that, have been doing that this year. Quickly move from token maxing to value maxing — I like that. Do you see the rest of the industry, non-tech industries, getting on board? How quickly is that happening? And what I mean by that is, you know, not just the frontier labs, but a mixture of models, some of the open source models bringing cost down.
It's — we're starting to see it. It's early. I want to be, you know, frank. You're right about the fact that digital natives and AI natives are all over it. Enterprises, they're just starting, and that's why we're excited about, as an example, the partnership that we have with Palantir. Palantir is going to be helping us to bring this capability into the enterprise, giving them the resourcing they need to be able to look at the alternatives so that they can adopt the open source models and move off of the costly models they're on today.
Right. Carmen, last question to you. Do you think that if more providers — and we've heard that there's interest — start moving towards real-time or auction-style pricing, what does that do to the overall GPU pricing environment?
I think anything that introduces transparency and a variety of options is good for the customer and clients. So to everyone's previous point: if you want reserved capacity, great, they do bare metal, right, longer-term contracts. If you want something — hey, I just want it for the next hour or two, great, you can do on-demand, right, contracts. People want options, they want transparency. The more transparency everyone has, it's easy for financiers to predict the future revenues and also hedge the risks.
I completely agree, and actually I'd go a step further and say the transparency is actually driving prices up right now, because the fact that we have a market indicator when we're having conversations with customers, it's no longer us justifying the price. It's like, that's actually what the market is telling us, and it makes it a lot easier for us to have an honest opportunity created with that customer.
Right. So that translates into more demand, right? More supply constraint.
Well, they go even further. The customer has greater conviction when they know that they're spending at a fair value.
Right. Carmen, Mark, thank you both very, very much. It was a great discussion.
Deirdre, it's our pleasure, and it was a great time.
Thank you so much. Let's move now from the cost of computing to how you put that power to work. Take a look at Sakara.
We help scientists solve some of the most challenging problems using computational analysis. I'm Evan Floin. I'm CEO and co-founder of Sakara. Our top customers, they're primarily trying to solve problems around drug discovery.
If you look at them, like for example Sierra, the development of new medicines, and in some cases even use that information to generate personalized medicines for things like cancer vaccines. And what we're enabling them to do is to write software that does analysis, and that's able to scale up that analysis really across very large data sets and kind of spin that up into different compute environments.
Early on with Nextflow, we saw more and more there was customers who were wanting to go into structure prediction. But how are we going to take some of the protein design work and use some of the latest models and run them with GPU resources that were required? As part of that exploration work, we came across the Nebius cloud, and we saw that there was this ability to link in Nextflow executions in this way using some of the separation of compute and configuration. Through that, we were able to put some of those workloads and even run them through some of the NVIDIA NIMs. That gave us huge increases in terms of speed up, really sort of 4x over launching containers natively, and something that we see customers now starting to adopt more broadly. In my personal experience in terms of working with Nebius, work that we've done with joint customers, that's something which I haven't seen before in terms of working with cloud providers. So something that we know we're really excited about expanding the partnership and being able to, you know, run even faster. With the speed that AI is moving, I predict we're going to be using a whole lot more compute. Scientists are going to be able to run many, many more experiments. Much of the grunt work that's involved today in terms of handling files and configurations, even writing code, those aspects are going to melt away very quickly. Combine that with the sort of reasoning that you can apply to your analysis, we're going to be able to see a massive, huge scale-up in the number of experiments that we can run. I'm personally very excited for the future of this.
Welcome back. We have talked a lot about the demand for AI compute, but for enterprise customers, access to GPUs is just part of the equation. They also need storage, software, infrastructure to run increasingly complex AI workloads reliably and at scale. Nebius and Vast Data are working together to address that and to make it easier for businesses to put AI to work. Joining me to discuss this are Jeff Denworth, co-founder of Vast Data, and Laurel Roseman, VP of global partnerships at Nebius. Good morning to both of you. Thanks for making it out to Nebius's house.
Jeff, maybe we'll start with you. When an enterprise commits to AI workloads at the scale that they do these days, which is thousands, maybe hundreds of thousands of GPUs, what does it need to trust about the infrastructure before making an investment?
Well, first of all, I think if you're going to make that type of investment, we're not talking about small amounts of money here. We're talking about, you know, millions to, in certain cases, hundreds of millions of dollars. You know, our perspective is things need to work first and foremost. And this is a market that has very deep roots in high-performance computing, a market that has very deep roots in research and development, right? And when you make the transition to productionizing enterprise AI, people just expect 24/7, always-on infrastructure, and that is really, really hard, right? And so building the right systems and deploying them intelligently with partners like Nebius is really important to get right. Otherwise, you just wasted all your money.
Right. It's kind of the whole goal, right, Laurel, is making sure that it's reliable and always online. So how do you sort of assure your partners — obviously through your history — but what do those conversations look like?
Well, when you speak about enterprise customers, to Jeff's point, right, they're not like an AI native where they want to control the entire infrastructure stack. They want to know that it's a large investment. It's probably a new investment for them getting into this AI transformation space. They want the outcomes. They want the value. They want the ROI. And making sure that they're choosing a reliable partner to help them do that is critical to it. And just buying GPUs, putting it in a colo — the hardware alone is not going to help them get the outcome that they want in the performance. So choosing a partner like Nebius and working with partners like Vast are really crucial to making sure the customers are getting the utmost performance of the infrastructure, because it's all the way to the software and application layer.
Right. And the companies that you're partnering with, how much do they know about what it takes to put this on? Or are they kind of like relying on you to serve that to them?
I think it's a give and take, right? Like it's a new industry for a lot of people. To Jeff's point, they've been in the HPC space for a long period of time, but this is uncharted territory for a lot of people. And they've been partners with these customers for decades. And part of their digital transformation journeys are looking for that education from them and that enablement from them. And being a partner with Nebius, you have to provide that level of enablement and education to help them.
Right. Go as deep as they want or serve it all to them.
Game changes like almost every week, right?
You need to work with some of the smartest and most nimble organizations just to be able to keep pace with what's happening, right? Whole new processor platforms on an annual basis, new white paper that comes out like every two weeks about how... new architecture, everything changes. So you have to work really quick, and you have to do that responsibly too.
And you have to have the right partners who are able to move quick as well. I mean, Jeff, talk a little bit about what Vast does, because it's very fascinating. It's really like bringing efficiency into the equation too, right?
Yeah, we bring a lot of stuff to the table. So the company's 10 years old, actually founded a few weeks after OpenAI was founded. And we set out to basically redefine data infrastructure with an idea that deep learning would change the entire computing stack. And so we've been at it for a decade, arguably the largest player in the space today, work with some of the best possible AI cloud partners like Nebius. And we're basically redefining how data is managed — so storage, database services, compute on top of all of that — in a way that is designed for extreme scale, with enterprise-grade resilience, service-provider-grade industrial kind of capability and security, and all put together in one simple package for customers.
And how long have you been working with Nebius?
Three years. Three years, I think we're at it now. Maybe two and a half.
I mean, we've been talking a lot about how Nebius has been so early to this, and how you guys were thinking about all the different layers early on. And there's a lot of NeoClouds, data, whatever you want to call them, coming online, and not all are created equal, of course. Jeff, yeah, talk about that a little bit, and like, have more been interested in working with Vast, trying to move up the stack?
So I think it's fair to say that we have probably the most experience with the AI clouds of anybody in the market, and you see such a wide spectrum. Like half the people coming to the table now are just bankers. Somebody has $500 million and they want to get it, put that capital to work. You know, I think of capital as the sixth layer in the five-layer cake. And there's just such a wide spectrum of capability, and people don't understand the complexity associated with deploying all of this. And so, you know, it's no secret that Nebius has really figured out the game and does things ultra-professionally. And that's really important when big dollars are on the line, which a lot of customers are coming to the table and spending.
So Laurel, you know, how do you know that partnership is delivering value? What matters most to an enterprise?
You know, a partnership is delivering value when the customer keeps growing with us, right? And customers, especially the enterprise, will start small and they'll test the waters, see what works. And if you start seeing they're bringing different workloads and continuing to increase their projects on Nebius, experimenting with new things, continuing to optimize and leverage us for different, you know, products that we're bringing to market, is how you know partners are bringing value, because they're consistently staying with the customer to help them grow in our platform.
Right. And adding more. You're helping the businesses on the back end as well, I'm sure. Jeff and Laurel, thank you so much. Great discussion.
Oh, wonderful. Thank you for having us.
Thank you, Jeff and Laurel. One area where we are seeing growing demand for AI infrastructure is healthcare and life sciences — from medical research to drug discovery. Nebius has been expanding its work in this space in particular. Have a look.
For the last couple of decades, drug discovery has gotten more expensive and less efficient, and exponentially so every single year. And only like 10% of the drugs actually make it to market. The complexity of biology is far greater than the complexity of the written word. One of the biggest challenges for healthcare is its data. I want to see us as an example of a company in the bioeconomy space that can turn a very difficult environment into forward momentum. The world we're living through now with these advances in AI is a super exciting time for healthcare, because it's changed so dramatically from 10 years ago. If we design a new molecule and put it in the market in less than 24 months, that is absolutely insane. It's really about taking the ideas that exist in science and really being able to turn them into code. The work that goes on in this lab generates petabytes of data that is used to train and fine-tune our foundation models. We're storing petabytes and petabytes of large sequence data. We're working with that across tens, hundreds, up to thousands of GPUs. This is a superpower that we're giving healthcare professionals. This is a superpower we're going to give patients. But this superpower comes with great expectation. Who can be the fastest to provide compute at a really large scale? And that's where Nebius comes into play. I don't have to think about what am I provisioning, how many instances do I need. I just log in and it works. We could reduce the time from like months down to just a few hours. We don't have time for infrastructure. We need to do science if we actually want to change the world the way we want to. Not only are we bringing the chips to market, we're also bringing the innovation and changing the mindset. We want to democratize care, and the solution for that is through AI. A very close friend of mine, her mom has been recently diagnosed with Alzheimer's disease. Knowing that I'm working at the frontier to help, it's really empowering. By 2030, I expect we will see truly great advancements in biology across all of the domains. We are going to need an exponentially higher amount of compute and computational capacity to take us to the next level. It's incredibly exciting. I'm personally very excited for the future of this.
Healthcare is one of the areas where AI could have a huge impact, but it's also where the stakes are highest. It's one thing to build a powerful AI model, but it's another to make it reliable enough for use in hospitals, where patient safety, privacy, clinical outcomes — of course, they matter. Joining me to discuss where we're actually seeing progress are Thomas Knox, co-founder and CEO of Vitvio, and Dr. Ilia Burkov, global head of scientific AI and healthcare at Nebius. Guys, welcome. Good morning. Thanks for being here. Thomas, let's maybe start with you. Give us a little bit on what Vitvio does.
The long and the short of it is we use ambient sensing computer vision to break down the operating room into 3D, to then deploy agents that act like another staff member within the space to help take over administrative work. Basically get the staff in the room back focused on the patient as much as we possibly can.
And what have you been observing? What decision or workflow does that improve?
I mean, a lot of this is about the clinician time spent. Like, for example, every hour that you spend on an operation, it's often, you know, one to two additional hours of documentation and paperwork. And we've been trying to make surgery safer and, ironically, more...
efficient, but the go-to has always been with these like human manual input, and it's just stacking over and over. So the whole goal is transition back to letting the clinicians be clinicians, right? Um, Ilia, you know, when we think about healthcare, we think so much matters with the model and the iterations and the feedback loops and so on. But how can infrastructure actually accelerate scientific discovery, um, without encouraging teams to skip validation?
So when you look at models, when you look at the multimodality of, uh, healthcare, you need to take into account a lot of information, a lot of data. Making sure that the model is the right one, making sure that the infrastructure and the model fit, uh, is absolutely key. So you're dealing with very sensitive data, with very patient-identifiable information and, um, clinically important information. So making sure that the model is the right one and making sure that, uh, the information that you're working on is going to be critical for that particular modality is key. Um, infrastructure is a big influence within that, having the right GPUs matched with the right workloads. When you're looking at simulation works, you're going to use a different type of GPUs.
If you're looking at training a model, you're going to look at different types; inference, different types. It's getting those things in line, uh, that makes, you know, the importance of what we do. Um, and we're working with teams like Fitvio and others to align on the workload specifically. Um, that...
Right, it's not only about like what's the right model, right process, because that's changing all the time, right? It's a moving target. Um, so, you know, how do you evaluate a system, um, when everything is changing, performance is changing all the time?
I mean, for us, it's all about the impact. Like, stop looking at the benchmarks of certain things because, you know, if you overoptimize for that, it doesn't really give you a clear picture. But what is the impact that it's having, especially in healthcare? Um, if you have a beautiful model that doesn't actually make an impact, it's worthless. So we try to focus on like, how do you drive that impact and extend it rather than optimizing on sometimes just the technical evaluation of it.
Right? Um, what is that impact? Is that time saved for the doctors? Is that how patients feel?
So healthcare organizations in general are facing like an existential crisis right now. Um, you know, they have aging populations, rising costs, lack of medical staff. Like, I think WHO published, um, that by, I think it's 2031, we're going to have like, I think it's close to 11 million, uh, clinical, like, shortage in clinicians. That's not a challenge that you can solve other than by implementing technology that can help make the clinicians that we already have more effective. So that can be clinician time, but it's also patient outcomes, because if you have a clinician that is squarely focused on, like, on the patient...
Naturally, they are going to be able to provide better care to that patient as well, because right now we try to do our best with it, but what you end up getting is a clinician that's very distracted. They have a lot of things that they have to document. It's not to say those things shouldn't happen, just shouldn't be them doing it.
Right? So if you solve like one piece of it, it often benefits another piece downstream or upstream or whatever.
Absolutely. The future that we see is like, honestly, I think that if a clinician is touching a computer, and you know, 5 to 10 years from now, we've all failed.
They should be seriously sitting there. They should walk in, it should be very like, no, I mean like the interface should be, they walk in, all of this is running in the background contextually, they get what they need, and they don't have to be sitting there typing on a keyboard anymore.
They're spending time with the patient.
Yes. Spend time... spending time with the patient. Right.
Like we talk about like AI, AI should be driving the most human connection we can possibly have. And when you talk about shortages, I mean, one of the most famous examples, I think Gavin brought this up in the pregame, was radiology, right? Everyone thought that AI was going to lead to the end of radiologists, but in fact, we've had more. Ilia, do you think that that's a trend that we're going to be able to see in other parts of healthcare?
Absolutely. It's a tool. It's an enablement for a lot of the work that's being done. Um, radiology is a perfect example of that innovation, that clinicians don't need to look at every single thing. It's more like the tool is saying, 'Have you checked this? Have you seen this? Pay attention to this.' They highlight, um, carcinomas, they highlight various, uh, features that they've been trained on to say, 'Look here.' And radiologists use those tools to make their lives easier. And you're seeing as a radiologist 10, 20 scans every hour, right...
...of a patient, of different patients. And it's incredibly difficult. 10% of, um, fractures are missed in radiology. Not because of the incompetence of the radiologist, but just because they're overstretched...
...with the work.
You know, I have my own like personal example of this. Um, my OB/GYN, she's like 70. She was on the verge of retiring, and she was using Open Evidence, right? And that actually gave her so much like empowerment. She was not going to retire. She's going to become a primary care doctor. And in the US, right, those are in such shortage. So it was interesting how it's just a small anecdote, but you know, a personal one where AI and these tools are enabling more in healthcare. Um, Thomas, one of the tricky parts of this obviously is privacy. That's why some doctors, some systems, medical systems, hospitals, you know, are not or are maybe a little cautious. So how do you protect the privacy of patients and staff while collecting enough useful information in operating rooms?
Yeah, I think it comes down to like two pathways. First one, uh, is more of a policy of how you work within a system. We have gone through multiple union reviews and we've never had a single challenge. We've never had to have a single, like, a second conversation afterwards because we follow something that we call extreme, uh, transparency, which is I'm going to tell you exactly what I'm doing and I'm going to tell you exactly what I will not do. And as long as you agree to that and you actually keep to that plan, people start to say, 'Okay, great.' Because no clinician on earth wants to be doing documentation. Nobody went into this...
...So you tell them like the payoff.
Yeah. Here's what you're going to get out of it. But we promise we're not going to look at these things because the whole point is that it's not, we don't want to be coming in and be a big brother. We want to come in and actually help you, and you, by doing this, it will drive every other indicator. The second thing is zero trust, zero trust, uh, infrastructure, which is anonymize at the point of capture, minimize as much as you're taking, and allow the hospital to be the controller that tells you what is good and what is not, and follow their lead.
And yet this is an industry with a lot of like old school thinking. So have you had a challenge in getting that message across, or has it been mostly smooth?
I think that, you know, when we first started the company, there was, um, you did take, uh, some convincing, you had to find the right people, but in the beginning of any, you know, uh, new venture, it always comes down to finding the people who see the future of where it's going. Now I would say it's probably, it's not a challenge at all. Like we don't have that conversation very much anymore. I think hospitals, hospital executives, clinicians in general, they all see where this is going and they now are starting to recognize, okay, we want to be a part of that rather than push against what they consider now to be inevitable.
Right? Um, that's interesting. So, you've really had a move towards these tools. Ilia, last question for you. Um, I always find this so fascinating in the Bay Area that startups are like often asked about their compute strategy. Of course, that is so important, but biotech particularly, um, what does a credible answer look like? Right? You have to balance like, yes, there's going to be demand, but it's expensive and you have to plan for the future. You have that cone of uncertainty that Dario Amodei and others talk about a lot.
Yeah. Uh, the thing that we always say is plan in advance and get more than what you planned for, as always, because of the challenge.
You need to have a lot of money for that though, right? And the uncertainty.
It's uncertainty, but it is, uh, reassurance of what you're doing. And many VCs that we talk to now, they're saying as part of the raise, they need a guarantee of compute...
...already in place. Uh, because there's no point in raising 50 million, 100 million and having to wait for half a year for your compute to be available. So it kind of comes together nowadays, and we're seeing a lot of this, and what we advise is get ready as soon as possible.
Yeah. Uh, Thomas and Ilia, thank you very much. Great discussion.
Thank you very much, Andrew.
Okay, we are moving on. Now, just before we went live this morning, I caught up with Oliver Vince, co-founder of Base Camp Research. They are developing an AI model called Eden, which is trained on evolutionary data with the aim of discovering entirely new approaches to medicine. The company has just raised $140 million to advance its work on therapeutics. I asked Oliver what that funding will allow them to do next.
So at Base Camp we are trying to build large AI models that ultimately can ingest disease biology and output potential therapeutics for patients. Right? So the idea is that, you know, there are lots of diseases today that we can't treat, we can't access, we're not on track to treat, and this is a, you know, potential exciting new way of developing, um, medicines, treatments, protocols for patients that currently don't have any.
Right. And so when you say you're developing a large model, that makes me think immediately of how much compute is needed, and I think you trained Eden on a thousand GPU run with Nebius. Talk a little bit about getting the data, how much you need for this, and why that's important.
Yeah. So the basic idea behind Base Camp is we're building large language models, but rather than on human language and French, English, Chinese, anything like that, it's on DNA. So DNA is the fundamental language of biology. And so what we do is we train larger and larger models on DNA to read DNA and to write DNA. And the reason we're trying to do that is because the disease biology is described in DNA and the treatment for that disease biology is given out in DNA. Right? So you're trying to get a large language model that can map between those two things.
And, you know, ultimately that requires scaling up the models. They follow the same sort of scaling laws as we get in natural language. The larger the model, the more DNA it's seen, the better it can do both of those things. And so obviously working with Nebius on some of the compute has been an exciting journey on that.
Right. How do you actually train a model on DNA? Right? So much of, you know, what I've been told in AI is like making it understandable for humans. We're trained on the world's, you know, internet. Exactly. Um, so what does training it on DNA actually look like? Like what's the input?
So the big challenge we have in biology is that the internet of biology, the comparable, like, large resource of everything known, is non-existent, right? So 68% of all information we have about life on Earth comes from about five species, right? So we are incredibly biased by what we know about life on Earth. And there are trillions of species that are simply undiscovered. They're out there in the natural world and we don't know where they are.
So we run expeditions to 32 different countries at the moment, uh, expanding beyond that. We've run, you know, hundreds of expeditions all over the world. This is teams of like explorers going out to the volcanoes, ice caps, rainforest to collect this information and then feeding it into large language models. And the really, really cool thing is that this is done with what's called access and benefit sharing agreements, so that those who contribute the data, they get upfront investment, they get scientific partnership, and also when we make money out of it, they get a royalty back from the data and from the outputs of the models.
What are some of the more exotic places Base Camp has gone to?
Well, we have public partnerships in places like Malawi, Cameroon, Costa Rica, Chile, Argentina, uh, all over the world, right? So this is, you know, this is not just working sort of close to home where the vast majority of data has already been collected. This is not just in the West. This is working all over the world with people on the ground who want to contribute to the scientific effort, uh, and can see some reward from doing so.
So, what does this look like if it's successful? Are you going to be solving certain diseases? Is that the goal?
So, the easiest way to think about it is there's lots and lots of diseases for which we do not have a treatment today. And getting to a treatment, as far as we know, is far beyond the level of complexity or progress of our current medicines. However, we do understand that there, you know, if we could truly program biology, we, you know, we could like read and write with perfect fidelity, there is a technically feasible path to curing some of the more complex diseases. Now rather than, you know, guaranteeing this is going to work on certain time frames, it's more that there is so much potential to treat those people, to deliver that impact, and there is such a clear technically feasible route towards that that it's our mandate and our responsibility to prosecute on that.
Can you give me some of the examples of diseases you think are, I guess, is the word solvable, earliest solvable?
So the diseases that we're focused on initially are those that are diseases of the DNA. Right? Our model understands DNA and speaks DNA, and therefore it's diseases like, I mean, particularly cancer where it's DNA that's gone wrong in your lifetime, or rare genetic disease where it's DNA that has gone wrong or like, you know, is causing some, uh, harm before, like, when you were born, right? So both of those diseases are categories of diseases where we know kind of the causative drivers or we have a good idea of what they are, and we have a good understanding of what it would look like to fix it, but the level of complexity of doing that is far too... And so what you need ultimately are models that can, you know, work in that language with a high degree of understanding and can design solutions to those problems.
And so you're building your own model. What is that based on? Are you starting from scratch or do you start from like an open weight model?
No. So everything we do is fully, um, pre-trained ourselves. So we're one of the very few, um, biology AI companies that both has its own unique data set, pre-trains from scratch, and then translates that into medicine. And so there's a whole series of challenges in doing that, especially as a small team. But, you know, I think what we're doing over time is figuring out a lot of the bottlenecks that are going to come for the larger companies over time, uh, but doing that sort of ahead of where they are.
I feel like we've been talking about bottlenecks all week at GTC, but operating in healthcare is a particular, you know, challenge. I think we started off the interview saying it's one of the most regulated industries. How is that piece looking? Are you there yet?
