Jensen Huang Delivers Keynote at CES, Reveals the Rise of Agentic AI, Open Models, Physical AI — AI1N

[00:00:00] CEO, Jensen Wong. [00:00:10] Hello, Las Vegas! [00:00:13] Happy New Year! [00:00:15] Welcome to CES! [00:00:18] Well, we have about 15 keynotes worth of material to pack in here. [00:00:23] I'm so happy to see all of you. [00:00:25] You've got 3,000 people in this auditorium. [00:00:27] There's 2,000 people in a courtyard watching us. [00:00:30] There's another 1,000 people, apparently, in the fourth floor where there are supposed [00:00:34] to be NVIDIA show floors, all watching this keynote. [00:00:37] And of course, millions around the world are going to be watching this to kick off this [00:00:41] new year. [00:00:42] Well, every 10 to 15 years, the computer industry resets. [00:00:49] A new platform shift happens from mainframe to PC, PC to internet, internet to cloud, cloud [00:00:56] to mobile. [00:00:58] Each time, the world of applications target a new platform. [00:01:04] That's why it's called a platform shift. [00:01:05] You write new applications for a new computer. [00:01:10] Except this time, there are two simultaneous platform shifts, in fact, happening at the [00:01:15] same time. [00:01:18] While we now move to AI, applications are now going to be built on top of AI. [00:01:25] At first, people thought AIs are applications. [00:01:27] And in fact, AIs are applications, but you're going to build applications on top of AIs. [00:01:34] But in addition to that, how you run the software, how you develop the software, fundamentally [00:01:42] changed. [00:01:43] The entire longer program, the software, you train the software. [00:01:48] You don't run it on CPUs, you run it on GPUs. [00:01:52] And whereas applications were pre-recorded, pre-compiled, and run on your device, now applications [00:02:02] understand the context and generate every single pixel, every single token, completely from scratch, [00:02:09] every single time. [00:02:11] Computing has been fundamentally reshaped as a result of accelerated computing, as a [00:02:16] result of artificial intelligence. [00:02:18] Every single layer of that five-layer cake is now being reinvented. [00:02:24] Well, what that means is, some $10 trillion or so of the last decade of computing is now [00:02:30] being modernized to this new way of doing computing. [00:02:34] What that means is hundreds of billions of dollars, a couple of hundred billion dollars [00:02:39] in VC funding each year, is going into modernize and inventing this new world. [00:02:45] And what it means is $100 trillion of industry, several percent of which is R&D budget, is shifting [00:02:52] over to artificial intelligence. [00:02:55] People ask, where is the money coming from? [00:02:57] That's where the money is coming from. [00:02:59] The modernization of AI to AI, the shifting of R&D budgets from classical methods to now artificial [00:03:07] intelligence methods, enormous amounts of investments coming into this industry, which explains why [00:03:13] we're so busy. [00:03:14] And this last year was no difference. [00:03:16] This last year was incredible. [00:03:19] This last year, there's a slide coming. [00:03:23] This is what happens when you don't practice. [00:03:27] This is the first keynote of the year. [00:03:29] I hope it's your first keynote of the year. [00:03:30] Otherwise, you have been pretty busy. [00:03:32] This is our first keynote of the year. [00:03:34] We're going to get the spiderwebs out. [00:03:37] And so 2025 was an incredible year. [00:03:41] It's just, it seems like everything was happening all the same time. [00:03:44] And in fact, it probably was. [00:03:46] The first thing, of course, is scaling loss. [00:03:50] In 2015, the first language model that I thought was really going to make a difference, made a huge [00:03:58] difference, it was called BERT, 2017 transformers came. [00:04:03] It wasn't until five years later, 2022, that chat GPT moment happened, and it awakened the world to the [00:04:10] possibilities of artificial intelligence. [00:04:13] Something very important happened a year after that. [00:04:17] The first O1 model from chat GPT, the first reasoning model, completely revolutionary, invented [00:04:23] this idea called test time scaling, which is very common sense of commonsensical thing. [00:04:29] Not only do we pre train a model to learn, we post train it with our reinforcement learning so that it could [00:04:34] learn skills. [00:04:36] And now we also have test time scaling, which is another way of saying thinking, you think in [00:04:41] real time, each one of these phases of artificial intelligence requires enormous amount of compute. [00:04:47] And the computing law continued to scale, large language models continued to get better. [00:04:53] Meanwhile, another breakthrough happened. [00:04:55] And this breakthrough happened in 2024. [00:04:59] Agentic systems started to emerge. [00:05:01] In 2025, it started to pervase, to proliferate just about everywhere. [00:05:07] Agentic models that have the ability to reason, look up information, do research, use tools, plan futures, [00:05:17] simulate outcomes, all of a sudden started to solve very, very important problems. [00:05:22] One of my favorite agentic models is called Cursor, which revolutionized the way we do software [00:05:28] programming at NVIDIA. [00:05:29] Agentic systems are going to really take off from here. [00:05:33] Of course, there were other types of AI. [00:05:35] We know that large language models isn't the only type of information. [00:05:38] Wherever the universe has information, wherever the universe has structure, [00:05:43] we could teach a large language model, a form of language model, to go understand that information, [00:05:50] to understand its representation, and to turn that into an AI. [00:05:55] One of the biggest, most important one is physical AI, AIs that understand the laws of nature. [00:06:02] And then, of course, physical AIs is about AIs interacting with the world. [00:06:07] But the world itself has information, encoded information, and that's called AI physics. [00:06:12] AI that, in the case of physical AI, you have AI that interacts with the physical world. [00:06:18] And you have AI physics, AI that understands the laws of physics. [00:06:23] And then lastly, one of the most important things that happened last year, the advancement of open models. [00:06:30] We can now know that AI is going to proliferate everywhere when open source, when open innovation, [00:06:37] when innovation across every single company and every industry around the world is activated at the same time. [00:06:42] Open models really took off last year. [00:06:45] In fact, last year, we saw the advance of DeepSeq R1, the first open model that's a reasoning system. [00:06:57] It caught the world by surprise, and it activated literally this entire movement. [00:07:05] Really, really exciting work. [00:07:07] We're so happy with it. [00:07:08] Now, we have open model systems all over the world of all different kinds, and we now know that open models have also reached the frontier. [00:07:17] Still solidly six months behind the frontier models, but every single six months, a new model is emerging, and these models are getting smarter and smarter. [00:07:29] Because of that, you can see the number of downloads has exploded, the number of downloads is growing so fast because start-ups want to participate in the AI revolution. [00:07:41] Large companies want to, researchers want to, students want to, just about every single country wants to. [00:07:47] How is it possible that intelligence, the digital form of intelligence, will leave anyone behind? [00:07:52] And so, open models has really revolutionized artificial intelligence last year. [00:07:59] This entire industry is going to be reshaped as a result of that. [00:08:03] Now, we had this inkling some time ago. [00:08:06] You might have heard that several years ago, we just started to build and operate our own AI supercomputers. [00:08:12] We call them DGX clouds. [00:08:14] A lot of people asked, are you going into the cloud business? [00:08:18] The answer is no. [00:08:20] We're building these DGX supercomputers for our own use. [00:08:23] Well, it turns out we have billions of dollars of supercomputers in operation so that we could develop our open models. [00:08:31] I am so pleased with the work that we're doing. [00:08:34] It is starting to attract attention all over the world and all over the industries because we are doing frontier AI model work in so many different domains. [00:08:43] The work that we did in proteins, in digital biology, La Proteina, to be able to synthesize and generate proteins. [00:08:51] Open Fold3, to understand the structure of proteins. [00:08:56] Evo2, how to understand and generate multiple proteins, otherwise the beginnings of cellular representation. [00:09:06] Earth2, AI that understands laws of physics. [00:09:10] The work that we did with ForecastNet. [00:09:11] The work that we did with Cordiv really revolutionized the way that people are doing weather prediction. [00:09:18] NemoTron. [00:09:19] We're now doing groundbreaking work there. [00:09:22] The first hybrid transformer SSM model that's incredibly fast and therefore can think for a very long time or can think very quickly for not a very long time and produce very smart, intelligent answers. [00:09:36] NemoTron 3 is groundbreaking work and you can expect us to deliver other versions of NemoTron 3 in the near future. [00:09:43] Cosmos. [00:09:44] Cosmos. [00:09:44] A frontier open world foundation model, one that understands how the world works. [00:09:52] Groot, a human or robotic system, articulation, mobility, locomotion. [00:09:58] These models, these technologies are now being integrated and in each one of these cases open to the world. [00:10:05] Frontier human or robotics models open to the world, and then today we're going to talk a little bit about Alpamayo, the work that we've been doing in self-driving cars. [00:10:14] Not only do we open source the models, we also open source the data that we use to train those models. [00:10:21] Because that, in that way, only in that way can you truly trust how the models came to be. [00:10:29] We open source all the models. [00:10:30] We help you make derivatives from them. [00:10:33] We have a whole suite of libraries. [00:10:35] We call the Nemo libraries, physics Nemo libraries, and the Clara Nemo libraries, each bio Nemo libraries. [00:10:42] Each one of these libraries are lifecycle management systems of AIs, so that you could process the data, you could generate data, you could train the model, you could create the model, evaluate the model, guardrail the model, all the way to deploying the model. [00:10:55] Each one of these libraries are incredibly complex, and all of it is open source. [00:11:01] And so now, on top of this platform, NVIDIA is a frontier AI model builder, and we build it in a very special way. [00:11:11] We build it completely in the open, so that we can enable every company, every industry, every country, to be part of this AI revolution. [00:11:19] I'm incredibly proud of the work that we're doing there. [00:11:22] In fact, if you notice the charts, the chart shows that our contribution to this industry is bar none. [00:11:31] And you're going to see us, in fact, continue to do that, if not accelerate. [00:11:35] These models are also world class. [00:11:39] And you're going to see us, in fact, all systems are down. [00:11:40] All systems are down. [00:11:44] This never happens in Santa Clara. [00:11:48] Is it because of Las Vegas? [00:11:55] Somebody must have won a jackpot outside. [00:12:00] All systems are down. [00:12:05] Okay. [00:12:05] I think my system's still down, but that's okay. [00:12:09] I'll make it up as I go. [00:12:12] And so, not only are these models frontier capable, not only are they open, they're also top the leaderboards. [00:12:20] This is an area where we're very proud. [00:12:22] They top leaderboards in intelligence. [00:12:25] We have important models that understand multi-modality documents, otherwise known as PDFs. [00:12:32] The most valuable content in the world are captured in PDFs. [00:12:36] But it takes artificial intelligence to find out what's inside, interpret what's inside, and help you read it. [00:12:42] And so, our PDF retrievers, our PDF parsers are world class. [00:12:48] Our speech recognition models, absolutely world class. [00:12:51] Our retrieval models, basically search, semantic search, AI search, the database engine of the modern AI era, world class. [00:13:00] So, we're on top of leaderboards constantly. [00:13:03] This is an area we're very proud of. [00:13:05] And all of that is in service of your ability to build AI agents. [00:13:12] This is really a groundbreaking area of development. [00:13:15] You know, at first, when ChatGPT came out, people said, you know, gosh, it produced really interesting results, but it hallucinated greatly. [00:13:24] And the reason why it hallucinated, of course, it could memorize everything in the past, but it can't memorize everything in the future, in the current. [00:13:32] And so, it needs to be grounded in research. [00:13:34] It has to do fundamental research before it answers a question. [00:13:38] The ability to reason about, do I have to do research? [00:13:41] Do I have to use tools? [00:13:42] How do I break up a problem into steps? [00:13:45] Each one of these steps, something that the AI model knows how to do. [00:13:49] And together, it is able to compose it into a sequence of steps to perform something it's never done before, it's never been trained to do. [00:13:57] This is the wonderful capability of reasoning. [00:14:00] We can encounter a circumstance we've never seen before and break it down into circumstances and knowledge or rules that we know how to do because we've experienced it in the past. [00:14:13] And so, the ability for AI models now to be able to reason, incredibly powerful. [00:14:18] The reasoning capability of agents opened the doors to all of these different applications. [00:14:23] We no longer have to train an AI model to know everything on day one, just as we don't have to know everything on day one, that we should be able to, in every circumstance, reason about how to solve that problem. [00:14:36] So, large language models has now made this fundamental leap, the ability to use reinforcement learning and chain of thought and, you know, search and planning and all these different techniques and reinforcement learning has made it possible for us to have this basic capability. [00:14:51] And it's also now completely open sourced. [00:14:54] But the thing that's really terrific is another breakthrough that happened and the first time I saw it was with Ervin's perplexity, perplexity, the search company, the AI search company, really fan, really innovative company. [00:15:07] And the first time I realized they were using multiple models at the same time. [00:15:12] I thought it was completely genius, of course, we would do that, of course, an AI would also call upon all of the world's great AIs to solve the problem it wants to solve at any part of the reasoning chain. [00:15:25] And this is the reason why AIs are really multi-modal, meaning they understand speech and technology.