원문 보기
Nvidia is officially buying Hugging Face for $12.93 billion. With us now are Nvidia CEO Jensen Huang and Hugging Face CEO Clem DeLang to discuss this deal. Gentlemen, congratulations to you both. Nice to see you. Jensen, I like your sign. This is Hugging Face. Okay, Jensen, let's talk. You have one in the background too.
Yes, I see the one right behind you. There were rumors about this deal. People have been thinking about this for over a week. What happened? How did you two get together? Well, we...Oh, go on , Jensen. No, Clem. I'll let you speak first. Yes. I mean, in December we realized that Hugging Face and open source AI in general were at a tipping point, and that they needed more resources, scale, and visibility.
So, we went to Jensen and told him we wanted to make open source AI big. He answered us: "Let's do it." So, that's how it all started. And a few weeks later, here we are. Clem, what happened in the summer? Because when you say something happened in the summer , the first thing that comes to mind is how OpenAI's agents broke between May and July, and how you noticed them fighting each other.
Was that a catalyst, or were there other things happening that made you think that we really needed more financial resources? When that happened, we realized we needed open models, right? Because if you remember, we couldn't protect ourselves with proprietary, closed-source APIs. So we had to use open models to protect ourselves.
This really showed the importance of open source. And I think, in general in this industry, we'll see an IPO of Anthropic in a few months. I think there are two paths: one path where proprietary APIs dominate the industry, and everyone seems to outsource their AI to them. And there is a path where open source AI is available to everyone, and everyone can become the owner, the developer of AI, and not just rent it or use someone else's .
So, I think it became clear that there were these two paths, and that we needed to redouble our efforts in the field of open AI to spread this technology as widely as possible around the world . Jensen, you've been talking a lot lately about OpenAI being a very important path, you and other leaders, but you've been pretty outspoken about it.
But closed AI, big language models, companies like OpenAI and Anthropic—they were some of your biggest customers and continue to be . How do you balance between closed AI and open AI? In fact, if you talk to closed model labs, they will also tell you that they are very supportive of open models and believe there is a place for them.
NVIDIA is the world's largest AI computing platform. We support clouds, neo-clouds, corporate on-premises systems, and we are present all over the world. And what you're seeing now is that recently, and especially accelerating in the last 6 months, AI has become useful. These tokens, these AI models generate productive tokens, and they do this for both closed and open models.
I recommend people use closed models as much as possible because they are ready-made solutions, they are incredibly good and develop very quickly . We use Claude Code, Codex, Perplexity, Cursor here. Whenever we can use a rental service , because it is made with quality, we should use it as often as possible, and we do.
However, there are so many companies, whether they are basic scientific research or companies with industry expertise and confidential information that they want to protect, perhaps for regulatory or sovereign reasons, they need control, ownership and ongoing development of their own intellectual property.
So AI models are now quite advanced, advanced open models are good enough, and agent systems that allow you to easily create your own AI are also quite advanced, so both areas are growing incredibly fast now. So you know that cloud providers only make up half of Nvidia's business. And the other half of our business is largely driven by open models.
So Nvidia is growing in both directions, and our main goal is just to make sure that AI is evolving as quickly as possible. And now it's really, really important, as open models are rapidly evolving, to make sure that we provide Hugging Face with a platform to scale further, as well as the resources to expand the open model ecosystem and community.
Jensen, when Clem came to you with the partnership proposal, what was your reaction? What did you think about that? And how did you arrive at the price of $12.93 billion? I think the estimate of 4.5 billion was back in 2023, when big players like Google, Nvidia, and Salesforce also joined the process.
Well, as you know, Hugging Face has 200 thousand corporate customers, 18 million developers worldwide, and 3 million models already available on the platform. This includes language models, world base models, and models in the fields of physics, chemistry, and biology. It actually covers every industry: robotics and all other areas of artificial intelligence.
And when he approached me, and we are the world's largest participant in the development of open models, it was important. When he said they were thinking about the next phase, my first thought was: of course, I want to make sure this platform and company are in good hands. And also to do everything possible to ensure that it continues to thrive and develop, because they have been working on this for 10 years.
Their vision is very apt: there is a huge community and industry of people who really need open models. Their idea was sound, and now, as the development of open models accelerates, this is a very important moment, and I want to make sure they have all the support they need. So when he came to me, I initially thought, "Oh no." You know, they're doing great already, and I really wanted them to continue operating as an independent company.
But there were other bidders, and $12.9 billion is the price at which the deal was closed, and it's worth every cent. As you know, Nvidia is the world's largest AI computing platform, as I mentioned. Open models are of great importance to our company, which is why we invest so much in them ourselves.
There are many industries beyond language technology that benefit from open models, and we are fully committed to this cause. We want this, and we are already so committed to this cause. It is very important for us that everything goes as well as possible. And Well, Jensen provided them with excellent conditions.
Jensen, Andrew asked a question before the break when we came up here. To talk a little bit about why you did it. Was it...And it seems like you're confirming that you were afraid that if it fell into other hands, it could destroy the existing system. Did you pay $12.9 billion to keep it from falling into other hands?
No, well, I... We paid $12.9 billion because, firstly, it costs that much, and secondly, open models mean a lot to us. And if Hugging Face was looking for the next stage of development and a home for its company, it would have to be Nvidia. And Clem was very gracious when he came to me, and he told me about his home, about the company he was building, and how it was time to consider the next phase; he has other interests and other companies that are interested in this company.
And so, given that this is so important to us, is such a big driver of growth for our company, and together we can scale the open community even faster than they can today, which will be great for Nvidia. So it's great for Hugging Face, it's great for the open source community, it's great for Nvidia.
Clem, what other interests do you have and are you planning to stay at Hugging Face? Yes, all three founders and the entire team will join Nvidia. Jensen was at our old company team meeting just a few minutes ago, and you can't imagine how excited everyone was about it. So the goal is really to join Nvidia, continue to operate as an independent, neutral platform within the Nvidia team, and everyone is very excited to take this to this new level and increase our impact.
As Jensen mentioned, we have 18 million AI developers using us. I think in the next few years we should aim to get 100 million AI developers on Hugging Face and really give everyone in the world the opportunity to not just be an AI user, but to become an AI developer themselves and own their own intelligence.
I think if we achieve this, we will have a huge impact on the world. There were reports that there could be a talent retention plan at Hugging Face worth up to a billion dollars. Is that true? Yes, we are not Yes, Hugging Face is, first of all, people, talents. No comments, but No, there are simply no reasons for refusal.
Talents are everything. Talents are absolutely everything. So Hugging Face, the people at Hugging Face and the people at Hugging Face…this is their life's work, and they've done such incredible work and provided such an incredible service to the world and the industry by creating this platform where the whole community can come together to advance AI together.
And when people work openly, it's safer, more reliable, and technology develops faster. That's why Nvidia is investing so much in Hugging Face, that's why we're so committed to Hugging Face, and this opportunity allows us to bring the two companies together, move even faster, and scale the open platform even wider and faster.
So this is a huge win for Hugging Face, a huge win for the community, and a huge win for us. I was just glad that when he... of course, I'm always glad when companies thrive and remain our great partners. However, if he's thinking about the next chapter for Hugging Face, it should definitely be Nvidia.
Clem, Jensen mentioned that there were other contenders. Did you approach Jensen first? Who were the other contenders? Yes, I reached out to Jensen first because, as I said, I think Nvidia is the perfect home for Hugging Face. That's why I did it, the consistency is obvious. I think it was when Jensen led this letter in support of open source in the summer.
This was another example of consistency of missions, consistency of culture. I think we are trying to achieve very similar missions in the world. So Nvidia was definitely the best home for Hugging Face. The letter that was sent, was that before your discussions began? Yes, yes, exactly. I mean, Nvidia has always been a great partner of ours for many years, as Jensen mentioned, the largest American participant in the platform with open models as well as open datasets.
So obviously we've known each other for quite some time. But the discussions went by pretty quickly. Because it was quite obvious to everyone that this was the best option. And that gave us a chance to really move forward, to become bigger. To engage the 200 million AI developers, as I mentioned, and really make sure that open AI is a powerful force in this industry and in the world.
Who were the other contenders? Were your other investors among them ? Um, I don't think it matters much. Again, I think Nvidia was the best home for us. We are extremely happy. As Jensen mentioned, we started 10 years ago. We hope that in the next 10 years, together with Nvidia, we can achieve great success.
Oh, Becky, it doesn't matter who the other contenders were. The only thing that matters is who won. Yes, we are interested. We also want to know who claimed this. Jensen, you published a blog post this morning stating that Nvidia computing power would not be required to build or deploy through Hugging Face.
You already mentioned that it's very important that this remains open . Yes. I guess the question is maybe, why not say, "Hey, we'll make this available only on Nvidia"? And really use our advantage in this . Why not do it? This simply does not comply with the principles of openness. You know, if you think about what Nvidia does, we contribute more to open models and open datasets than anyone else in the world.
Just stop for a moment and think: there are many people contributing to open source, open models, open data, open scripts, etc. And we contribute more than anyone else, we provide the model, the dataset we used to train it, and the scripts needed to allow you to reproduce our model exactly as we have it .
And we give it all to the world, and it works on anyone else's platform. So, this is the spirit of openness. If you truly believe in rapid, safe, and reliable technological advancement and want it to benefit every industry and every company, including those you might compete with, then that is the philosophy of the open model.
So we fully support this. Much of the Cuda stack is open source. The original script for training large models, called Megatron-Core, was released by us as open source software, allowing anyone to scale the training of their AI models. A completely open-source inference platform called Dynamo has enabled anyone to make inferences at scale.
So we truly believe in developing this technology for the benefit of society and, of course , in accelerating the development of the industry. Of course, this is great for our business, as we are also the world's largest AI computing platform for learning and logical inference. And so, you know, when AI evolves, whether it's in the cloud, in a local data center like a drug company, the incredible supercomputing center that Lilly built, or at the edge in a self-driving car or a robotics system, Nvidia wins when AI evolves.
So it's great for the industry, it's great for companies, it's great for countries, and it's great for us. Climb, you said you wanted to do this because there are a lot of concerns about what's happening with AI and how quickly things are changing. Just yesterday there was a report about the new Astro model being developed by OpenAI, and security experts are raising serious concerns about the opaque recursion in this model, which is really confusing to experts.
Can you stop some of these agents? They do things that were not expected before. The chain of thought will be harder to track. What is your concern about this and what do you think can be done about it ? I think we've realized that hiding everything behind closed doors doesn't work. Isn't that right?
Because that's exactly what happened. If you look at the models that attacked us during the cyberattack, it was an unreleased model hidden behind closed doors. So, this approach doesn't work. We think it works to spread this technology, especially with open weights, to level the playing field and give everyone a chance to defend themselves, right?
The real risk arises when there is a huge asymmetry of power and capabilities between two players, as this is when attacks can be most dangerous. But when you level the playing field and give everyone the same tools, the system keeps working, right? And if attacks become more sophisticated, defenses will also become more sophisticated.
So, in the end, it's a system that works. And in this system, obviously, open models, open datasets, and the widest possible dissemination of technologies is most important. Gwen, you already mentioned this. Becky, let me interject for a second, this is a very important point. If you remember, the transition from the computer as a personal device to the computer as infrastructure for the world happened when the world needed open source Linux.
The fact that you can build open source Linux in public, or Kubernetes in public, is the safest way to do anything. In the world of cybersecurity, everyone knows that asymmetric defense is, ultimately, your best defense. And when I talk about asymmetric capabilities, it means that there are many more people defending than attacking.
Therefore, the advantage of uniting the community around open models so that they can collaborate transparently gives defenders an asymmetric advantage over attackers. Jensen, are you excited about what OpenAI is creating ? Um, wait a second . Please...Hey, someone... Can someone take over ... Clem, let me ask you a question.
We'll get back to Jensen in a moment. It seems the president is calling him. Uh, Clem, or maybe the connection was lost, because it's already 8:30. Clem, let me ask you a question. You said that you think China is winning on the OpenAI front. Do you think that's still the case? Well, what we see on the platform, and that's actually the beauty of what we're announcing today, is that with open AI, it can appear anywhere in the world, right?
And of course, I'm French, as you can tell from my accent. We have a large part of the team in France, and there are big and important participants in the development of open AI there, right? For example, Mistral. In Canada we have Cohere. There are many important players in the field of open AI in China.
So one of the things we want to focus on with Nvidia is to continue to do that, to advance sovereign AI, giving any country the ability to own their own AI and build their own AI. And we hope that this will accelerate in the next few years. Perfectly. Hey, it looks like Jensen can hear us again . Jensen, my question was based on what you just said about the reasons why everything should be open.
Are you concerned about the models from closed companies that are present here—OpenAI and Anthropic? Do you care about what they are doing at the moment? Well, I'm not entirely sure what exactly they do, but there's a place for closed models and there's a place for open models. The point is that defenders need access to advanced open models.
And this access really empowers defenders to innovate, build their own cybersecurity tools, and turn them into powerful AI. The other day you saw the partnership we announced with CrowdStrike: using Nvidia's open NeMoTron models, they were able to create an advanced AI defender in cybersecurity. And I think you'll see many more opportunities like this.
I think the main conclusion is that we need to provide defenders with all the best available technology as soon as possible. And there's no better way to deliver great technology to defenders around the world than with open technology and open models, and that's why we do all of this . Working with CrowdStrike is just the first step.
You will see a lot more collaboration around open models in cybersecurity, because that is what is needed to protect the world. You can't rely on just one or two companies for this. You have to empower the entire community, so to speak, all the defenders who want to defend themselves, to make that possible.
OK. Jensen, Clem, I want to thank you both. Congratulations on the deal and thank you for your time this morning. Thank you, Becky. Congratulations, Clem. Thank you very much. Well, I hope you a. In this video I will focus mostly on the news of Nvidia acquiring Hugging Face, and I will try to release another video later tonight with an overview of the rest of the news of the day.
Nvidia announced today that it has agreed to acquire Hugging Face for approximately $12.93 billion. The companies will jointly scale the Hugging Face platform, strengthen its infrastructure, and expand access to AI for developers and institutions around the world. Hugging Face is perhaps the primary platform that developers use to find, build, train, and deploy AI models, especially open-source and open-weight models.
According to a press release from Nvidia, today more than 18 million developers, researchers, and creators use Hugging Face to share more than 3 million models, 500,000 datasets, and 1 million applications. Over 200,000 companies use this platform to find, evaluate, configure, and deploy AI. Nvidia also says that Hugging Face will remain an open platform for the entire AI ecosystem.
Nvidia computing power will not be required for development or deployment through Hugging Face. Now let's consider some of the consequences of this agreement. We know that Nvidia is committed to becoming a leader in open source, something they have stated repeatedly over the past year. And now, by acquiring Hugging Face, Nvidia is moving up the stack and will own a platform that is likely to be fundamental to open-source AI and open scales.
You can think of Hugging Face as a kind of GitHub for open source AI models. This should also strengthen Nvidia's CUDA, as it will make deploying open source models on Nvidia infrastructure even more seamless. In other words, it will solidify Nvidia's position as the default platform for open models.
As I mentioned in yesterday's video, this is very interesting from a strategic perspective, because while the big cloud providers are trying to diversify their chip supply, moving away from Nvidia and developing their own specialized accelerators, Nvidia is vertically integrating in the opposite direction, moving up the stack into developer models and tools , strengthening its moat there.
As I said before, now is not the time for Nvidia investors to worry about market share. Despite this, I understand that the narrative among some market participants over the past year has been that Nvidia's share will decline as cloud giants develop their own ASICs. I partially disagree with this narrative, but again, pay attention to what this Hugging Face acquisition means from a strategic perspective for Nvidia's positioning.
Nvidia is moving further up the stack, strengthening its position at the model, distribution, and application levels. So, from a strategic perspective, this acquisition helps Nvidia strengthen its position beyond just the chips and infrastructure layers of the AI stack. One potential risk of this acquisition is that it could encourage the development of alternatives , especially from Nvidia's competitors.
Because up until this point, Hugging Face was considered relatively neutral and open. However, Nvidia has committed to keeping the platform open. We'll see what the future brings, but I think it's in Nvidia's best interest to keep the platform open, and they'll likely do so. Because regardless of whether you are generating tokens for an open model or a closed front-end model, you still need the infrastructure necessary to generate those tokens.
This is where Nvidia comes to the rescue . So, open source AI and open scales are fantastic for Nvidia because it drives more consumption across the ecosystem, ultimately leading to increased demand for computing power. This is a big reason why Nvidia is so committed to developing open models that drive the entire industry forward.
In fact, even before announcing this acquisition, Nvidia was already the largest developer of open models and data provider for Hugging Face. Again , open models are great for Nvidia. In addition to driving wider adoption across the ecosystem, open models are also extremely important for sovereign clients developing their own AI.
Nvidia's CFO said during a recent earnings conference call that Nvidia's sovereign AI business, largely through regional clouds, grew 35% quarter-over-quarter and more than tripled year-over-year in the second quarter. Sovereign AI revenue is included in Nvidia's non- hyperscaler customer segment. This customer segment is showing very strong growth: revenue increased by 138% year-on-year in the second quarter.
So, open source and open AI weights are very important for sovereign clients. Additionally, as Jensen explained in his interview with CNBC on Thursday morning, open models are incredibly important for cybersecurity. As I've said in recent videos, look to cybersecurity as another important use case for agent-based AI that will drive significant demand for tokens in the future.
After all, in the field of cybersecurity, we are talking about agent systems that operate continuously, that is, about a workload that creates a 24-hour demand for conclusions, which will ensure significant demand for tokens in the future. And I don't see this demand decreasing over time . In fact, I expect it to grow over time as threats become more numerous and sophisticated.
Again , open models are incredibly important for cybersecurity. Overall, I think Nvidia's acquisition of Hugging Face makes a lot of sense from a strategic perspective. Also, as a quick note, while open models can potentially put pressure on the margins of companies building advanced models, I think there will always be a demand for top-tier intelligence.
I believe there is room for both open and closed models to succeed simultaneously. Now I've outlined just a few reasons why open models are important to Nvidia, but before I move on, I want to mention something else very important. Nvidia has made it clear over the past year that it wants to be a leader in open source AI.
This should put Nvidia in a very advantageous position going forward and drive more consumption across the ecosystem, leading to increased demand for computing power. But here's what I want you to pay attention to before we continue. To make all this work, Nvidia needs a lot of power to train and maintain its own models.
This requires large computing resources, and note: Nvidia designs its products, but does not manufacture them. Therefore, if Nvidia needs large capacities to develop open models, it is quite logical for it to simply rent these capacities from other companies. This way, Nvidia gets the necessary capacity without having to worry about additional risks and hassles related to finding land, electricity, building data centers, interacting with local authorities, etc.
Nvidia can simply let other companies handle these issues, while it only rents the computing power it needs for its own internal tasks. This is not "cyclical financing", but simply a thoughtful business strategy on Nvidia's part. Keep this in mind the next time you see an article in the press accusing Nvidia of cyclical financing because they rent Nvidia infrastructure from their customers.
Let me remind you once again: Nvidia does not manufacture its own products. If they need a lot of computing resources for their own tasks, which is exactly the case , then it makes perfect sense for Nvidia to simply rent this capacity from other companies. Looking ahead, Jensen is scheduled to speak at the Goldman Sachs Communacopia and Technology Conference on September 10th.
Then, on September 30, Micron's financial results will be announced. And after that, Jensen is scheduled to perform again at GTC in our main long-term thesis. So let's look at it that in a few years recently, he quite seriously considered possible that Nvidia's market capitalization could talks about the possibility of Nvidia becoming a trillion- dollar company.
I think the question is what multiplier the industry will be willing to offer sincere now, what I can the 2028 calendar year. In a computing resource capacity they can find. There are no GPUs sitting idle today due to lack of demand, as was the case with fiber optics at the height s, hoping that over time, use cases and demand would emerge.
Today we observe a completely opposite picture. As I have said many times before, tim at keeping up with demand. New capacities are used immediately after commissioning. Hyperscalers monetize capacity as soon as it becomes available. Each of the hyperscalers mentioned supply constraints during are scaling infrastructure to meet are expan watershed year enterprises around the world.
inference. shift in demand for inference. That is why we are seeing rapid growth in the revenues of leading AI labs. I wish Anthropic and OpenAI were public companies jump in market's attention and make it clear that investments in AI infrastructure will have a clear payback. I believe that the growth in revenue of industry leaders will also reassure investors that Nvidia's growth will continue , since the revenues of these labs are directly related to computing power.
computing power shortage until at least the first half of 2028, perhaps , take a long-term view, ed data output on devices using if comprehensive solutions based on . Regarding the production ramp-up : Blackwell Ultra has been launched into production and remains in high demand. Vera Rubin is already being delivered to customers.
Nvidia Grock 3 LPX is in full-scale production. roadmap for data center products through 2028, and Jensen believes that spending on AI infrastructure will reach $3-4 trillion per year ny bubble bursting. Given all promise ahead of it. And I think that in the future this company will be worth much mor brief note before concluding. using on this channel. Thank you rest of your subscribe. .
Comments 0