Jensen Huang Exclusive: AI Financing Fears & Rising GPU Rents — GTC Berlin

[00:00:00] Speaker 1: Hi, everyone. I'm Deirdre Bosa, live in Berlin, and I am pleased to be joined by NVIDIA CEO, Jensen Huang, fresh off the stage here at GTC Berlin. Thank you for coming right to me. I appreciate that. [00:00:16] Speaker 2: Delighted to be here. [00:00:17] Speaker 1: How's the European orbit? Welcome to Berlin. Thank you. [00:00:19] Speaker 2: Incredible. Incredible journey. We started in, I guess we started in France or London. London to France to Sweden to here. And in the last year, we've invested about $10 billion in companies here, and this week we announced 14 more. So we're in the process of building the AI economy here. [00:00:40] Speaker 1: Right. And some of your message has been just start building, start going, overthinking, stop overthinking, start building. Do you think that message got through? [00:00:47] Speaker 2: I hope so. And, you know, Chancellor Mertz also told everybody to lean in. Let's go build this thing. You know, I think the fact of the matter is, is the technology had to be invented for the last 15 years when we're kind of in invention mode. But the six months ago, you know, we really transitioned to AI being useful mode. And the thing that I wanted to do during keynote was to demystify what makes AI useful, what are the components of AI, and why they could build it, you know, incredibly here. [00:01:18] Speaker 1: Right. And you talked a lot about that on stage just moments ago. I want to talk about one of your announcements. Yeah. DSX Pod, an AI factory in a shipping container. Is that a fair sort of representation of it? You know, it's fascinating because it gets at the supply chain, but it also gets at this financing question that's been sort of rattling the AI trade over the last few weeks, especially the neoclads. So if you can give lenders essentially more certainty about what a data center costs and when it starts earning money, when it comes online, do you think that makes it easier to finance? [00:01:48] Speaker 2: This is going to make it way easier. Let me tell you why. So a gigawatt of data center is like $50 billion. And that $50 billion, from the moment that we ship the technology components, it goes through an OEM, those computers land at a data center, from that point to operationalizing, could be four to six months. And during that four to six months, you're wiring these computers up, you're getting them all booted. And so what if we were to do some of the manufacturing of that in the supply chain? The advantage is that you might be able to shorten that four to six months to maybe three to four months or take three to four months out of it. And as a result of that, your time to revenues is going to be accelerated, but also that idle cash, because you had to pay for it, that idle cash would be reduced. [00:02:40] Speaker 1: Right. So it's sort of narrowing the cone of uncertainty, right? And that has been the major concern. Right now, markets are sort of, they're volatile, we'll say, worried about neoclads, then optimistic. What do you think the chances of a shakeout, any kind of shakeout are? [00:02:53] Speaker 2: I think this is going to get solved. And the reason for that is because the demand is so good. You know, just last night, SpaceX just announced, and I think Elon said that demand is many times the capacity he has. And he has a lot of capacity. And so he's building more because the amount of demand is just incredible. CoreWeave is kind of the same way. Nebius is the same way. Nscale is the same way. I think the challenge, of course, is that, and this is kind of what you were trying to get at, from the moment that the computers are shipped to the moment that they start earning could be many months. And the amount of float, because we're talking about a fairly significant investment, the amount of float is quite large. And so we need to help these companies reduce that. And I think reducing the supply chain, making it easier for them to stand up the computing is going to make a big difference. [00:03:42] Speaker 1: Right. So you're doing this on the NVIDIA side. But what about some of the physical bottlenecks like power? How much of a timeline can you really take out of the equation if the customer, you know, is waiting years for grid connection? [00:03:53] Speaker 2: This is really one of the reasons why we're so resilient to this, because we have so many different customers. You know, we're the only computing company that supports every single AI model, every single CSP, every single computer company, across neoclouds in every single country. And so there's some, there might be some bottlenecks somewhere, but because the overall demand is so great, we'll be able to get the computing out to the market wherever they need it as soon as possible. [00:04:22] Speaker 1: Right. And what is your outlook on the power question on some of those physical bottlenecks? Do you think that gets solved in the next few years? What, you know, there's a lot of efficiencies on the chip side on what you're doing with the DSX pods, but what is sort of the bigger picture there? And do you think that's the biggest bottleneck? [00:04:38] Speaker 2: It is one of the major bottlenecks. The bottlenecks keeps shifting. And the reason it started with wafers and the co-was to HBM memories to silicon photonics to even connectors. And as those bottlenecks are, they're going to continue for probably a couple more years. Now that we're ramping up these AI factories around the world, now power becomes increasingly so. But, you know, oftentimes the power is available, but it's the land power and shell is either maybe a construction limitation or maybe the power generator is on being delivered or connection to the grid or, you know, whatever, you know, a lot of construction challenges. Because the number of jobs that this has created for construction workers around the world is incredible. And so, you know, there's limitations there. And so, anyhow, I think the bottlenecks just, these are physical bottlenecks and they're part of the physical world. [00:05:36] Speaker 1: Are you sort of getting at some of the permitting bottlenecks as well, getting these online? [00:05:41] Speaker 2: Some countries have permitting issues, but some companies, some countries don't. Some states do, but some states don't. And so, I think the NVIDIA diversity really makes us quite resilient to this. You know, we'll be able to ship it somewhere. And whoever has a bottleneck, we'll work with them to make sure that as soon as they get the bottleneck resolved, we'll ship computers to them as fast as possible. Meanwhile, we'll ship the computers to somebody who can use it. [00:06:07] Speaker 1: Right. And you're standing ready. I want to talk to you a little bit about compute as a commodity. It's starting to trade like one. There's GPU futures on the CME. Nebius has moved to spot pricing. They can change every 15 minutes, still early days. What does it mean for NVIDIA when computing power gets traded like capacity? [00:06:25] Speaker 2: Well, this is one of our advantages because NVIDIA's computing systems are so universally used and the performance is so good that it's productive for so long. And because we're software programmable and we keep enhancing the software throughout the entire life of the asset, we keep refreshing it so that it's productive for a very long time. And so people are discovering that what people thought used to be a four year life cycle really is probably eight to maybe more and irrespective of the clients, the customers that use the compute, they're always companies that can use the compute. And so there's just so many off takers. So it's productive, it's fungible, and it's durable. Now, when that happens, you might be able to underwrite a loan against the asset itself. Today, most computers, you know, don't have value, don't have very little residual value, but people are now recognizing NVIDIA's compute has a great deal of residual value. And so you might be able to use that and, you know, backstop the debt itself. [00:07:31] Speaker 1: Are some of the financial instruments and derivatives around the AI space, does anything about that worry you? [00:07:38] Speaker 2: Well, I think that one of the pillars for this growth, because we're talking about infrastructure build out, the capital market will have to think about how to support it. And now there'll be different capital instruments. But I think the real advantage that we have is because of the durability, because of the productivity and fungibility of our computing system, the risk involved in underwriting NVIDIA's asset in our NVIDIA computing is the lowest of any asset out there. And so I think this is an area that's going to really play into our advantage. [00:08:12] Speaker 1: You mentioned the durability, one of the most surprising things of the last few years has actually been the durability, the lifetime of chips, even older generations. There's some really great data from silicon data that shows, you know, Blackwell rental prices are still rising. I think you mentioned durability, did you say seven, eight years a moment ago? Is that what you think the durability is for Blackwell? Probably. What do you think that is for Verirubin, the next generation? [00:08:40] Speaker 2: Well, I think that Verirubin is probably going to be the same. It'll be, you know, seven, eight years. And, of course, the technology is changing. So long as we continue to upgrade the software, this is where NVIDIA's CUDA has such a big advantage. We're introducing new algorithms, new software layers all the time. And when we invest in enhancing the install base, you'll be able to continue to use it for a long time. [00:09:07] Speaker 1: Let me bring in Gavin Baker. He's waiting in the wings. He was my GTC pregame host. [00:09:12] Speaker 2: Is that right? Apparently, you guys did pretty amazing. Superstars. [00:09:16] Speaker 1: We had a lot of fun. We talked about safety. We talked about open source models. One topic that we didn't get to a huge amount, which I've sort of been talking to you about now is sort of... [00:09:26] Speaker 2: Gavin. How are you, my friend? Good to see you, brother. [00:09:29] Speaker 1: You know, this is why we do this live, because we can have surprises here, right? So, Gavin, thank you for joining us. You know, I've been talking to Jensen a little bit about some of the jitters in the markets, which, you know, always come around. I wonder how seriously do you take these ones now, and are we sort of out of it? Comments from Elon Musk, as you mentioned on the earnings call, we've got Jensen here right now saying that demand is through the roof. Is there still that financing risk that's being overestimated or underestimated? [00:09:59] Speaker 3: When I think about financing risk, I read something incredible last night. An open source model, DeepSeq, said that they are monetizing at $38 per GPU hour. Last time I checked, like, I haven't heard anybody renting GPUs anywhere near that high. So maybe this explains Blackwell prices up 17 days in a row. [00:10:22] Speaker 2: Yeah. [00:10:23] Speaker 3: How do you think about the intersection, what is clearly extraordinary demand, with supply from a capital perspective? [00:10:34] Speaker 2: Well, you know, the single problem that nothing can solve is demand. The fact that there's so much demand across the world, I mean, this isn't just from one model. We're talking about OpenAI's got incredible demand, Anthropocene's got incredible demand, Metamuse's got incredible demand, you've got GrokBot, incredible demand, Gemini 4 just coming out. You just mentioned DeepSeq, Moonshot's models have got great demand, GLM's models have got great demand. And so you've got all these different models, closed models, open models, the demand is just extraordinary. And we're just, we're going into this, we're really actually going into the inflection point of useful AI as we speak. And people are, more people are using it, they're using it to do more useful things, and that's the reason why token demand is so high. I think all of the other stuff is, they're going to get resolved. You know, without demand, nothing is resolvable. Absolutely. Yeah. With demand, everything will be resolvable. [00:11:32] Speaker 1: Demand can be incredibly high and usefulness can be going up. But if the financing is ill-timed, and you've heard people like Daru Amode, Satya Nadella, others talk about this, that can sort of trump everything else. How much are you thinking about sort of the macro environment? [00:11:50] Speaker 2: We have to make it easier for people to finance these build-outs. And one of the areas that we have a real advantage in is what you and I were just talking about, the fungibility and the durability of the asset. You could use the asset to underwrite the loan. And so, the fact of the matter is, you know, the residual value of NVIDIA's compute is really, really high. And at the moment, nobody gives us credit for that. [00:12:19] Speaker 3: I do think we should, you guys touched on it a little bit earlier, but the DSX pod, I actually think it's pretty important here. Like, if we think, let's use a trillion dollars as a nice round number of AI CapEx, seems like we're going to go much higher very soon. But a trillion dollars, if it's a six to nine-month cycle time, that's a lot of wasted capital. That's right. It's unproductive. So, maybe talk about the DSX pod and how that changes. The moment it gets shipped, somebody has to pay for it. [00:12:47] Speaker 2: But the moment, then you have to install these AI factories, and they're gigantic. And putting these AI factories in the service takes months. If we could reduce that months, you know, from the moment that the systems are shipped to the moment that it's deployed and being operational, if we can reduce it from months to weeks or even a month, that will be completely game-changing for all of these AI factories. Right. Having a trillion dollars, having a few hundred billion dollars sitting idle while you're waiting for revenues is obviously a challenge. [00:13:17] Speaker 1: And can give your customers a little bit more control over that sort of very pressing, very important financing question. Gavin mentioned open models. We've had sort of a year where it's just exploded in terms of adoption in the U.S. It's been a lot of Chinese open models in the spotlight. But over recent weeks, we've got some really interesting big developments from Western companies. Reflection, Mistral, Meta, of course, there's NemoTron that you talked about in the keynote. How do you see the balance right now in open-weight AI? [00:13:47] Speaker 2: The world's going to use closed models and open models. In the last six months of this year, several hundred billion dollars, I think $600 billion worth of venture financing went into AI-native startups. 80% of those AI-native startups are using open models. And so you're able to take what is close to a frontier open model, post-train it with your expertise and your data, turn it into essentially a super AI that is good at your particular domain of use, and build a company around it. And so you're starting to see this all over the place, whether it's in legal services or entertainment or manufacturing and design. And all these different applications are now jumping on. And so I think the world is going to have both closed models and open models. You're just going to use the best model where it makes sense. [00:14:40] Speaker 1: And we heard from, I think, every single guest in the pregame that they're using a mixture, they're using a number of models. So far, it's been the Chinese ones that have seen a lot of adoption. Do Western models catch up? Do we see sort of them gain momentum given, you know, the releases of the last few weeks? [00:14:57] Speaker 2: Yeah, no doubt. I mean, NemoTron with post-training gives you near frontier. You know, when you're trying to build something that's domain-specific and application-specific, you don't need to have the frontier of general intelligence. You can just have something that's close to, something like reflection or TMLs is pretty great, NVIDIAs is great. And you take these models, you post-train it with your own data. When you do that, they become super intelligent at the application space that you want to target. And so this is, you know, I think the open models have reached the frontier level. And I think at this point, we're going to see a lot of super intelligence built. [00:15:38] Speaker 1: Where does that leave the frontier labs? [00:15:40] Speaker 2: They're still going to be off-the-shelf incredible. And I would recommend that everybody rent a frontier lab, closed AI model, whenever you can. And you build your own AI only when you must. And so for a company, it could be one of the companies here in Europe, Dassault, they have to build their own AI. Cadence has to build their own AI. Siemens has to build their own AI. They need sovereign AI. That's right. They need sovereign AI. So they need the ability to build their own custom proprietary super agents. And all of the technology, all the resources are now put together. Everybody should be able to do it. And so, you know, I would just continue to use off-the-shelf whenever you can and build your own super agent whenever you must. [00:16:25] Speaker 3: Well, it feels to me like what a lot of, at least our portfolio companies are doing, is they are combining their own specialized intelligence. They're generally taking, actually, an American open source model, post-training it, making it amazing at the tasks they use it for. And then you combine it behind a router with your favorite closed frontier model. That's right. So I think the future is hybrid. That's right. Behind a router. [00:16:48] Speaker 2: Yeah. Just as computing today is hybrid cloud, today in the future, AI will be hybrid AI. Yeah, absolutely. [00:16:54] Speaker 1: Yeah, I want to get back to sort of the China question as well. You know, just saw a recent report that DeepSeq and Huawei are working together on chip designs. And you have long said that restrictions would actually accelerate China's own industry. How do you read this moment? And do you think it's time to take China's chip ambitions more seriously in America? [00:17:15] Speaker 2: We should always take China seriously. If there's one country that could manufacture things at scale, whether it's manufacturing, of course, manufacturing, of course, ships at scale, manufacturing open models at scale, or even, very importantly, manufacturing AI researchers at scale. Nobody does it, you know, more seriously at a larger scale and better than China does. And so we've got to be, you know, we have to be, we need to recognize that we have very, very serious competition. But, you know, the United States, we're, you know, we've got a lot of things going for us. And our frontier labs are the world's best. Our technology, our chips, our computing systems, our infrastructure, all the world's best. And, you know, we're going to continue to make sure that we stay ahead. [00:18:02] Speaker 1: Right. We should compete on innovation. Yeah. Has that message been getting through to Washington, to lawmakers, especially as we approach sort of a midterm? [00:18:10] Speaker 2: Well, to some. I mean, you know, obviously, the whole point of politics is that they're different people with different thoughts and different beliefs. But I think that at the moment, I would say that the U.S. government would encourage us to go compete around the world. There's a reason why we're in Africa. There's a reason why we're in Central Asia. We would like to compete in Brazil. We'd like to compete in Egypt and different parts of Africa. We want to compete everywhere. And so I think the policies, the environment makes it possible for United States, American companies to be the infrastructure for European countries, the global south, so on and so forth. And I think we ought to go compete for it. [00:19:00] Speaker 1: One more question sort of in how this time, how AI is perceived in the West versus the East and in the U.S. You know, there's a lot of pushback on AI when it's mentioned in a graduation speech. Of course, there's, you know, the data center permitting pushback, et cetera. China is a very different story. And, you know, they're embracing it, going back to useful AI. They find it useful. Why do you think it's developed that way? [00:19:25] Speaker 2: Well, I think the narratives around AI, of course, safety is important. And we have to talk about safety. And the most important thing about safety is to build the technologies so that we could deploy, we could build these AIs and deploy the AI safely. And so OpenShell is about that, Sentry is about that. We've dedicated an enormous amount of R&D so that we could create these safety systems that the world could deploy safely. And so obviously safety is important. However, we also know that this is really incredibly powerful technology. And we need to make sure that we direct it in an optimistic way so that its potential and its positive impact could be enjoyed by society. And we ought to talk about that more. I think some of our narratives about jobs obviously scare people. The facts are that AI has created more software jobs. The facts are that AI has created more healthcare jobs. The facts are even AI is creating legal jobs. If you look across the entire five-layer cake that I talk about, the energy sector is booming. All the companies in the energy sector are booming. This is the best chance to invest in sustainable energy at no better time than, you know, now in the last hundred years. And so whether you're a nuclear fission or fusion or large or small nuclear reactors, it's all being invested in. Battery technology, solar technology, it's all being invested in. So I think the energy sector is doing great. Obviously, the chip sector is doing great. The infrastructure sector is doing great. Construction workers, we're re-industrializing the United States. We're going to re-industrialize a lot of Europe. And so those manufacturing jobs are doing great. And so each one of these layers are creating new jobs. So the narrative about jobs is almost precisely backwards. It's precisely wrong. And so I think that the facts are it creates jobs. And I think that people are leaning into it. [00:21:26] Speaker 1: I mean, there's a narrative problem, but there's also the way that sort of tech has developed. And there's certain things. Microsoft, for example, took away the NDA, nondisclosure agreement, providing more transparency to these communities. Are there things that can be done, not just said, that would help sort of the reputation and the blowback? [00:21:45] Speaker 2: You bring up a really good point about these AI data centers and AI factories. You know, when you go into a community and you want to build into a local community, before you start building, you ought to extend yourself out to the local community and partner with them. Let them know that what you're about to do is going to bring, of course, all the benefits. But it's going to, of course, bring new neighbors. And we ought to be good neighbors. And I think that the first couple of years in building out the AI factories and data centers around the world, people were a bit too much of a hurry. Everybody has now taken a step back, whether it's the NDAs or contributing to the communities, whether it's building shops and centers and community centers and improving schools. And in many cases, even supporting lower taxes and, you know, all kinds of benefits that can be brought to the local communities. I think those are all good ideas. And each one of them ought to be tailored for the needs of the community. But the most important thing is you ought to be a good visitor. You know, you should be a good guest. And before you go into a community, stop by and see if you could be helpful. [00:22:59] Speaker 3: AI is clearly creating a lot of jobs. It's doing amazing things for the world. We're re-industrializing America and Europe, helping to make America and Europe greener. But to ordinary Americans and Europeans, like how can AI make your life better today? What are a couple of tangible examples? Yeah. [00:23:16] Speaker 2: In a lot of ways, almost everybody's using AI already. And maybe people are looking for that extraordinary transformative face shift. And all of a sudden, everything is different. I think everything is just going to be better, as in technology over years. For example, our search has gotten better. Our ability to use computers to write, to design, to create, to solve problems have gotten better. Obviously, our healthcare, the ability to diagnose disease has gotten better. The number of examples of discovering, you know, early detection of colon cancer or breast cancer, so many different examples, so many stories of AI helping diagnostics, that's gotten better. Obviously, we're going to make cars safer. Without AI, that's not possible. We're going to make planes more energy efficient, fuel efficient, not possible otherwise. And so, designs of new wings and new engines. But I think that the simplest idea is that AI is a tool that's going to amplify your capabilities. And I don't know anybody who wouldn't want to have this incredibly useful tool to amplify their own abilities. If you were a small shop somewhere, and you know a lot of things about the shop, the business you're in, but you don't know very much about software programming, and you would like to create a website. And this website, you know, is going to have this incredible digital marketing expert inside. You could now get an AI to write that for you. So, every single small, medium business, whether it's one of the medium industrial companies, Mittelstand, here in Germany, if you don't have software capability, now you have software capability. And so, everybody has all of this available to them. And so, I think, yeah. [00:25:22] Speaker 1: It amplifies capability, but I think what many Americans are worried about is it amplifies risk as well. And it's always that trade-off on even like a micro level. You can see that with the rise of personal agents, right? Like Muse, GrokBot, Instinct, and some others. [00:25:37] Speaker 2: That's what OpenShell is created for. [00:25:38] Speaker 1: Exactly. Yeah. Do you think that ultimately, you know, defenses can keep up with that risk? We talked about this in one of our panels. But, like, when does that happen, and what needs to happen for people to actually feel safe? [00:25:51] Speaker 2: We should apply the philosophy that is captured in OpenShell instantly. And basically, it works like this. AI is an autonomous, these agents are autonomous systems. And when you have an autonomous system on your computer, autonomous agent on your computer, you should first start with zero trust. Right. And give it minimum privileges, minimum access, minimum use of tools. Maybe they could access a file, but they can't delete the file. Maybe they can access an internet, but they can't, you know, send stuff out. And so there's a lot of different things that you could specify. We just need those practices about how to deploy agents and deploy autonomous systems has to be adopted across the board right away. [00:26:38] Speaker 1: And I think I've heard you use the analogy of, like, if you have a one-year-old, you put them in a playpen, and you don't give them a hammer or anything that's going to hurt them. But the example that many people are now familiar with are agents or swarm of agents escaping this playpen or escaping that sandbox. I wonder how much damage did that do to, like, the narrative to trust? And was it a one-off, or do you expect more of these things to happen? [00:27:02] Speaker 2: It shouldn't have happened, and it didn't have to. Yeah, it shouldn't have happened. It doesn't need to happen. If OpenShell was used, it probably wouldn't have happened. And Hugging Face developed and exercised a simulator, and they came to the conclusion if OpenShell existed, their website wouldn't have been attacked that way. And so I think the technology now exists. The methodology of delivering software in a zero-trust environment, that methodology has existed for some time. We ought to take advantage of that. And so I think we ought to build on existing technology. We ought to create some new technology. The way to think about safety is we have to accelerate AI safety. And safety technology and capability, to me, is one and the same. You know, how could you have a car that's not safe? And so as much as it is capable, you want it to be safe, the two of them are really one and the same. And so in the future, as we develop these AI capabilities, safety will come along with it. [00:28:08] Speaker 1: And this is very much your philosophy, NVIDIA's philosophy. Do you trust the rest of the industry to have that same focus, focus on capability, but safety along the way? [00:28:18] Speaker 2: Yeah, I think so. I mean, look, OpenShell was announced. We have over 100 companies supporting it. I highlighted one particular collaboration, which is what I'm very happy about, which is Anthropic, using OpenShell, working with NVIDIA to create cloud-managed agents. Really terrific work. And we're working with all the labs and working with all the operating system companies. I think the incident, unfortunate. But I think that it's very, very much transformative for the industry. [00:28:51] Speaker 3: And we have an existing system of laws and liabilities that incentivize companies to use the best playpins available. And the playpen just got a lot better in the last month. And it just, history has not been kind to unsafe companies. No. Safe and fast. [00:29:08] Speaker 2: No customers. And many of these customers are businesses, other businesses. And so these other companies can't put themselves at risk. [00:29:16] Speaker 1: So you're saying there's like a market reason to be safe and those that are not will drop out, which I think is fair. Jensen, thank you so much. We really appreciate you coming right off stage here and joining us live. Gavin, thank you for stopping by. [00:29:30] Speaker 3: Thank you. Thank you for having me. Both of you. [00:29:32] Speaker 2: Good times. Yeah, fantastic. Thank you. Thank you.