Hello everyone, welcome to Nebius Live, coming to you from Berlin where the AI industry has gathered for Nvidia GTC. We are live this time from Nebius House, just down the road from where Jensen Huang delivered his keynote only less than a few hours ago. There is a lot to unpack from it. AI infrastructure that is making headlines almost daily. You've got enormous investments, growing demand for compute, and you know, big questions about what comes next. We have got a great lineup of guests joining us to make sense of it all. Here's a look at just a few of them that are coming up.
We've got Lars Sophie Botour from The Tech Translator, Andrei Kolenko from Nebius, Alex Banks from The Signal, and Kelly Kong, investor at Accel, many more as well. It doesn't even stop there. We will be live again tomorrow from 10:00 a.m. Berlin time with another incredible lineup. We start though, of course, with reaction to Jensen Huang's keynote.
Nvidia is at the center of the AI infrastructure boom. Billions pouring into new data centers. Demand for compute continues to grow. The industry really grappling with how to build capacity fast enough to keep up with demand. That was certainly a theme throughout the day. A lot of demand. So what did we actually learn from Jensen today? Here's a few of the key highlights. Nearly $10 billion invested in Europe's AI ecosystem. New AI factory design cuts deployment from 6 months to just 14 days. Those containers that you might have seen.
Same AI output, up to 30% less power. Nvidia also launched architecture to secure AI agents. Safety is something that Jensen talked about at length. Incredibly important at this moment. So what does this all mean for the companies that are buying and deploying that compute? Joining me to break it all down is Nebius chief revenue officer Mark Burditsky. Mark, hello. Welcome. Thanks for being here.
Thanks for having me.
Dietrich, it's a pleasure, man. We are thrilled to be hosting you here at Nebius House. And this is the first time that we've had this kind of an opportunity. Nebius House is new in our overall execution, and it's great to have you in this inaugural experience.
Great. If the viewers could kind of like see beyond me, there's this whole house, and there's people working. It feels almost like a hackathon house or something in San Francisco.
Pop-up Eastern European headquarters for us.
You're right. Actually, there's dozens of customers coming through all the time. More interest and attention than I could have imagined when we put this idea together. And I think this is going to be a standard in our playbook going forward.
Right, because this is the first GTC, right? In Berlin.
This is, and the first Nebius House, yeah.
House, which makes a lot of sense. So, let's talk about the keynote. By the way, this is the postgame. This is great. I did the pregame, postgame. We're going to break it all down. He covered a lot of ground in that keynote. I wonder what stuck out to you.
Well, first of all, you got to thank Jensen. He brings extraordinary energy and focus on what's going on. He elevates the entire market's focus and attention on what we need to be doing next. And we can already feel it in the conversations I'm having with customers. He's definitely brought everybody along. And he pointed out the fact that Europe's making big strides. And we're seeing it as well in terms of the kinds of customers that we're winning that are building the future with models and next-generation inference capabilities.
Really thrilled about the early adoption that we're starting to see with industrial AI. Jensen focused on key industries like the healthcare highlights and the manufacturing highlights. And obviously, he had a great time laying out what's going on with physical AI and robotics.
The robots on the end, I think, were very cute.
I think he brought it to a new level. I mean, he humanized in an interesting way that...
It's not as threatening when you're playing with a little robot on stage, right?
Right. He brought it back down. These were agent swarms. They were cute little ducks.
Again, he made... yeah, cute little ducks, but at the same time, he may have just sold 20 million of them.
Well, he said right away on stage, he said, get on their website, sign up now, and you know, this company of 10 people are going to get 20 million orders from everybody that's watching this streaming.
I would rather one of those ducks than like an actual humanoid at this point, but you know, we were talking to some of the people and they said maybe 3 to 5 years before you have those humanoids actually in your house, which I thought was like earlier than I expected.
It's a, you know, it's been an interesting promise. I mean, we've been watching it in popular movies for a long time and we have the good fortune of working with companies that are providing the engineering and enabling technology that are building the models, but also building the robots. I think you're right that it's a couple of years out. I for one wouldn't mind the help at home. I think my wife is looking forward to somebody else taking care of the laundry.
But the reality is, you're right. It's not tomorrow. We are seeing the fleet of robots in industrial applications. You know, we've all seen the videos of the Tesla robots and the Amazon automated warehouses. And I do think that more industry is looking for that and we're hearing it as well from manufacturing customers in key industries.
Mhm. So, what do you think the thing your customers are going to be calling you about tomorrow? What do you think resonated with them? Or maybe you already know.
Well, we already know because it's already happening. I mean, the conversations we are having at Nebius House: urgency. Urgency. I think the reality is people have been forecasting eventually supply will catch up with demand and that's just not happening.
And it seems like every one of these GTCs, people are just re-forecasting when that's going to happen. And folks that have any kind of plans are realizing they need to be thinking now about how they're going to take care of those plans. So, we're hearing from customers about what they need late in '27, early in '28. And candidly, we're going to be in the same situation as everybody else, which is we don't have enough capacity to support them.
And so, you guys are doing something very, very interesting. We talked about this before. I think you guys just made the official announcement. I have the website up myself because I'm just fascinated by this. You have GPU spot prices and auction. Can you explain how that works?
Certainly. Really simply, well, first off, let me add a little bit more context because I think you did a really good job with the lead-in, but the reality is pricing in this market is opaque.
You have a lot of deals being done in order to be able to secure the capacity that people need. And then vendors turning around and explaining what things are worth, and there's no confirmation. You know, in some respects, it's sort of like DeBeers and diamonds. It's what the buyer is willing to pay as opposed to what it's really worth. And Arkady has been very clear: we have to switch the paradigm from vendors picking customers to customers picking vendors. Return to a more transparent type of experience. And one of the critical factors that's missing is pricing.
You know, that sort of DeBeers diamond experience, that needs to go away. We need to create a visible recognition of reinforced value by the customer. So, spot pricing is where a customer can get access to self-service, preemptible compute for short, focused projects. So, these are engineers that are willing to accept the fact that maybe the compute will turn off after a certain amount of time.
They set a price range. As long as the algorithm-driven price is inside their price range, they get access to compute. So, effectively, we have a real-time price-driven recognition — I mean, a real-time demand-driven recognition of what price people are willing to pay. And you can see on our website now where those prices are going.
It's pretty amazing. I'm looking at like a B200 up almost 80% in the last 7 days. I know it goes up and down, and it's such an interesting idea because, and Arkady was talking about this earlier today, but how sort of the history of Nebius allows you to do that. Used to be, you know, Yandex and advertising and that has auctions. Do you think others in the industry can or will get on board with this kind of spot pricing and auction?
There's a lot of curiosity out there. I can tell you — oh, yeah, oh, yeah. I can tell you a couple of hyperscalers actually asked us about our experiences. We had previously announced an auction that we had completed and actually we're doing auctions in addition to the spot pricing. And there's a recognition that the opportunity to provide customers a more clear, validated price is — everybody's looking for it. So, yes, we are getting a lot of interest and attention from across the board.
Okay, that's interesting. So, any other takeaways from his keynote? What's next?
Well, what I loved about the keynote was he actually also spent a fair amount of time about Vera's. And Vera Rubens are critical to the industry right now. Folks are excited about the next generation compute. We're hearing it mostly from model builders. But now we're also hearing it from the broader portion of the market. So, getting Jensen to share, you know, what's valuable and why Vera's are important was super helpful. We have a gigawatt worth of Vera's coming next year and already have significant demand for them and it's great to be having Nvidia leading the way making sure that folks know that this is an important next step for all of them.
All right. Well, it's been a pleasure getting your feedback and your insights on the keynote. I'm going to talk to you again tomorrow, right?
That's right, Deidra. I'm looking forward to it.
Make sure the audience tunes right back in. This is a two-day postgame.
That's right. That's right.
Thank you to Mark Burditsky. Now, while we've been breaking down Jensen's announcement here at Nebius House, we also caught up with attendees outside of the keynote to get their first reactions. Here's what they had to say.
This is my first time in GTC. Everything is amazing.
Jensen's speech is a fantastic speech.
I really like the ending, too, because he said that it's all about Europe, that we are all together.
Very proud about Europe technology. I thought that we were more behind, but really, really impressed.
Very, very interesting for me, especially for the robotics part. What they are developing together with Agile Robots.
What excites me a lot is the physical AI, the robotics. I think this guy has huge value.
It was really awesome, like seeing my products there, like working at Hugging Face. Like I think they are quite huge.
Number one key takeaway is that it's not difficult. So, these are tools that anyone can use and they are providing the infrastructure to make it easy to develop applications, to scale, to enter in the physical world with AI.
I really liked the introduction of the DSX superpod. So, I'm very interested to see that in the digital twin and also play with it.
Jensen reflected on how far we've come and how far, even in 2016 and 2019, he was talking about OpenAI and how, you know, he was giving 101 petaflops of compute to OpenAI in 2019, and today you can get that same compute on, like, a laptop, you know, that you can fit in your backpack.
What Nvidia built with sandboxing is really the future and they think it's really important to focus on security.
There's no doubt that AI is going to be evaluated and continue growing, and especially for the AI agentic. And also I'm working on the AI infrastructure and the requirement also of building more powerful interconnect between GPUs and GPUs.
I'm very excited that it's going to be very acceptable and very easy to put in practice, like the robot trying to help others. For example, my father had cancer and the thing for radiotherapy was very, very — for me it was too much. Like I was very impressed. I'm very glad that things like this are going on. Yeah.
Also watching from the keynote floor was Laura Sophie Botor. She spends much of her time translating complex technology into something that the rest of us can understand. She has built a large following doing just that. Laura, thank you so much for coming straight over to Nebius House. I was in the room as well. There was a lot of energy and this is your third GTC. Is that right?
How did it compare to some of the others?
Thanks, Steve, for having me here. Well, it is different. It has been a little smaller, more cozy, you know, in sweet Berlin here in Germany. I have been in Paris, in San Jose, now here. But still, it doesn't need to be very big, you know?
It was nice and intimate, I thought.
Intimate. And I think that there is some, you know, people really want to get on track, you know? And I believe that Jensen is also a big believer in Germany, as he knows that we have great talent and we have great industry knowledge. And I think you need to now implement AI into industries and that's what he sees. That's why we also need big players like Nvidia coming into the game and I think Germany has great potential there.
He highlighted a lot of the other big players in sort of the Nvidia ecosystem, which was great to see. You explain this stuff for a living. So, what is the version of the keynote you would, you know, tell your friends over dinner?
Well, I love that Jensen is pretty good also in explaining technologies. He's always putting that into — this is the five-layer cake of AI economy, you know, with compute and energy and AI factories and so on. Those are the five things you need for a safe agent. And he's always breaking it down very well. Still, he's pretty, you know, human, while he's a genius. And then he's also bringing robots on stage all the time, you know. So in general I can tell that he believes, and I believe that too, that every company needs to become an AI company. And the next stage is that we all need to integrate physical AI. And that's where we are right now.
Right. And what is your audience? What were they hoping to see out of Jensen? What do you think resonated with them?
Well, as I'm a big believer in European technology companies, I was hoping, and I think my community too, for some optimism, you know, from him.
But I saw that and I could also tell from use cases he was demonstrating.
And there were — I love to see like all those different companies from Europe that are on stage, like Deutsche Telekom for example, like Siemens, like SAP, but also like Agile Robots. He is now investing 10 billion dollars in the European ecosystem and he's collaborating with plus 12K partners from here. And that's what I enjoy seeing because Europe belongs at the table and is, as said, pretty good in his industry knowledge. And for example with Agile Robots, they are now also training 2,000 robots right in the heart of Europe in Munich.
And there are great investments, great partnerships coming to life.
You know, he was in Europe for a stint, I think a week or more. Berlin was obviously GTC, but he's been throughout sort of the continent. And his message seems to be like: reach your potential, right? Do you get a sense from the companies you cover here, from the startups, that they're reaching their potential or are there bottlenecks, which has been a big theme today as well?
Well, I think we are really on it and we're so willing to be part of the global market and stage. And I'm also working closely with wonderful companies here from Germany. For example, Neurabotics is also collaborating. They just built up their Neuragym also in Munich. So they also start to train robots. I have great hope for Germany and I believe Jensen too. I once asked him in San Jose what he believes, like what is the difference between Europe and the US. And the US, of course, they are good in software because you need to ship it before it's ready. And this is not how we work, you know, this is why we're good in industries. We like to build everything step by step by step by step and then it's very safe, but still we're not making a big show out of it. So I believe that with the two of us, we can be good partners because the US has great software, we too. We also start to build great AI models from Black Forest Labs, for example, Mistral AI. But then we always need the industry knowledge. So it doesn't make sense to just have the models, just have the AI. It needs to go hand in hand.
And so there was a whole section of his keynote devoted to physical AI where those things can go together. What was your takeaway? Progress on that front? Challenges?
So I believe that physical AI, as we're coming from great industry knowledge, also from the automotive industry, can be a great chance for Germany. And we're on it to build. I believe that with physical AI, it's all about training those robots. Because for example, you know, if I grab the microphone right now and I can feel it, I can hear it when it's falling down and I can — you can sense everything in it naturally, but those robots need to learn it. And so you need both: you need artificial environments to train them at scale, but then also real data and real environments. And I think it's about that we need to understand that robots learn from us and they need to learn everything we do naturally now for the very first time. And it starts with very easy use cases, like take something from A to B. And then bring that into your homes, bring that into the factories. So I believe like from AI companies to AI agents, we are now at the era of physical AI everywhere around the world.
Yeah, that was certainly his message as well. Laura, thank you so much for sitting down.
Thank you to Laura Sophie Botor. Coming up, a closer look at the engineering behind the compute capacity rollout and what it takes to build and operate this infrastructure at scale. But first, across the show, we're going to be highlighting some of the AI companies building across Europe, starting now with Cambrian, based right here in Berlin.
We managed to design a new molecule and put it in the market in less than 24 months. That is absolutely insane. We are around my computer, we hit enter, we got a bunch of sequences of DNA, we ordered the DNA, put that into a strain, fermented that in a 15,000 L tank, got tons of product, sold that in 24 months using AI. This is the pace we need to impose on modern science, modern R&D. Oh, yeah, it's going to take 7 years to create a new molecule. No, it can take 24 months if you do it properly.
Most of the materials that we use on a day-to-day basis are unsustainable. You probably have like a polyester jacket somewhere, you're probably using plastic foam shoes for your running. At Cambrian, we use AI to design the chemistry of tomorrow that will yield better materials for the future of the planet.
You can see all the atoms flying around and towards the end, they look like a real protein.
Currently, I'm standing in our lab in the middle of Mitte, Berlin. We put everything up here and everything that works out, we send for scale up to finally roll out on hopefully the shelves of our customers later on. A lot of companies just focus on having AI make predictions. What we do is we verify, make predictions, create experiments, and learn from these experiments to improve the quality of the AI. And that is what allows us to design new proteins, new materials, new polymers in record time. And for that, Nebius was quite empowering because it reduced the friction between idea and execution.
If we go fast, then maybe we can actually launch foreign cars next year.
When you have a happy commercial team.
Just since we've started using Nebius, there's been a notable increase in performance, but also the time it takes to get to the answers that we want. So, we have these protein design campaigns where we are designing tens of thousands of proteins, and then we filter them to see which ones have the best traits. And to build this pipeline, you need several models, right? You need some sort of structural design model at the start, some sort of sequence design, a folding, and then you need the comparison of all of them. And what we did by employing the NIMs provided by Nvidia on Nebius' platform is we basically managed to speed up this process fourfold. So, what would usually take us two to three weeks, we can get done within a week. And they understood that I was scientist first and engineer second. They offered a lot of support, and this really, really made the process much easier.
We're in the part of the lab where we do next generation sequencing. So, this is where we can, once we have engineered our microbes, sequence their genome and understand what these edits that we're trying to introduce have actually achieved. Is the gene in the right location? Do we have any mutations? Is the strain what we think it is? And once you understand what features are important, you can start building them into your models to be able to get better results down the line.
We use the phrase, we're sort of repeating the work of evolution, but we're speeding it up.
We don't have time to deal with like EC2 scheduling and so on. 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.
It's not an easy field. It's not a simple task that we're trying to solve. I think there's a lot of like startup magic happening at Cambrian. We all have to wear several hats, and I think the general energy of like can-do attitude, let's just like solve every problem, is what it is to work at Cambrian.
Welcome back to Nebius Live. We have talked a lot about the scale of investment going into AI infrastructure, but as the industry races to add capacity, the engineering challenges, they're becoming just as important as access to capital and power and GPUs. AI data centers, they're fundamentally different from traditional facilities. We're talking about increasingly dense computing systems, enormous power and cooling requirements, networks that are connecting tens of thousands of GPUs, and increasingly the economics of AI infrastructure depend on getting all of those systems to work together efficiently. To explore that, I'm joined by Jeremy Aliyahu Ontiveros, managing director at SemiAnalysis, and Andre Karalenko, chief infrastructure and product officer at Nebius. Gentlemen, welcome. Thanks for being here.
Andre, let me start with you because I think Mark brought this up at the very top of the livestream. Speed is of the utmost importance on so many different levels. So, what is Nebius's move to go quickly? What does that look like in practice?
Well, probably the most important part is that we as a team, as a company, already went through that curve. So, we know how the good result looks like, feels like, how good and bad execution feels. And so, it's a slightly different technology, but the same idea, the same priorities, the same set of people, basically. And we have this experience.
The next I would highlight is that building that basic infrastructure is the highest priority for the company. That's where a lot of our efforts and priorities just go. We're just all in on it. And probably the... Also, we are very open to different solutions. So there was a technology shift, there are some technology shifts coming on. So we do not just replicate and build whatever we used to do or what the market is used to build and to execute. So now we are very open to pretty different approaches to how we deliver the capacity.
So, if you see a different approach, you'll pivot or you'll remain agile to be able to move with it. You won't like...
Absolutely. These days you almost never replicate what you've done yesterday. It's like the...
Yes, every single project is much different from your previous one. Even if the gap between the start of those projects is like a quarter.
And you know, some of the stuff that Jensen Huang showed on stage, right? He showed the Vera Rubin. He showed like the actual hardware of it. And then he showed that massive container, right, Jeremy? You know, we hear news all the time of these huge new AI sites or data centers. We can't build them fast enough because we are supply constrained. What did you think about the keynote, right? He talked about DSX as kind of a solution, almost like containerized data centers that you can drop in that reduces that time to build and get online, I guess.
Yeah, look, there's different things. DSX, I would say, is extremely helpful for the whole ecosystem because what we've seen since the beginning of the AI wave is just a huge wave of new entrants, right? New data center developers, new neo clouds. And obviously having this sort of centralized blueprint produced by Nvidia which has the full system view on like what needs to be done, the cooling at the power level and so on and so forth. That helps everyone building AI infrastructure. So, I think it's a great strategic move by Nvidia to make sure that everyone is just overall better at building AI infra. Obviously you have companies that like Nebius start much bigger, have a lot of engineers, can take on more than themselves. And there's other companies that are newer to the space, still have certain assets like maybe they have a great site with power and so on and so forth. Maybe they have a great technical team, access to capital, but maybe they lack the data center expertise and that's where DSX is in the market.
You would want to use DSX, but it's not for everyone, is that right, do you think?
I think it's a blueprint that shows how to do things, but there's still some amount of flexibility. It's not like DSX is this one thing. And for example, you talked about container data centers. That's one way to sort of build facilities, but there's obviously many other ways. These container data centers can be extremely fast to deploy. They can have some drawbacks as well. Some companies, especially for example hyperscalers, they want to build for fungibility, they want to be able to have, you know, GPUs, CPUs and maybe TPUs or Trainiums, all different kind of chips, and so they're going to have different types of designs that some of these container solutions may not be the best at the time.
And so the container solution, it makes, you know, the lead time a little bit shorter to get the data center online. I think it was something like four months down to just two weeks or something like that. But you know, this is something I asked Jensen as well: you can have the data centers there, but you need the power to actually power them. And I know that Nvidia is trying an early trial of Nvidia DSX Flex, which looked at that exact problem. Let AI clusters adjust their power use in response to the grid. How does that work and, you know, what has it enabled you guys to do in terms of speed?
Well again, I'll refer to the history. Back in the days, like a couple of decades ago, when the previous wave started, everyone was building for a certain money and the project took two or three years. Going forward like 15 years, the cost of the build-out was like 30% of the initial ones and you could do it in 12 months or eight months instead of two or three years. So that's kind of how the previous technology evolved. But that was a very gradual growth. Now with the AI stuff, you have to scale so much, you don't have that time to basically optimize everything yourself and to gain efficiencies. So the ecosystem itself — it includes Nvidia, us, our colleagues, vendors, and the design companies, everyone — they have to share this information, to share the approach, to get all these cues, to get all the production. Again, to get to the more optimal and more effective scale faster, much faster, because the investments are bigger, you need it faster.
Right. Everything has scaled from capital to the data centers.
So that is the biggest value, basically, of getting everyone together, saying like, 'This is the blueprint. This is how it can be done. This gets to everyone.' And then everyone can take it and just evolve on top of that.
Now, Jeremy, SemiAnalysis has sort of a big number that it's predicting for, you know, gigawatts coming online. What is it, 43?
added in one year in 2027.
Which is just a wild number.
Excluding China, by the way.
Excluding China. How is that possible? What do you think the biggest bottleneck is? What could stop us from reaching that?
There's plenty of bottlenecks, of course, across the whole stack. If you talk to people building data centers, they will tell you labor is a big struggle. Maybe workers aren't showing up. Maybe they're getting other offers elsewhere at much higher prices. Maybe the cost is going up continuously. Supply chains. All of these OEMs, you know, the Vertivs, the Schneiders of the world, they have to expand capacity tremendously. The capital has been a major issue as well. And capital has actually been, I would argue, probably one of the key reasons why '27 is going to be such an acceleration versus '26 or '25. It's because since the middle of 2025, roughly speaking, we've seen materially more capital being deployed in this space. One of the indicators we track is like lease contracts, you know? Microsoft signing with QTS for QTS to build a data center for them. These lease contracts — we track $1.3 trillion of lease contracts signed year-to-date in 2026, when 2024 was like 200 billion.
Right. You know, I'm wondering like where does this capital come from? We've already seen hyperscalers, many of the safest, best fortress cash balance sheets possible, dip into debt financing. How do we find all that capital for next year?
Well, I mean, if you look at, for example, Microsoft, if you assume like the debt for a data center lease is like, let's say, 6.5% for a Microsoft exposure. And then until a few weeks ago or a few months ago, the 10-year risk-free rate in the US was, you know, 4.5%. Then you get like a pretty nice spread for a triple-A rated company, which is higher than the US government, by the way. It was pretty attractive. Now, obviously, since the last few months, the risk-free rate has moved up. Spreads have moved out a bit, which is a consequence of, hey, maybe the money that threw all of this capital first, maybe that's starting to be a bit tapped out.
And so, you just need to increase a bit the interest rate to enable more folks to come in. However, when you talk to all of the data center builders in the space, they will still tell you that hyperscale off-take is actually fine. People are still very much willing to deploy capital for a hyperscale off-take, a 20-year contract, for example.
Right, right, right. With a 3% escalator on Microsoft. What is most struggling these days to finance is AI labs, for example. OpenAI and Anthropic signing 5-year contracts or, you know...
Think about all the — they also need to tap a lot of capital in public markets, probably, right?
And I think there was too much enthusiasm perhaps, where folks thought like, hey, I got a $20 billion contract from OpenAI. It's going to be more than enough to make sure I get funding at, you know, 90% loan-to-cost. I think that was kind of delusional to assume this could carry the build-out. They're not investment grade. OpenAI is not yet profitable. Anthropic is. And they have massive compute commitments. So, obviously, what's happening is interest rates are moving up. People are asking them to do maybe higher prepayments, maybe get some backstop from the chip makers. And what is interesting is the chip makers, they're making so much money off selling the chips that it actually makes sense for them to provide maybe, you know, a 5, 10 billion dollar credit support per gigawatt if they're going to sell 30 billion dollars per gigawatt.
Sort of just very quickly to you, what is the biggest bottleneck in your opinion to building data centers?
Like Jeremy said, all of that.
All of that. All of the above.
If you had to choose one.
It's not also the bottleneck. You have a moving target. So, every single day you want to do it fast and more and more. So, it's like when you're dealing with a moving target, you basically will always have almost everything bottleneck. If everything is bottleneck, then you're in a good position. It means you're not being strangled by something specific.
I actually think that sums it up. So, great answer. Thank you guys so much. Jeremy from SemiAnalysis and Andre from Nebius. Thanks, guys.
Still to come, we are turning to physical AI and the growing effort to bring AI out of the data center and into the real world. We will be joined by Nebius's head of physical AI and the co-founder and CTO of Saract. For now, here's another look at the AI companies building across Europe. Here's Rota AI.
There's a clear gap between imagination of the technology world and reality. There have been decades of these cool demos of robots doing backflips or humanoids carrying items around, but at the same time, you know, no one has a humanoid in their house. No one is running humanoids in their factories. No one had an AI model that was good enough to where the robot can operate in the real world.
One of the things that Nebius, apart from...
other cloud providers is actually how much they care about our success. From the onset of the company, we identified that the biggest problem in robotics is a data problem.
The way that robot models are conventionally trained is, it's actually kind of surprising. They're typically trained by someone standing behind a robot and teleoperating it. So, essentially puppeteering the robot and doing that for thousands and thousands of hours. And then the robot learns from that data and kind of imitates what it learned from that data. Problem with that is that as soon as you then take that robot and put it in a new environment, it no longer has the data to know how to operate in that environment. And so, you end up with these very brittle policies that just practically don't work in the real world.
And humans don't need this kind of data and they learn very accurately how to manipulate the world. We felt like there should be a way to be able to do this if we can construct our models in the correct way.
We view data sort of a pyramid where the width of the pyramid describes how available that data is and how much of that data there is. The height of the pyramid describes how close it is to the direct data that you would collect from a robot.
And our team's insight was that you're never going to solve a problem as challenging as robotics by trying to collect your own data. There's only one data set that even comes close to that and that's internet-scale video. We can extract the intelligence about the physical world from video data and we can use that intelligence to then teach robots to do tasks that were previously believed to be impossible to automate.
Nvidia's helped us double our compute capacity at a really critical inflection point for the company. To train these types of models, you need a lot of compute, but it's not just quantity of compute. The first is the reliability of nodes. The second thing is how they're interconnected and something that stood out to us about Nvidia's was when we ran tests, we found that the reliability of the chips was actually very high.
We leverage managed Kubernetes Slurm for being able to access these resources, utilize them for training our models.
Resulted in a cluster that we can train very large models on with pretty high levels of reliability.
And really enable us to kind of not worry about the compute infrastructure and worry about ultimate performance on our robots.
What was really great about our partnership with Nebius is that, even though they didn't necessarily have in their road map plans to put petabytes of data right next to our compute cluster, they actually did the work to build that storage on the cluster for us so that we could then use it to train our models.
One of the things that to me has set Nebius apart from other cloud providers is actually how much they care about our success. Nebius has the strong desire to maintain it so that we are successful.
Every time we work with a real customer is really fun because you learn about what it actually takes to deploy a robot.
We want the robots we're developing here to ship to real customers and have a real impact.
AI is really going to be able to enable humans to focus on the things that they derive value from.
My real hope and something that we care deeply about at Rodo is that robots are used to enhance people's lives.
Physical AI is one of the big themes here at GTC this week and we are starting to see the technology move beyond research labs and demonstrations into real-world applications, particularly in robotics. Joining me to discuss this next phase are Mark Tishler, co-founder and CTO of Seriac, and Evan Halda, head of physical AI at Nebius. Welcome to you both. We were just having this conversation. We got to the bottom of the decks at the end of the keynote. They were the hugging face stocks. Give us a little context behind that. Like what do they do and why did Jensen bring them out?
So, I think it's just like some fun robots everyone can use in their house.
Yes, exactly. I also ordered a couple of them.
And what do they do? Just
Whatever you want. Haven't arrived yet, but yeah, they're walking around and clean house.
Cute. They're cute and they're Thomas.
I think it's a way to try to just inspire people and get builders hands-on keyboard. It's an open-source platform, right? That they're building to help train robots. So, it's a way to educate people on what's possible.
Right. And so much of the robotics race has been these demos that are really amazing and capture imagination. But, Mark, give me like the boring version of this. What happens when you get robots to do the same job all day?
I think when it comes from demos to deployment, I think that the boring version of it is that in deployment, robots have to be super reliable, super accurate, and quite fast. So, and also like the customer often doesn't really care how we make it happen. So, we love our new world action models, vision language action models, but the customer in the end doesn't care. They care about KPIs. I think that's the boring version of it. And at the same time, we have to make sure because with physical AI these days, what the models brought us is capabilities that were unimaginable 5 years ago.
For me as a roboticist, I'm doing robotics for the last 12 years. And 10 years ago, I couldn't have imagined that we will end up here.
What about you, Evan? Are you surprised by some of the breakthroughs that we have, you know, robots and industrials, humanoids? Talk a little bit about what it takes to actually put those out there too on the Nebula side.
Yeah, I think people are being surprised every week. I mean, just recently you had the Astra model from Open AI come out that was in everyone's mind an LLM. And all of a sudden it has like VLA world model type capabilities. And I think there's these emergent properties that are coming from the bitter lesson, right? More data, more compute, and amazing things are happening. And so, yeah, I think we're seeing models that can reason across tasks, models that can reason across form factors, and it makes, if I'm building a robot, it begs a lot of questions about where value is going to be captured and if I should train my own model, if I should just use off-the-shelf models, if I should bet on getting the hardware right. And so a lot of open questions if you're someone like Seer React, but for us, we're here to try to help really accelerate every stage of that learning loop across all kinds of different.
And I feel like, right, and so for a robot to work when you talk about safety and all these things, you have to have the right models behind it. You need the compute power behind that. How computer intensive is it to run these robots?
I think it's quite computer intensive. Like first of all, training a new task, training a new use case always requires training.
Give me a few of those use cases.
Like we currently do return handling in warehouses. Meaning we process returns, goods people sent back.
Yeah, it's our robots doing it. You have to unpack it, unfold it, inspect it. Like an AI model inspects the thing and then decides whether we can restock it or we have to send it to a checking or quality assurance unit. So and then we fold it again, put it into a bag, and then we send it.
Oh, so it's like Dexterous. It's actually doing the folding as well.
Like first opening and then unfolding, inspecting it, then folding it again, and then putting it into the bag.
Are these mechanical hands then? It's not a full humanoid. It's kind of, it doesn't look like fingers, too.
It has finger grippers, like two finger grippers. And it's a stationary dual arm system.
Right. What's something that's easy for a person but still surprisingly hard for these robots?
I think actually the whole task is quite easy for a human. You can learn pretty quickly.
But like it has so much variety that, you know, a t-shirt could be folded differently. Someone could have packed, you know,
There's many many ways to fold it.
many. Yeah, and someone could have packed something totally different. And a human would recognize easily. And it's quite hard for a Roman robot to manipulate the things.
On the other hand, recognizing stains, recognizing damage, and so on is much faster and much more accurate than a human could.
Right. So, there's sort of like give and take a little bit of that. Evan, what are robotics teams having to build themselves that should be easier by now?
Everything. [laughter] A lot of them have really talented technologists and builders. So, you see a lot of them go out there and they have a mindset, they're confident they can build a lot of the back-end elements themselves. What we do see though is a lot of them end up spending a lot of time twirling with infrastructure and in the cloud when a lot of that time could be spent on the actual learning part and the actual setting up the experiments and doing the research. And so, that's our big focus is how do we abstract away a lot of that infrastructure orchestration work, the pipeline, staging, and we're finding ways now to actually build out frameworks that allow agents to do that for them. And so, we can get into that if you want to. We're launching things we're launching here.
Yeah, sure. Well, Mark mentioned the KPI for the end user and, you know, the main KPI we're seeing, tell me if this is wrong, but I think the main KPI for you is iteration speed. Like how many times can you get through a learning loop and into that next training run based upon the real data you're collecting and the failure modes you're seeing. Is that...
I think that's a KPI we are tracking, yes. If we can increase that speed, like we will be fast.
I'm applying that to your example, it's like it took X amount of hours to get the robot to fold the clothes a certain way, a shirt a certain way, to fold a pair of pants, that it should take less iterations, should learn that new task faster. Is that how that works?
So, the iteration speed we're talking about is upstream of that where there's numerous steps to go into a training run or a learning loop is what you'll hear about, right? So, they have a robot in the wild. It's collecting real-world data. There's now all these steps to go and take that data, curate it, process it, and find the failure modes, and then get it ready for training. And then there's the next stage of training.
Then you got to take that new model that came out, and you got to do eval, evaluations. Does it do what you think it's going to do? You can put it in sim, simulation, you can put it in the real world. That's its own pipeline and workflow. Then you got to deploy it, and you got to do it all over again. And what we're seeing is that there's all these different tools at each of those stages.
And those should be a more connected workflow, a more connected pipeline, a little more automated pipeline, but today it's not, right? There's a lot of glue code you have to write between tools, the tools don't talk to each other. There's different cloud environments sometimes between them. There's the edge environment. And so the question is, how do we create more of a single pane of glass where we can orchestrate those different tools and cloud environments as more one uniform pipeline? And fortunately, we now have agents that are really good at a lot of the engineering work that does the glue code, that does the infrastructure provisioning. So what we've launched here is what we call the Physical AI Workbench. It's part control plane for Physical AI, meaning it's a black pane of glass to go in and to procure infrastructure. It's also part agentic harness.
Right? And you think about what a harness is, it gives agents the ability to be effective in the world. To do that, you need access to infrastructure. You need shared context across the tools and stages. And you need access to tools.
So what we're doing is we're taking all that packaging into a framework and allowing customers to point an agent at it and use language to say, "Hey, build me a workflow with these steps, these tools, these GPUs, run it for this amount of time." And boom, it just goes. And so it goes from what would take days if not weeks to build out one full pipeline down to hours. And so we're working with Seer React on this, and they actually told us that it's been saving 40% of engineering time just to build out workflows and pipelines. So that's a pretty huge leap, and that's what creates iteration speed.
And that's a lot on the back end. I mean, now I understand what you said. Everything is not the sexy... I'm sorry. It's everything I was thinking just full like it to get to that point of a robot folding a shirt. Let me ask you guys maybe a little bit of a broader question that fits into, you know, Jensen Huang's tour through Europe and some of the stuff he's been saying. Like, we're sitting here in Germany with all of this industrial expertise around us. Is there anything missing to turn that into even more successful robotics companies or do you think that, you know, Germany is already on its way?
So, I think there could be a lot more robotics companies if they would take a bit more risk because I feel like automation and robotics will change massively because before we had to write a lot of code for systems that were basically state machines. Whenever you had an error, you had to write some code that fixes that error. Now, one model call has replaced all of that. And I think that was quite impressive. And that was the change. And so you could potentially automate everything.
And I think if we would use the industrial network we have in Germany or in the DACH region, I think many, especially industrial facilities, have a lot of automation already, which facilitates an easier entry. If you don't have to walk around and get stuff out of shelves, but you can get it from a conveyor, that's already an easier entry for a new kind of robotic system.
So, is the idea you need to embrace that more? Evan, what do you think? Same question to you.
