Hey everyone, it's Thursday, March 19th. Welcome to another CNBC live stream. I'm Deirdre Bosa. Well, GTC just wrapped and everyone is talking about new chips, Vera Rubin, the Grok integration, usual GPU horsepower. But I think the most important thing Jensen Huang did this week actually had nothing to do with silicon. He launched Nemotron Claw, jumping on that open claw frenzy bandwagon. It's a free open source platform for AI agents. That tells you where Jensen Huang is looking next. Now that well-used Gretzky saying applies here. Skate to where the puck is going, not where it's been. So while every other CEO in semis is focused on building a better chip, Jensen Huang, he's also looking to the next layer. Of course, he's building the next great chip as well. He's looking to an operating system for agents. Let me explain. Nvidia dominated the training era of AI through lock-in. Its software ecosystem made it nearly impossible to switch to a competitor. That's CUDA. But the industry, it's now moving from training models to running them. And when you're running models, that lock-in, it doesn't hold the same way. Google is building its own chips. Amazon building its own chips. Broadcom is designing custom silicon for the hyperscalers. So that moat that made Nvidia the most valuable company in the world, that is getting thinner. So what does Jensen Huang do? Well, he builds a new moat at the platform level. Nemotron Claw, it's free and it runs on competitors' chips too. That sounds like a crazy reversal for a company that became dominant thanks to that lock-in until you think about past technology shifts. Google, right? It gave away Android because the product wasn't actually the operating system. It was search. Same idea here. Jensen's product isn't Nemotron Claw, the compute and GPUs underneath it. Now, there's another layer to this. If Nemotron Claw becomes how companies actually deploy their AI agents, it commoditizes everyone above it. So, makes it harder for OpenAI or Anthropic to get big enough to squeeze Nvidia on pricing or build their own chips. Now, Jensen isn't just building a platform here. He's making sure that no single customer ever gets powerful enough to walk away or even just get the upper hand. So, that's what we're going to dig into today. I have the perfect guest for this conversation, Clement Delangue, CEO of Hugging Face, which is essentially the home of open source, where developers go to find, download, share AI models. Clem, welcome. It's great to be chatting with you again today. And congratulations. You're coming off paternity leave with not one but two babies. You had baby girls, twins.
Oh, thank you. That's so nice. I mean, my wife has to be congratulated. She did all the work. I was just here to support. But yeah, as you said, super exciting week, super exciting announcements at GTC. As you said, I think Nvidia recognized that they need everyone to be able to build AI, not just a few players. And so that's why they're supporting open source AI so much. That's why they became, in my opinion, the American king of open source AI. And that makes it super exciting to see what they're going to do next.
I really love that framing. I know you tweeted that, too, and I retweeted you. You're officially calling it. Nvidia is the new American king of open source. Who's he dethroning? Would that be Mark Zuckerberg, Meta?
Well, I mean, they're dethroning, and they're competing first with Chinese players, right? Because the reality is that so far the ones dominating open source AI are Chinese companies. We released a few days ago a state of open source AI that shows that for the first time in 2025 the volume of downloads of open source models from Chinese providers was higher than from American providers. Right? Even in the US now the majority of open source model downloads in the US are from Chinese providers. And so I think what Nvidia is doing, which is really great, is kind of like creating an American competitor to that and providing American models to Americans but also to the world.
So back up a second. Why do we need an American champion? I've been looking at the same data with you and you know, as you have, and you know, DeepSeek, a lot of folks in America want to say that was sort of like a moment, a glitch, and it's been reversed. But the data doesn't show that, right? The adoption of open source models is only growing. So explain to the people who are watching why we need someone in America to be looking at this.
Open source and open source models are really kind of like the foundation to build all technology, right? Everything is built, like the chatbot that you're using every day is built on these open source foundations, right? And so I think what we need is every single country in the world to be able to build their own models so that everyone else can build on their values, right? For example, in the US it's democratic values, it's entrepreneurship values. And so that everyone has their own control, right? That they can decide their fate, decide what to do, not to do with the technology. Versus if you have only one country, for example, only China that is dominating all open source, then it creates a bunch of risks, right? It creates a bunch of biases, meaning that some of the biases that you can see in Chinese models could be replicated in the applications that you're using every day. And so that's not something that I think is positive for the world. You want basically every single country in the world to be able to train their own foundation models. We've seen that also recently with the announcement of Nvidia again, Jensen playing like in all the different fields with South Korea announcing really big initiatives for sovereign AI and for creating their own open source foundations for the ecosystem to thrive on local foundation models.
Right. But for the last few years, it's felt like American companies and especially a lot of the CEOs that I speak to, especially public companies, say that they want to buy a closed model from OpenAI or Anthropic or Google. Does inference and this next sort of phase of AI where you're actually running it, does that change? And do you think that appetite is growing now among corporate America for these open source models?
I think so. In the study we released we saw that almost 30% of the Fortune 500 now is using Hugging Face and using open models sometimes in addition to closed models, right? They're going to use like a big API for their more generic use case and then they're going to use open source for smaller use cases when they need things to be cheaper, faster, more secure. So that's what we're seeing, a world where companies use a mix of proprietary APIs and models that they train themselves based on open source. You're seeing that with Cursor today, right? Which for a long time has been kind of like the poster child of using APIs, right? Famously built on top of Anthropic API at the beginning. Today they released their new models that they built themselves, of course, based on open source.
Which company? I missed it. Which company did you... Cursor?
Cursor, yeah. The code editing. They announced it, I think today, their own models. So the future is really companies both using APIs and building, owning, fine-tuning their own models based on open source. And we're seeing more and more companies doing that.
Cursor is a good example, right? Because its costs are high when you're using an Anthropic model. So part of the point is to bring it in and be able to control, get leaner, get more efficient. And that's really like what Open Claw is about too, right? This phenomenon that has basically seen compute go through the roof because these agents just require so much more of it. It started on Anthropic and then I know you know Kimi in China was big for Open Claw. So let's get to like the Nvidia of all of this. So yes, it's been releasing open source models, but what it did with Nemotron Claw during GTC this week is launching free software. So Clem, I wonder if this feels like a gift to the open source community or do you read it more as a land grab?
It's alignment between their incentives and I think what's good for the field. You know, as I said, I think Nvidia needs as many people as possible to be able to train AI themselves, to build AI themselves, right? They don't want to end up in the world where only OpenAI and Anthropic are training and building models themselves, and everyone else is just using the APIs, right? That's really not a world that is beneficial to them. It's also a scary world for me. So that's why also Hugging Face is trying to fight that. But at the same time it's really great for the field and what we're seeing on Hugging Face is that people are really, and AI builders are really thankful to the contributions of Nvidia. They're now kind of like the biggest organization on Hugging Face in terms of numbers of followers for a big technology company. They release on Hugging Face more than one new model or data set every day for the past year. So the volume of their contribution to the field and to open source is really unprecedented. And people love that. You see how popular Jensen got at GTC, really giving his keynote in a stadium like a rock star. And I think part of this popularity comes from the contributions to community.
Right. He filled the whole hockey stadium. I know, it's amazing how much it's grown. It's like the Super Bowl here. Clem, you talk about this sort of community, right? And that's what sort of past open source technologies have really been built on, right? Like Linux as well was more community, but this is a company with incentives. You called them aligned incentives, but can we be certain that they're always going to be aligned? Because on one hand, developers, they get a free tool that works on any hardware, but on the other hand, if everyone is standardizing on Nvidia's platform, Nvidia, Jensen Huang, they end up controlling how AI agents get built and deployed. So, do you worry about those incentives always staying aligned?
I think it's a matter of kind of like looking what's better than other things. There are no pure kind of like, you know, openness, no control, and there's no kind of like full control, no openness. It's more a gradient. And the more it goes towards open source, the better for the world. The beauty of open source is that the control of something like Nemotron Claw is not as strong as it would be, for example, if it was an API, because people can fork it, they can contribute to it, they can modify it, they can own it themselves, right? So if at some point they're not happy about the work that Nvidia is doing, well, they can build their own, right? And for the field in general, the beauty of open source AI is that it's distributed and many, many players are playing a role, right? We talked about Chinese open source, but for example, in the US, Google is doing an amazing work on open source. They're releasing tons of models that are really, really used by the community. OpenAI released a model last summer. You have Allen AI that is contributing many models, data sets.
Let me push back a little bit on that front. Yes, OpenAI and Google are releasing open source models but it's not their best models the way that China is. I've been asking folks too about the OpenAI open source model and I get kind of like lukewarm responses. It was supposed to be Meta, right? Have they just completely fallen off this open source race? And I mean, are there any significant competitors especially versus what China is putting out there other than right now Jensen?
Well, I mean, building frontier open source models is hard, right? And I think even a big technology company like Meta sometimes can do very, very well, sometimes can struggle a little bit. I think I'm really thankful for the contributions that they've done to the field, you know, but others can continue and can keep pushing kind of like the frontier in open source. It doesn't need to be exactly the same frontier as closed source, right? Because open source has some unique advantage that makes it kind of like different than closed source. For example, you can run a closed source model on your hardware which makes it actually free. So even if open source is a little bit not as good as closed source, it's tremendously useful for the community and for a lot of use cases. In terms of players, I'm really excited about, you know, new American startups. Obviously, entrepreneurship is the main strength of the US. And I think if we can see more startups starting to contribute to open source, maybe you've seen Reflection AI that raised big rounds in the US and doing sovereign AI with South Korea. If we can see more and more American startups contributing to open source, we can definitely catch up to Chinese open source and have a big impact in the world.
Yeah, Reflection is certainly one we have our eye on. Hoping to talk to them soon. Good question from one of our viewers right now that asks if you can comment on AMD's open source platform compared to Nvidia's.
They're getting better. Their software is so much better than it was maybe two, three years ago. And you know, they really emerge as kind of like a good alternative for a lot of these workflows. And I think they're just getting started really. And we expect them to make a lot more progress in the next few months.
Okay. I know you dodged it a little bit earlier so I'm gonna ask it again. I think like we talked a while ago and we talked about sort of Zuckerberg building the ecosystem for open source models and I wonder, you know, and tell me if I'm thinking about this in the wrong way, right? Like Mark Zuckerberg has always sort of lamented the idea that Apple and Google created the operating system for smartphones and he had to build on top of that and you have like disputes even today that Apple, you know, taxes developers on top of its platform through the App Store. Could that scenario happen for open source AI? Could Jensen, you know, build the ecosystem and then find ways of charging people? And would you rather Jensen Huang be doing that than Mark Zuckerberg?
I think it's much, much harder to do that on open source than in any other things because again open source, it's super easy to fork. It's super easy to modify yourself. It's super easy to host it yourself. And so in a way you don't really build a moat, you don't really kind of like lock the community or lock the users with open source. Or it has to be, you know, thousand times bigger and better than everything else but I don't really see that happening. So in my opinion the risk is much less of concentration of power for open source, it is more for APIs and for agent APIs. Right now we're seeing a lot of concentration happening on code for example and overall Caplacropic solutions, but in open source it's almost impossible to do just by the nature of the technology that everything is available and so everyone can fork it, modify and switch very easily if they want to.
