NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang

[00:00:00] I look for things like these [00:00:02] Whoa, oh, oh, I look for, I look for [00:00:07] Whoa, oh, oh, I look for things like these [00:00:11] Oh, we're sky high up here rolling [00:00:15] And we won't turn around [00:00:17] Yeah, we're on our way up [00:00:19] And ain't nobody can bring us down [00:00:22] We turn up, we're coming through [00:00:24] We won't stop, no way, we can't lose [00:00:30] This is how intelligence is made [00:00:46] A new kind of factory [00:00:50] Generator of tokens [00:00:53] Building blocks of AI [00:00:55] Tokens have opened a new frontier [00:01:02] Turning data into knowledge [00:01:05] And knowledge into action [00:01:09] They decipher nature's code [00:01:15] To reveal hidden cures [00:01:23] Mirror our cities [00:01:25] And tirelessly navigate our world [00:01:34] To keep us moving [00:01:38] Tokens learn from the way we move [00:01:44] Deliver the care we need [00:01:55] We walk into danger [00:02:01] To protect the ones we love [00:02:07] To protect the ones we love [00:02:07] They see what we cannot [00:02:11] So the steadiest hands [00:02:18] Never work alone [00:02:19] Every idea tokens spark [00:02:25] Every success they share [00:02:29] Puts what matters [00:02:32] Within reach [00:02:33] Tokens are helping us break new ground [00:02:47] On a scale never seen before [00:02:52] On a scale never seen before [00:02:53] So we can reach [00:03:00] For the sky [00:03:04] And beyond [00:03:06] Together [00:03:16] Together [00:03:16] Together [00:03:16] We take the next great leap [00:03:18] Into a bright new future [00:03:21] Built [00:03:23] For all mankind [00:03:25] And here [00:03:34] And here [00:03:34] Is where it all begins [00:03:36] Welcome to the stage [00:03:51] NVIDIA founder and CEO [00:03:53] Jensen Wong [00:03:54] Hello [00:04:03] Hello Berlin [00:04:04] Hello Berlin [00:04:06] Hello Berlin [00:04:08] Welcome to GTC [00:04:14] It's great to be here [00:04:15] Before we start [00:04:17] Let's give it up [00:04:19] For our pre-game show hosts [00:04:22] Smart and adorable [00:04:34] Smart and adorable [00:04:36] They spoke about AI infrastructure [00:04:38] Sovereign AI [00:04:40] AI safety [00:04:42] And physical AI [00:04:44] They covered an enormous amount of ground this morning [00:04:47] And we are going to take it one step further [00:04:49] It's really interesting [00:04:52] What do Oriel, Revolut, Spotify, Merc, Vodafone, Zara, Porsche, Nestle, and HSBC have in common? [00:05:06] Well [00:05:07] They're all European companies [00:05:09] And they're all working [00:05:10] With NVIDIA on AI [00:05:12] Everybody [00:05:15] Everybody [00:05:16] Here [00:05:17] Is all activated [00:05:19] On the AI ecosystem [00:05:20] And working towards [00:05:23] The AI economy [00:05:24] We've been working here in Europe now [00:05:26] For 33 years [00:05:28] You probably don't realize [00:05:30] That [00:05:32] NV1 [00:05:33] NVIDIA's first chip [00:05:34] NV2 [00:05:35] And NV3 [00:05:37] Which became [00:05:38] Riva 128 [00:05:39] The chip that saved our company really [00:05:41] And launched our company [00:05:42] Were all built [00:05:44] In Europe [00:05:45] Our first partner [00:05:47] Was SGS Thompson [00:05:48] We manufactured a chip [00:05:50] In Crow, France [00:05:51] Assembled it in Malta [00:05:53] And shipped it around the world [00:05:55] Our first [00:05:57] Largest customer [00:05:58] The largest customer [00:06:00] Of G-Force days [00:06:01] In the beginning [00:06:02] Was a company called [00:06:05] MediaMark [00:06:06] Do you know them? [00:06:11] MediaMark [00:06:12] Had a spectacular idea [00:06:14] About building PCs [00:06:16] As you remember [00:06:16] And they were [00:06:18] Our largest customer [00:06:19] It was ELSA [00:06:21] A German company [00:06:22] That brought NVIDIA [00:06:23] Into the workstation industry [00:06:25] In the year 2000 [00:06:26] Now NVIDIA [00:06:27] Is in every single workstation [00:06:30] All over Europe [00:06:31] Designing cars [00:06:32] Everything from [00:06:34] Styling [00:06:35] Design [00:06:35] Engineering [00:06:36] Simulation [00:06:37] All the way through [00:06:39] To even [00:06:40] Inside the car [00:06:41] Which a European company [00:06:43] Brought us into [00:06:44] Audi [00:06:45] In the year 2005 [00:06:46] Then in year 2020 [00:06:49] We started working [00:06:51] On autonomous vehicles [00:06:52] With Mercedes [00:06:53] As you could see [00:06:54] NVIDIA's life journey [00:06:56] The entire 33 years [00:06:58] I've spent a lot of time [00:07:00] In Germany [00:07:00] So I'm delighted [00:07:02] To be back [00:07:03] Thank you [00:07:04] I'm delighted [00:07:09] To do GTC here [00:07:10] This year [00:07:12] This last year [00:07:13] We started the new journey [00:07:15] To building a new industry [00:07:17] Here in Europe [00:07:18] The beginnings of the AI economy [00:07:21] In the last year alone [00:07:23] We've invested [00:07:24] Nearly 10 billion dollars [00:07:26] In Europe [00:07:27] Across startups [00:07:28] Big partnerships [00:07:30] Small partnerships [00:07:31] All over Europe [00:07:33] This last week alone [00:07:35] As I journeyed from [00:07:38] France and UK [00:07:40] To Sweden [00:07:41] And now here in Berlin [00:07:42] We announced the investments [00:07:44] Of 14 startup companies [00:07:45] Of 14 startup companies [00:07:46] Overall [00:07:47] Since the beginning [00:07:49] We now have [00:07:51] 12,000 plus [00:07:52] Partnerships across Europe [00:07:54] And we're all working [00:07:56] Towards one goal now [00:07:57] After decades [00:08:01] After decades [00:08:02] Of advancement [00:08:02] In the computer industry [00:08:04] We are now [00:08:05] At the beginnings [00:08:07] Of a new industrial revolution [00:08:09] This image [00:08:10] Was generated by a computer [00:08:12] It was prompted by me [00:08:13] Nice simple prompt [00:08:15] About AI powering [00:08:18] A new industrial revolution [00:08:20] Generated this [00:08:21] I gave it a few partners [00:08:23] That I wanted to highlight [00:08:24] Companies I were working with [00:08:26] And this is an extraordinary thing [00:08:28] In the mind of the AI [00:08:30] This is what it means [00:08:31] To build a new industry [00:08:33] Generation of tokens [00:08:36] The production of intelligence [00:08:39] Which integrates [00:08:41] Into every single industry [00:08:42] Powering a new industrial revolution [00:08:45] A new industrial revolution [00:08:47] Happens [00:08:47] When [00:08:49] The new technology [00:08:51] Affects [00:08:54] How we build something [00:08:55] And what we can build [00:08:58] How we build something [00:09:00] And what we can build [00:09:02] I think in every single way [00:09:04] Artificial intelligence [00:09:06] Will affect everything we build [00:09:08] And in fact [00:09:09] Enable us to build new things [00:09:10] In this way [00:09:12] Because everyone will use it [00:09:14] Every company will be powered by it [00:09:16] Because it's digital intelligence [00:09:19] The foundation of everything [00:09:21] Every single person will use it [00:09:23] Every company will be powered by it [00:09:25] Every industry will be transformed [00:09:27] And new ones will be created [00:09:29] In this way [00:09:31] Like the electricity [00:09:32] It is a new industrial revolution [00:09:35] And we're at the very beginnings of that [00:09:38] The question is [00:09:40] What is AI? [00:09:43] And this is a big deal [00:09:45] If you want to be part of [00:09:47] Creating this future [00:09:49] Understanding [00:09:50] What is AI [00:09:51] Is vitally important [00:09:53] And this is really [00:09:54] A simple framework [00:09:56] We were early to observe [00:09:58] That in fact [00:09:59] AI is a manufacturing industry [00:10:01] AI is not just software technology [00:10:04] It is software technology [00:10:06] But not just [00:10:07] AI [00:10:08] Of course [00:10:09] Runs on a computer [00:10:10] But the way that [00:10:11] That computer operates [00:10:12] Is fundamentally different [00:10:14] Rather than a tool [00:10:16] Rather than a tool [00:10:16] That's used by us [00:10:17] It is now [00:10:18] Part of a machinery [00:10:20] An infrastructure [00:10:21] Factories [00:10:23] That are producing [00:10:24] Intelligence [00:10:25] That are integrated [00:10:27] Into every single industry [00:10:29] And application [00:10:30] A simple model [00:10:32] A simple model [00:10:32] Is to think about [00:10:33] Think of AI [00:10:34] As a five layer cake [00:10:35] Energy [00:10:36] Powers [00:10:37] Advanced computers [00:10:38] That are part of an AI factory [00:10:41] That generates tokens [00:10:43] Digital intelligence [00:10:45] From the models [00:10:46] That run on it [00:10:47] That are integrated [00:10:48] Into every application [00:10:50] A five layer cake [00:10:52] To be part of this [00:10:54] AI economy [00:10:54] You don't want to just be [00:10:56] Consuming [00:10:57] The AI [00:10:58] You want to be [00:10:59] Producing [00:11:00] The AI [00:11:01] We could do all of that [00:11:03] Here in Europe [00:11:04] And in fact [00:11:05] We've already started [00:11:06] Europe is building [00:11:08] Its AI economy [00:11:10] And its [00:11:11] Five layer cake [00:11:12] Represented here [00:11:13] We've got great partnerships [00:11:15] And when you look across [00:11:16] This five layer cake [00:11:17] You can recognize [00:11:18] That in fact [00:11:19] It is a new industry [00:11:20] Because [00:11:20] There's growth [00:11:23] At every single layer [00:11:24] We're creating jobs [00:11:26] Creating growth [00:11:27] At an extraordinary rate [00:11:29] Companies here in Europe [00:11:30] Schneider [00:11:31] Siemens [00:11:32] Electric [00:11:33] Siemens Energy [00:11:35] Doing incredibly well [00:11:37] Around the world [00:11:37] We're the computing layer [00:11:39] We integrate into [00:11:41] AI factories [00:11:42] That are now flourishing [00:11:43] All over Europe [00:11:45] Startup companies like [00:11:47] Nibia is doing incredibly well [00:11:48] A new company [00:11:50] Called Firebird [00:11:51] Which is setting up [00:11:52] A new cloud [00:11:53] In the Central Asia region [00:11:56] And even [00:11:57] Cassava [00:11:58] In Africa [00:11:59] Every single telco [00:12:01] In the future [00:12:02] And I predict this [00:12:04] Every single telco [00:12:05] In the future [00:12:05] With 100,000 sites [00:12:08] Representing some [00:12:09] 100 gigawatts [00:12:10] Of distributed [00:12:11] Telco base stations [00:12:13] And data centers [00:12:14] Will in the future [00:12:16] Will in the future [00:12:16] Evolve into [00:12:17] A distributed [00:12:18] AI network [00:12:19] So the AI infrastructure [00:12:21] Layer is being built out [00:12:23] Which is then [00:12:24] Above that [00:12:25] The AI models [00:12:26] It's really important [00:12:27] To recognize [00:12:28] That long before [00:12:29] The AI models [00:12:30] Emerge [00:12:31] The infrastructure [00:12:32] Has to exist [00:12:33] Just as before [00:12:34] You couldn't have [00:12:35] An internet [00:12:35] You have to have [00:12:36] The networking [00:12:37] Just as before [00:12:39] You could really enjoy [00:12:41] The capabilities [00:12:41] Of electricity [00:12:42] You have to have [00:12:43] Electrified houses [00:12:44] Electrified power plants [00:12:46] And so the models [00:12:48] Are the next layer [00:12:50] It's also interesting [00:12:51] To observe [00:12:52] That the model layer [00:12:53] Is not just [00:12:55] Language models [00:12:56] It represents [00:12:57] Language [00:12:58] Biology [00:12:59] Chemicals [00:13:00] Physics [00:13:01] Robotic actions [00:13:03] Animation [00:13:03] And it represents [00:13:05] Models of all sizes [00:13:06] These different industries [00:13:09] Are going to be powered [00:13:10] By different modalities [00:13:11] Of AI models [00:13:12] Which then integrate [00:13:14] Into the AI natives [00:13:16] The startups [00:13:17] And there are so many [00:13:18] Here in Europe [00:13:18] So the [00:13:20] Europe's AI economy [00:13:22] Is starting [00:13:23] And it's really [00:13:24] Accelerating [00:13:24] It's great to see [00:13:26] At the foundation [00:13:28] Of it [00:13:28] Where we contribute [00:13:29] Most people think [00:13:31] Is GPUs [00:13:33] It is true [00:13:33] We started [00:13:34] In 1993 [00:13:37] Building a graphics chip [00:13:38] Which emerged [00:13:39] Into the GPU [00:13:40] The programmable shading [00:13:41] GPU [00:13:41] Over time [00:13:43] We added CUDA [00:13:44] Acceleration to that [00:13:45] Reinvented computing [00:13:47] As we know it [00:13:48] Which then [00:13:49] Brought forth [00:13:50] Artificial intelligence [00:13:51] But NVIDIA today [00:13:53] Is really [00:13:54] A full stack [00:13:56] End-to-end [00:13:57] Complete [00:13:59] AI factory platform [00:14:00] The breadth of technology [00:14:04] Of our company [00:14:05] Over the years [00:14:06] Has really expanded [00:14:07] Everything from [00:14:08] GPU computing [00:14:09] Vera Rubin [00:14:10] Is our latest generation [00:14:11] To the [00:14:12] NVLink switches [00:14:13] That connect them [00:14:14] Into one giant [00:14:16] GPU [00:14:17] The Vera [00:14:18] Bluefield 4 compute [00:14:20] That we recently [00:14:21] Introduced [00:14:22] And I'll say [00:14:22] A few words [00:14:23] About that later [00:14:23] To Spectrum X [00:14:26] Which is our [00:14:27] Ethernet solution [00:14:28] Making NVIDIA [00:14:29] The largest networking [00:14:30] Company in the world [00:14:30] Today [00:14:31] To Grok 3 [00:14:33] A decode [00:14:34] Accelerator [00:14:35] For [00:14:36] Generation [00:14:37] Text [00:14:38] For token [00:14:38] Generation [00:14:39] To Vera [00:14:40] 4 [00:14:41] Vera Bluefield 4 [00:14:42] Storage Acceleration [00:14:43] With I'll say [00:14:44] A few words [00:14:45] In that [00:14:45] In just a second [00:14:46] And then [00:14:47] Vera RTX [00:14:47] Compute [00:14:48] And I'll mention [00:14:49] That in just a second [00:14:50] But this [00:14:51] Now represents [00:14:53] The breadth [00:14:55] Of technology [00:14:55] Necessary [00:14:56] To create [00:14:57] The AI factory [00:14:58] But nothing [00:15:00] Works [00:15:01] Unless you have [00:15:02] Software that [00:15:03] Activates it all [00:15:03] The first part [00:15:05] Is the system [00:15:06] Software [00:15:06] That runs [00:15:07] The AI factory [00:15:08] These systems [00:15:10] Are gigantic [00:15:10] A 1 gigawatt [00:15:13] Data center [00:15:13] 1 gigawatt [00:15:15] AI factory [00:15:15] Costs about [00:15:16] 50 billion dollars [00:15:18] Land [00:15:19] Power [00:15:20] Shell [00:15:21] All of the [00:15:22] Compute [00:15:23] Networking [00:15:23] Storage [00:15:24] Power [00:15:25] Cooling [00:15:25] All connected [00:15:26] Together [00:15:26] Multiple acres [00:15:28] Of AI factory [00:15:30] Just turning it on [00:15:32] And operating it [00:15:33] Is a marvel [00:15:34] In itself [00:15:34] The magnificent [00:15:36] Amount of software [00:15:38] Necessary [00:15:39] To energize [00:15:40] This factory [00:15:41] To put [00:15:42] It to use [00:15:43] And produce [00:15:44] Tokens [00:15:45] Is an extraordinary [00:15:46] Feat [00:15:46] NVIDIA [00:15:47] Offers [00:15:48] The entire [00:15:48] Full stack [00:15:49] Software [00:15:49] On top of that [00:15:51] Are what we call [00:15:52] The CUDAX libraries [00:15:53] The algorithms [00:15:54] Of our company [00:15:55] What NVIDIA [00:15:56] Truly is [00:15:57] Is an algorithm [00:15:58] Company [00:15:59] That designs [00:16:00] Silicon [00:16:00] To accelerate [00:16:01] Them [00:16:01] Of course [00:16:02] One of them [00:16:03] Is CUDNN [00:16:05] What many [00:16:06] Researchers recognize [00:16:07] As the library [00:16:09] That brought forth [00:16:10] That started [00:16:12] The big bang [00:16:13] Of modern AI [00:16:14] CUDNN [00:16:15] It's [00:16:16] CUDNN [00:16:16] Is to AI [00:16:18] What SQL is [00:16:19] To the world's [00:16:20] Database industry [00:16:21] It's a [00:16:22] Domain specific language [00:16:24] A domain specific library [00:16:25] Just as [00:16:26] OpenGL [00:16:27] And DirectX [00:16:28] Enabled real time [00:16:29] Computer graphics [00:16:30] CUDNN [00:16:32] Enabled [00:16:32] The AI [00:16:34] Computing industry [00:16:34] 350 other [00:16:37] Libraries [00:16:38] Those libraries [00:16:40] Are in fact [00:16:41] The crown jewels [00:16:42] Of our company [00:16:43] Everything from [00:16:44] Computational lithography [00:16:45] To signal processing [00:16:47] Image processing [00:16:48] Data processing [00:16:49] Particles [00:16:50] Fluids [00:16:51] Sparse [00:16:52] Solvers [00:16:53] Incredible [00:16:54] Breath of libraries [00:16:56] That many of [00:16:57] You in the audience [00:16:58] Have connected [00:16:59] Into your company [00:17:00] This is how [00:17:01] We open new markets [00:17:02] This is how [00:17:03] We connect [00:17:05] NVIDIA [00:17:05] Into algorithms [00:17:07] That enable you [00:17:08] To do computation [00:17:09] At incredible speeds [00:17:11] And then finally [00:17:12] On the right hand side [00:17:14] On that side [00:17:15] The models [00:17:18] This is the new software [00:17:21] The new engine [00:17:23] Of new applications [00:17:24] And it sits on top of [00:17:26] And it collaborates [00:17:27] With the CUDAX libraries [00:17:28] These models [00:17:29] NVIDIA have been working on [00:17:31] Across multiple domains [00:17:32] With these models [00:17:34] It opens up new opportunities [00:17:35] And I have lots to say [00:17:36] About that later on [00:17:37] Well, every single generation [00:17:40] NVIDIA started with [00:17:42] Pascal [00:17:44] The AI [00:17:45] Era started with [00:17:48] The Pascal [00:17:48] New [00:17:49] What is called [00:17:51] DGX1 [00:17:52] System [00:17:53] DGX1 [00:17:54] Brought to the world [00:17:57] The world's first AI supercomputer [00:17:59] I introduced it [00:18:00] 10 years ago [00:18:02] Delivered the first one [00:18:04] To open AI [00:18:05] Started their computation journey [00:18:08] From Pascal [00:18:10] To Volta [00:18:12] To Ampere [00:18:13] Then to many of you [00:18:16] Who heard about us [00:18:16] For the very first time [00:18:17] Hopper [00:18:18] The current generation [00:18:20] Blackwell [00:18:21] And then now [00:18:22] Vera Rubin [00:18:23] In full production [00:18:24] Notice six generations [00:18:26] Of architectures [00:18:27] With each generation [00:18:29] We advanced the technology [00:18:31] In the course [00:18:34] Of that last 10 years [00:18:35] We've improved [00:18:35] AI performance [00:18:36] By 1 million times [00:18:38] If you would take a step back [00:18:40] And think about [00:18:41] What would Moore's law [00:18:42] Have done [00:18:42] In 10 years time [00:18:44] It would have [00:18:46] The first five years [00:18:47] Would have improved [00:18:47] By a factor of 10 [00:18:48] The second five years [00:18:50] By another factor of 10 [00:18:51] 100x improvement [00:18:53] In 10 years time [00:18:54] In the same period of time [00:18:57] Because of the way [00:18:58] NVIDIA Extreme [00:18:59] Co-designs [00:19:00] Were innovating [00:19:01] Across every single computer [00:19:03] Across every single stack [00:19:05] Of the software [00:19:06] Optimizing new algorithms [00:19:08] Creating new algorithms [00:19:08] While creating architectures [00:19:10] To accelerate it [00:19:11] Working across [00:19:12] That entire domain [00:19:14] What we call [00:19:15] Extreme co-design [00:19:16] Where we're able [00:19:17] To accelerate AI [00:19:18] By a million times [00:19:20] And we're going to [00:19:21] Continue to do so [00:19:22] This is what NVIDIA [00:19:25] Builds today [00:19:26] And this is the foundation [00:19:27] By which we can enable [00:19:29] Europe to be part [00:19:31] Of the AI ecosystem [00:19:33] The AI economy [00:19:34] Now each generation [00:19:36] In the beginning [00:19:38] Pascal was GPUs [00:19:40] Volta was GPUs [00:19:42] Ampere was GPUs [00:19:44] And Hopper was GPUs [00:19:45] We invented the new [00:19:47] DGX form factor [00:19:49] Then it was Blackwell [00:19:51] That introduced [00:19:52] To the world [00:19:52] The first [00:19:53] Rack scale computing [00:19:55] And I'll show you that [00:19:55] In just a moment [00:19:56] Rack scale computing [00:19:58] Turned one entire system [00:20:02] Into one giant computer [00:20:04] One GPU [00:20:05] And now with Vera Rubin [00:20:08] We're extending it [00:20:09] To processing beyond models [00:20:11] Beyond large language models [00:20:13] To now processing [00:20:14] The entire agent [00:20:15] Now [00:20:16] We'll refer back to this slide [00:20:19] Later [00:20:20] But remember this [00:20:23] The agent [00:20:24] Has to think [00:20:25] The agent has to think [00:20:28] And that's where the model [00:20:31] Comes in [00:20:31] Perception [00:20:32] Reasoning [00:20:34] Planning [00:20:34] And using tools [00:20:36] All of that runs [00:20:38] On the large language model [00:20:39] And the models [00:20:40] Are designed to do that [00:20:41] But the agent [00:20:43] Also [00:20:44] Has what is called [00:20:45] A harness [00:20:46] Basically [00:20:47] This exoskeleton [00:20:49] Of [00:20:49] The digital agent [00:20:52] Just last [00:20:53] We have the brain [00:20:53] The rest of our body [00:20:55] Represents [00:20:55] Our exoskeleton [00:20:57] The agent [00:20:58] Is the exoskeleton [00:20:59] That orchestrates [00:21:00] The work [00:21:01] That allows [00:21:02] The agent [00:21:03] To actually engage [00:21:04] Memory [00:21:05] Active memory [00:21:07] Working memory [00:21:08] As well as long-term memory [00:21:09] And activating tools [00:21:10] Then the third part [00:21:12] Is [00:21:13] A computer [00:21:14] That processes [00:21:15] Knowledge [00:21:16] Designed [00:21:16] For data processing [00:21:18] And then lastly [00:21:19] A computer [00:21:20] Dedicated to [00:21:21] CudaX libraries [00:21:22] And its job [00:21:24] Is to run tools [00:21:25] And these computers [00:21:26] Now represent [00:21:28] The entire agentic system [00:21:30] So [00:21:31] Verrubin [00:21:32] Is in full production [00:21:33] It's nice to kind of see it [00:21:34] This [00:21:35] Is the first computer [00:21:38] In the world [00:21:38] That we've created [00:21:39] No cables [00:21:40] No wires [00:21:42] Completely 100% [00:21:44] Liquid cooled [00:21:44] And it's just really delightful [00:21:46] Whoops [00:21:47] Hang on [00:21:49] Really delightful [00:21:51] To be able to pull it out [00:21:52] And this [00:21:54] Is one [00:21:55] Verrubin computer [00:21:57] Okay [00:21:57] So this has [00:21:58] Two GPUs [00:21:59] Yep [00:22:00] What used to take [00:22:08] In the case of [00:22:09] Grace Blackwell [00:22:10] Two hours [00:22:11] To assemble [00:22:12] One node [00:22:12] And when you see it [00:22:14] You'll know what I mean [00:22:15] Cables and wires [00:22:17] Everywhere [00:22:17] 80% liquid cooled [00:22:19] Now Verrubin [00:22:20] Is 100% liquid cooled [00:22:22] And it takes [00:22:22] Less than five minutes [00:22:23] To assemble it [00:22:24] And so [00:22:25] Incredible [00:22:26] Incredible achievement [00:22:27] Of course [00:22:28] This translates [00:22:29] To higher cycle times [00:22:30] In manufacturing [00:22:31] More [00:22:32] Better [00:22:32] Much much better [00:22:33] Reliability [00:22:34] And of course [00:22:35] A lot better performance [00:22:36] This [00:22:38] Is the [00:22:40] NVLink [00:22:40] Switch tray [00:22:41] And what this does [00:22:43] Is it connects [00:22:44] 72 [00:22:45] Of these GPUs [00:22:47] There are four of them [00:22:48] In here right now [00:22:48] It connects 72 GPUs [00:22:51] Into one gigantic GPU [00:22:53] Okay [00:22:54] So each GPU [00:22:55] Is 20 petaflops [00:22:56] Connecting 72 of them [00:22:59] Into one gigantic system [00:23:00] Okay [00:23:04] Thank you [00:23:05] Thank you [00:23:08] And this [00:23:12] Is our new [00:23:14] Vera computer [00:23:14] It's okay [00:23:27] Go easy [00:23:29] Let her [00:23:30] Give her a chance [00:23:31] Everybody has a chance [00:23:33] To say something [00:23:34] But today [00:23:35] Let me [00:23:36] No no [00:23:45] Take [00:23:45] Just [00:23:46] I'll talk to you later [00:23:47] I'll talk to you guys later [00:23:49] Thank you [00:23:54] All important issues [00:23:59] All important issues [00:24:01] But let me get through [00:24:02] My keynote [00:24:03] We have [00:24:06] We have [00:24:06] So this [00:24:08] Auditorium is packed [00:24:10] We have [00:24:11] A hundred [00:24:12] Other watch parties [00:24:14] Around the world [00:24:15] A hundred watch parties [00:24:17] Fifteen thousand people [00:24:18] Are watching us right now [00:24:19] And so [00:24:20] I'll get through [00:24:21] The keynote [00:24:21] There's some really [00:24:22] Really fun things [00:24:23] I want to show you [00:24:24] In the end [00:24:25] Okay [00:24:25] So you just [00:24:26] You just have to [00:24:26] You have to [00:24:27] I have to torture you [00:24:29] Through some education [00:24:29] First [00:24:30] And then the [00:24:31] Entertainment is coming [00:24:32] This [00:24:34] This is our brand new [00:24:35] This is our brand new [00:24:36] Vera computer [00:24:37] Vera CPU [00:24:38] Is an extraordinary [00:24:39] CPU beast [00:24:41] Eight [00:24:42] Instructions per clock [00:24:44] It has [00:24:46] The world's first [00:24:47] LPDDR memory [00:24:48] Memory system [00:24:49] So that we can have [00:24:51] High performance [00:24:51] Very high bandwidth [00:24:52] As well as [00:24:53] Extremely excellent [00:24:55] Energy efficiency [00:24:56] Efficiency [00:24:56] The on-chip fabric [00:24:59] Is the fastest [00:25:00] Of any chip [00:25:01] In the world [00:25:02] As a result [00:25:02] Data processing [00:25:03] Is off the charts [00:25:05] Agentic [00:25:06] Sessions starting [00:25:08] And processing [00:25:09] Is off the charts [00:25:11] And this is [00:25:11] Designed for [00:25:12] Doing exactly that [00:25:14] This is our [00:25:14] Vera computer [00:25:15] And this computer [00:25:16] Here [00:25:16] This computer here [00:25:18] Is the [00:25:20] Spectrum X [00:25:21] This is the [00:25:22] World's first [00:25:23] Co-packaged [00:25:25] Optics [00:25:25] And you can [00:25:26] See it here [00:25:27] I'll show you [00:25:28] Some of this [00:25:32] Some of this [00:25:33] Area [00:25:33] Falls into the [00:25:34] Ground [00:25:35] And some of this [00:25:36] Area you can [00:25:37] Walk on [00:25:37] I'm just not [00:25:39] Sure which [00:25:40] Alright so [00:25:44] So if I [00:25:46] Accidentally [00:25:47] Disappear [00:25:48] You guys know [00:25:49] You guys know [00:25:50] Us [00:25:50] It's not a [00:25:52] Stunt [00:25:53] It's an accident [00:25:54] And so [00:25:55] You can [00:25:56] See here [00:25:57] These are [00:25:58] The lasers [00:25:59] And these lasers [00:26:01] Come up through [00:26:02] Optics [00:26:03] And it [00:26:03] Directly connects [00:26:04] Directly connects [00:26:06] To our chip [00:26:07] And it connects [00:26:08] Using a process [00:26:09] Technology [00:26:09] We invented [00:26:10] With TSMC [00:26:11] Called coop [00:26:12] And so silicon [00:26:13] And so silicon [00:26:13] Photonics [00:26:14] Are directly [00:26:16] Attached [00:26:16] Right onto [00:26:17] Our chip [00:26:18] Otherwise [00:26:19] What you would [00:26:19] See is [00:26:20] All of this [00:26:21] Would be [00:26:21] PC board [00:26:22] And PC [00:26:24] Boards [00:26:24] Transmission [00:26:25] Capability [00:26:26] Is much [00:26:27] Worse [00:26:28] Than optical [00:26:28] Cables [00:26:29] And so [00:26:30] And so [00:26:30] We now [00:26:30] Have extremely [00:26:31] High data [00:26:32] Rate [00:26:32] Extremely [00:26:33] High quality [00:26:34] Data [00:26:34] That comes [00:26:35] Directly [00:26:35] Onto [00:26:36] The chip [00:26:36] And this [00:26:37] Goes directly [00:26:38] Into [00:26:38] Otherwise [00:26:39] Transceivers [00:26:40] Without [00:26:41] Going through [00:26:43] Without going [00:26:43] Through a PCB [00:26:44] And a transceiver [00:26:45] Goes right [00:26:46] Into the optics [00:26:46] Okay [00:26:47] So this [00:26:48] Is the [00:26:48] World's [00:26:48] First [00:26:49] CPO [00:26:50] We're [00:26:50] In [00:26:50] Complete [00:26:51] Volume [00:26:51] Production [00:26:52] And [00:26:53] 512 [00:26:55] Port [00:26:55] At 200 [00:26:57] Gigabits [00:26:57] Per second [00:26:58] The world's [00:26:59] Most advanced [00:26:59] Ethernet [00:27:00] Switch [00:27:00] And we're [00:27:01] Really proud [00:27:02] Of it [00:27:02] Thank you [00:27:02] And so [00:27:09] So what [00:27:10] That [00:27:10] What that [00:27:11] Gives us [00:27:11] What that [00:27:12] Gives us [00:27:13] Is the [00:27:13] Ability [00:27:13] To [00:27:14] Process [00:27:15] The thoughts [00:27:16] The large [00:27:17] Language [00:27:17] Model [00:27:17] Process [00:27:19] The agent [00:27:19] Harness [00:27:20] Run [00:27:22] Data [00:27:23] Processing [00:27:24] For its [00:27:24] Knowledge [00:27:25] And then [00:27:25] Run [00:27:26] Accelerated [00:27:27] Computing [00:27:27] For all [00:27:28] Of the [00:27:28] Tools [00:27:28] That [00:27:29] Represents [00:27:30] Essentially [00:27:31] The agent [00:27:32] Workload [00:27:32] And I'll [00:27:33] Show you [00:27:33] More [00:27:33] In a [00:27:33] Second [00:27:34] While [00:27:34] Very [00:27:34] Rubin [00:27:35] Is in [00:27:35] Full [00:27:35] Production [00:27:36] And the [00:27:37] Performance [00:27:37] Is off [00:27:38] The charts [00:27:38] Every [00:27:39] Single [00:27:39] Generation [00:27:40] As I [00:27:40] Mentioned [00:27:41] Through [00:27:41] Extreme [00:27:42] Code [00:27:42] Design [00:27:43] We're [00:27:43] Able [00:27:43] To [00:27:43] Achieve [00:27:44] Performances [00:27:44] That are [00:27:45] Well [00:27:45] Beyond [00:27:46] Reason [00:27:46] Well [00:27:47] Beyond [00:27:47] What [00:27:47] Moore's [00:27:48] Law [00:27:48] Would [00:27:48] Have [00:27:48] Dictated [00:27:49] And as [00:27:50] You [00:27:50] Know [00:27:50] Moore's [00:27:51] Law [00:27:51] Has [00:27:51] Really [00:27:51] Largely [00:27:52] Ended [00:27:52] And so [00:27:53] We [00:27:53] Need [00:27:53] These [00:27:54] Extreme [00:27:55] Co-design [00:27:55] Ideas [00:27:56] Working [00:27:56] From [00:27:57] Chips [00:27:57] To [00:27:58] Systems [00:27:58] Systems [00:27:59] To [00:27:59] System [00:27:59] Software [00:28:00] To [00:28:00] Algorithm [00:28:01] To [00:28:01] The [00:28:01] The [00:28:02] Air [00:28:03] Fair [00:28:03] Fair [00:28:03] And so [00:28:04] Varirubin [00:28:04] Is [00:28:05] Another [00:28:05] Giant [00:28:06] Step [00:28:06] On [00:28:08] Top [00:28:08] Of [00:28:08] Blackwell [00:28:08] Blackwell [00:28:10] Was [00:28:10] About [00:28:10] 30 [00:28:11] Times [00:28:11] Faster [00:28:12] Than [00:28:12] Hopper [00:28:13] In [00:28:14] Doing [00:28:14] Large [00:28:16] Language [00:28:16] Model [00:28:16] Processing [00:28:17] And now [00:28:18] Varirubin [00:28:20] Is [00:28:20] Another [00:28:20] 30 [00:28:21] Times [00:28:21] Greater [00:28:22] Than [00:28:23] Blackwell [00:28:24] In [00:28:24] Just [00:28:24] Two [00:28:25] Years [00:28:25] Time [00:28:25] Let's [00:28:26] Take [00:28:26] A [00:28:26] Look [00:28:27] At [00:28:27] How [00:28:27] Varirubin [00:28:28] Turns [00:28:29] Into [00:28:29] What [00:28:29] Is [00:28:30] A [00:28:30] DSX [00:28:31] Full [00:28:32] Stack [00:28:32] AI [00:28:32] Factory [00:28:33] AI [00:28:38] Is [00:28:39] Creating [00:28:39] The [00:28:40] Largest [00:28:40] Infrastructure [00:28:41] Build [00:28:41] Out [00:28:42] In [00:28:42] History [00:28:42] The [00:28:43] Factories [00:28:43] Are [00:28:43] Massive [00:28:44] The [00:28:45] Investments [00:28:46] Are [00:28:46] Significant [00:28:47] We [00:28:47] Have [00:28:48] To [00:28:48] Build [00:28:48] It [00:28:48] Right [00:28:48] But [00:28:49] Great [00:28:50] Chips [00:28:50] Alone [00:28:50] Don't [00:28:51] Make [00:28:51] A [00:28:51] Great [00:28:51] AI [00:28:51] Factory [00:28:52] NVIDIA [00:28:53] DSX [00:28:54] Is [00:28:54] A [00:28:54] Full [00:28:55] Stack [00:28:55] AI [00:28:55] Factory [00:28:56] Reference [00:28:56] Architecture [00:28:57] Co-innovated [00:28:58] With [00:28:59] The [00:28:59] World's [00:28:59] Leading [00:28:59] Infrastructure [00:29:00] Companies [00:29:00] Compute [00:29:02] Networking [00:29:03] Storage [00:29:03] Software [00:29:04] Power [00:29:05] And [00:29:05] Cooling [00:29:05] Co-designed [00:29:06] As [00:29:07] One [00:29:07] System [00:29:07] Optimized [00:29:09] From [00:29:09] Design [00:29:09] To [00:29:09] Operation [00:29:10] DSX [00:29:12] Starts [00:29:12] In [00:29:12] A [00:29:13] Digital [00:29:13] Twin [00:29:13] Performance [00:29:15] Power [00:29:16] Water [00:29:16] And [00:29:17] Scale [00:29:17] Validated [00:29:18] Virtually [00:29:18] Before [00:29:19] Construction [00:29:19] Begins [00:29:20] PTC [00:29:22] Turns [00:29:22] Equipment [00:29:22] Designs [00:29:23] Into [00:29:23] Simulation [00:29:24] Ready [00:29:24] Assets [00:29:25] Dasso [00:29:26] Systems [00:29:27] Connects [00:29:27] Rack [00:29:27] And [00:29:28] Facility [00:29:28] Design [00:29:29] Through [00:29:29] Model [00:29:29] Based [00:29:30] Systems [00:29:30] Engineering [00:29:31] Schneider [00:29:32] Scheider Electric [00:29:33] Simulates [00:29:33] Power [00:29:34] Distribution [00:29:34] And [00:29:34] Liquid [00:29:35] Cooling [00:29:35] Jacobs [00:29:37] Brings [00:29:37] It [00:29:37] All [00:29:37] Together [00:29:38] In [00:29:38] The [00:29:38] Omniverse [00:29:39] Digital [00:29:39] Twin [00:29:39] Siemens [00:29:41] Evaluates [00:29:42] Thermal [00:29:42] Acoustic [00:29:43] And [00:29:44] Grid [00:29:44] Impacts [00:29:44] And [00:29:46] Procore [00:29:46] Supports [00:29:47] Construction [00:29:48] Phase [00:29:48] Planning [00:29:48] Power [00:29:50] Limits [00:29:51] How [00:29:51] Much [00:29:51] An [00:29:51] AI [00:29:51] Factory [00:29:52] Can [00:29:52] Produce [00:29:53] DSX [00:29:54] Max [00:29:55] LPS [00:29:55] Is [00:29:56] An [00:29:56] End [00:29:56] To [00:29:56] End [00:29:56] Control [00:29:57] Loop [00:29:57] That [00:29:58] Dynamically [00:29:58] Steers [00:29:59] Power [00:29:59] To [00:29:59] Match [00:30:00] Workload [00:30:00] Demand [00:30:00] It [00:30:01] Uses [00:30:01] 30% [00:30:02] Less [00:30:02] Power [00:30:03] For [00:30:03] The [00:30:03] Same [00:30:03] Output [00:30:04] DSX [00:30:06] Rethinks [00:30:07] Cooling [00:30:07] NVIDIA [00:30:08] Vera [00:30:09] Rubin [00:30:09] Systems [00:30:10] Are [00:30:10] Designed [00:30:10] For [00:30:11] Closed [00:30:11] Loop [00:30:11] Liquid [00:30:12] Cooling [00:30:12] With [00:30:13] Coolant [00:30:13] Temperatures [00:30:13] Up [00:30:14] To [00:30:14] 45 [00:30:14] Degrees [00:30:15] Celsius [00:30:15] In [00:30:17] Suitable [00:30:17] Climates [00:30:18] Outdoor [00:30:19] Dry [00:30:19] Coolers [00:30:19] Release [00:30:20] Heat [00:30:20] Into [00:30:20] The [00:30:21] Air [00:30:21] Without [00:30:21] Evaporating [00:30:22] Water [00:30:23] At [00:30:24] 2 [00:30:24] Gigawatts [00:30:25] DSX [00:30:26] Can [00:30:26] Save [00:30:26] 54 [00:30:27] Million [00:30:27] Liters [00:30:28] Of [00:30:28] Water [00:30:28] A [00:30:28] Day [00:30:29] That's [00:30:29] The [00:30:30] Daily [00:30:30] Water [00:30:30] Use [00:30:30] Of [00:30:31] 180,000 [00:30:32] Households [00:30:33] But [00:30:34] An [00:30:35] AI [00:30:35] Factory [00:30:35] Can [00:30:36] Take [00:30:36] Several [00:30:36] Months [00:30:37] To [00:30:37] Build [00:30:37] Until [00:30:38] It [00:30:38] Is [00:30:38] Operational [00:30:39] It [00:30:40] Produces [00:30:40] No [00:30:40] Tokens [00:30:41] Today [00:30:42] We [00:30:43] Are [00:30:43] Announcing [00:30:44] NVIDIA [00:30:44] DSX [00:30:45] Pod [00:30:45] Bringing [00:30:46] This [00:30:46] Full [00:30:47] Stack [00:30:47] Platform [00:30:47] Into [00:30:48] A [00:30:48] Factory [00:30:48] Built [00:30:49] Integrated [00:30:50] System [00:30:50] It [00:30:51] Starts [00:30:52] With [00:30:52] Vera [00:30:52] Rubin [00:30:53] NV [00:30:55] 7 [00:30:56] Liquid [00:30:56] Cooled [00:30:57] Compute [00:30:57] Racks [00:30:57] Connected [00:30:58] With [00:30:58] 17 [00:30:59] Kilometers [00:30:59] Of [00:30:59] Spectrum [00:31:00] X [00:31:00] Cables [00:31:01] Form [00:31:01] A [00:31:01] Vera [00:31:01] Rubin [00:31:02] DSX [00:31:03] Pod [00:31:03] Vera [00:31:05] Rubin [00:31:05] DSX [00:31:06] Pods [00:31:06] Are [00:31:06] Integrated [00:31:07] Fully [00:31:08] Cabled [00:31:08] And [00:31:09] Factory [00:31:09] Tested [00:31:09] Then [00:31:10] Delivered [00:31:11] In [00:31:11] Transportable [00:31:11] Containers [00:31:12] Dry [00:31:14] Coolers [00:31:14] Replace [00:31:15] Traditional [00:31:15] Evaporative [00:31:16] Cooling [00:31:16] 800 [00:31:18] Volts [00:31:18] DC [00:31:19] Connects [00:31:19] To [00:31:19] The [00:31:20] Racks [00:31:20] For [00:31:20] Efficient [00:31:20] Power [00:31:21] Delivery [00:31:21] Once [00:31:23] Site [00:31:23] Utilities [00:31:23] Are [00:31:24] Ready [00:31:24] Estimated [00:31:25] Onsite [00:31:25] Deployment [00:31:26] Goes [00:31:26] From [00:31:27] Several [00:31:27] Months [00:31:27] To [00:31:27] Days [00:31:28] NVIDIA [00:31:29] DSX [00:31:30] AI [00:31:30] Factory [00:31:31] Is [00:31:31] Responsible [00:31:32] By [00:31:32] Design [00:31:33] And [00:31:34] Now [00:31:34] With [00:31:35] DSX [00:31:35] Pod [00:31:36] AI [00:31:37] Factories [00:31:37] Can [00:31:37] Be [00:31:37] Built [00:31:38] At [00:31:38] The [00:31:38] Speed [00:31:38] Of [00:31:38] Light [00:31:39] More [00:31:40] Intelligence [00:31:41] Per [00:31:41] Watt [00:31:41] Less [00:31:42] Time [00:31:43] To [00:31:43] First [00:31:43] Token [00:31:44] What do you [00:31:50] think? [00:32:00] So [00:32:01] In 2016 [00:32:03] We introduced [00:32:04] The world's [00:32:04] First [00:32:05] AI [00:32:05] Supercomputer [00:32:06] It was [00:32:07] In [00:32:07] One [00:32:07] Node [00:32:07] One [00:32:08] GPU [00:32:09] Node [00:32:09] Then [00:32:10] Three [00:32:11] Years [00:32:11] Ago [00:32:11] We [00:32:12] Introduced [00:32:12] The [00:32:14] Grace [00:32:14] Blackwell [00:32:15] NVLink [00:32:16] 72 [00:32:16] Rack [00:32:17] Scale [00:32:18] Computer [00:32:18] Both [00:32:20] Times [00:32:21] We [00:32:21] Had [00:32:21] To [00:32:21] Reinvent [00:32:22] Not [00:32:23] Just [00:32:23] The [00:32:23] Computer [00:32:24] The [00:32:24] Software [00:32:25] And [00:32:26] How [00:32:26] It [00:32:26] Works [00:32:27] But [00:32:27] Also [00:32:27] The [00:32:27] Supply [00:32:28] Chain [00:32:28] In [00:32:30] The [00:32:30] Case [00:32:30] Of [00:32:30] NVLink [00:32:31] 72 [00:32:32] We [00:32:33] Took [00:32:33] The [00:32:33] Entire [00:32:34] Rack [00:32:35] Which [00:32:35] Would [00:32:35] Have [00:32:36] Otherwise [00:32:36] Been [00:32:36] Assembled [00:32:37] On [00:32:38] Site [00:32:38] And [00:32:39] We [00:32:39] Assembled [00:32:40] It [00:32:40] Inside [00:32:41] The [00:32:41] Supply [00:32:41] Chain [00:32:41] We [00:32:42] Essentially [00:32:43] Manufactured [00:32:44] Supercomputers [00:32:46] At [00:32:46] Scale [00:32:47] It [00:32:47] Is [00:32:48] The [00:32:48] Reason [00:32:48] Why [00:32:48] We [00:32:48] Are [00:32:48] Able [00:32:49] To [00:32:49] Move [00:32:49] So [00:32:49] Fast [00:32:50] And [00:32:51] Be [00:32:51] Able [00:32:51] To [00:32:51] Provide [00:32:52] For [00:32:52] Literally [00:32:53] Several [00:32:53] Hundred [00:32:54] Data [00:32:55] Center [00:32:55] Buildouts [00:32:56] Around [00:32:56] The [00:32:56] World [00:32:56] At [00:32:57] The [00:32:57] Same [00:32:57] Time [00:32:57] Every [00:32:58] Single [00:32:59] Computer [00:32:59] Manufacturer [00:33:00] In [00:33:00] The [00:33:00] World [00:33:00] Now [00:33:01] Manufactures [00:33:02] NVLink [00:33:02] 72 [00:33:03] The [00:33:04] Scale [00:33:05] At [00:33:05] Which [00:33:05] We [00:33:06] Are [00:33:06] Now [00:33:06] Producing [00:33:06] Computing [00:33:07] Is [00:33:07] Unheard [00:33:08] Of [00:33:08] Ever [00:33:09] Unimaginable [00:33:10] Well [00:33:11] Even [00:33:12] Then [00:33:13] When [00:33:13] You [00:33:13] Bring [00:33:14] These [00:33:14] Systems [00:33:15] Onto [00:33:15] Your [00:33:16] Sites [00:33:17] There [00:33:19] Are [00:33:19] Still [00:33:19] Months [00:33:20] And [00:33:20] Months [00:33:20] Of [00:33:20] Assembly [00:33:21] And [00:33:21] You [00:33:21] Just [00:33:22] Heard [00:33:22] 17 [00:33:23] Thousand [00:33:24] Kilometers [00:33:25] Of [00:33:26] Wiring [00:33:27] To [00:33:27] Connect [00:33:27] Just [00:33:28] Spectrum [00:33:29] X [00:33:29] For [00:33:30] The [00:33:30] Scale [00:33:31] Out [00:33:31] Computing [00:33:32] 17 [00:33:34] Kilometers [00:33:35] 17 [00:33:35] Thousand [00:33:36] Kilometers [00:33:36] It's [00:33:37] Crazy [00:33:38] Number [00:33:38] And [00:33:38] So [00:33:39] 17 [00:33:39] Kilometers [00:33:40] Incredible [00:33:40] Numbers [00:33:41] Now [00:33:41] Suppose [00:33:42] We [00:33:42] Were [00:33:42] To [00:33:43] Do [00:33:43] That [00:33:43] All [00:33:43] Inside [00:33:44] The [00:33:45] Supply [00:33:45] Chain [00:33:45] Run [00:33:46] Everything [00:33:46] Up [00:33:47] Build [00:33:47] Everything [00:33:47] Up [00:33:48] Make [00:33:48] Sure [00:33:48] That [00:33:48] It [00:33:48] Runs [00:33:49] Completely [00:33:49] Get [00:33:50] It [00:33:50] Qualified [00:33:51] Burnt [00:33:51] In [00:33:51] And [00:33:53] By [00:33:53] The [00:33:53] Time [00:33:53] We [00:33:53] Ship [00:33:54] It [00:33:54] It [00:33:55] Is [00:33:55] Fully [00:33:55] Functional [00:33:55] The [00:33:56] Amount [00:33:57] Of [00:33:57] Cabling [00:33:57] And [00:33:58] On [00:33:59] Site [00:33:59] Assembly [00:33:59] Is [00:34:00] Going [00:34:00] To [00:34:00] Be [00:34:00] Greatly [00:34:01] Reduced [00:34:01] But [00:34:02] There's [00:34:03] Something [00:34:03] Much [00:34:04] Much [00:34:04] More [00:34:04] In [00:34:06] Fact [00:34:06] What's [00:34:06] Going [00:34:06] To [00:34:06] Happen [00:34:07] Here [00:34:07] Is [00:34:08] That [00:34:08] By [00:34:08] Moving [00:34:08] This [00:34:09] Into [00:34:09] The [00:34:09] Supply [00:34:10] Chain [00:34:10] The [00:34:12] Amount [00:34:12] Of [00:34:12] Idle [00:34:13] Computing [00:34:14] The [00:34:14] Amount [00:34:14] Of [00:34:14] Idle [00:34:15] Computing [00:34:15] That's [00:34:15] Going [00:34:16] To [00:34:16] Be [00:34:16] Sitting [00:34:19] Backly [00:34:19] Remember [00:34:20] This [00:34:21] Each [00:34:22] Gigawatt [00:34:23] Each [00:34:24] Gigawatt [00:34:25] Data [00:34:25] Center [00:34:25] Is [00:34:26] 50 [00:34:26] Billion [00:34:27] Dollars [00:34:28] The [00:34:29] Revenue [00:34:29] Generation [00:34:30] You [00:34:31] Expect [00:34:31] Is [00:34:33] Tens [00:34:33] Of [00:34:34] Billion [00:34:34] Dollars [00:34:34] Per [00:34:35] Year [00:34:35] And [00:34:37] So [00:34:37] If [00:34:37] We [00:34:37] Just [00:34:37] Said [00:34:37] It's [00:34:39] 40 [00:34:39] Billion [00:34:40] Dollars [00:34:40] Per [00:34:40] Year [00:34:41] Of [00:34:41] Revenue [00:34:42] Generation [00:34:43] That [00:34:43] Comes [00:34:44] Out [00:34:44] Of [00:34:44] That [00:34:44] Infrastructure [00:34:45] 40 [00:34:46] Billion [00:34:47] If [00:34:48] Delayed [00:34:49] By [00:34:49] Just [00:34:49] Three [00:34:50] Months [00:34:51] That's [00:34:52] 5 [00:34:53] Billion [00:34:53] Dollars [00:34:53] Excuse [00:34:54] Me [00:34:54] What [00:34:54] Is [00:34:54] That [00:34:54] 10 [00:34:55] Billion [00:34:55] Dollars [00:34:55] Worth [00:34:55] Of [00:34:56] Revenues [00:34:56] 10 [00:34:58] Billion [00:34:58] Dollars [00:34:58] Worth [00:34:58] Of [00:34:59] Revenues [00:34:59] Will [00:35:00] Make [00:35:00] Anybody [00:35:00] Anxious [00:35:01] And [00:35:02] So [00:35:03] By [00:35:03] Moving [00:35:04] This [00:35:05] Into [00:35:05] The [00:35:05] Supply [00:35:06] Chain [00:35:06] The [00:35:07] Speed [00:35:07] By [00:35:08] Which [00:35:08] We [00:35:08] Can [00:35:08] Turn [00:35:08] On [00:35:09] Go [00:35:09] From [00:35:09] Installation [00:35:11] Arrival [00:35:11] Of [00:35:12] Computing [00:35:12] To [00:35:13] Production [00:35:13] Of [00:35:13] Tokens [00:35:14] Will [00:35:15] Go [00:35:15] From [00:35:15] Many [00:35:15] Months [00:35:16] We [00:35:17] Think [00:35:18] Something [00:35:18] Along [00:35:18] The [00:35:18] Lines [00:35:19] Of [00:35:19] Four [00:35:19] To [00:35:19] Five [00:35:19] Months [00:35:20] Down [00:35:20] To [00:35:20] About [00:35:20] A [00:35:20] Month [00:35:21] Okay [00:35:21] So [00:35:22] This [00:35:22] Is [00:35:22] A [00:35:23] Complete [00:35:23] Transformation [00:35:24] Of [00:35:24] The [00:35:24] World [00:35:25] Supply [00:35:25] Chain [00:35:25] And [00:35:26] This [00:35:26] Will [00:35:26] Change [00:35:27] The [00:35:27] Way [00:35:27] That [00:35:27] People [00:35:27] Install [00:35:28] And [00:35:29] Build [00:35:29] Data [00:35:29] Centers [00:35:30] And [00:35:30] Hopefully [00:35:30] As a [00:35:31] Result [00:35:31] Velocity [00:35:32] Goes [00:35:32] Up [00:35:33] The [00:35:33] Wasted [00:35:34] Idle [00:35:35] Compute [00:35:35] The [00:35:36] Wasted [00:35:36] Idle [00:35:37] Cap [00:35:37] X [00:35:37] Will [00:35:38] Reduce [00:35:38] And [00:35:39] Everybody [00:35:40] Becomes [00:35:40] More [00:35:44] NVLink 72 [00:35:45] Now [00:35:46] Into [00:35:47] A [00:35:47] Large [00:35:48] Cabinet [00:35:48] System [00:35:49] We [00:35:49] Call [00:35:49] It [00:35:49] DSX [00:35:50] Pod [00:35:51] Okay [00:35:52] So [00:35:52] This [00:35:52] Is [00:35:53] The [00:35:53] Next [00:35:53] Click [00:35:54] Of [00:35:54] NVIDIA [00:35:54] Computing [00:35:55] Now [00:36:04] What [00:36:06] You [00:36:06] Make [00:36:06] And [00:36:08] How [00:36:08] You [00:36:08] Make [00:36:08] It [00:36:09] Makes [00:36:10] You [00:36:10] Special [00:36:10] Just [00:36:12] As [00:36:12] Here [00:36:12] In [00:36:12] Europe [00:36:13] You [00:36:14] The [00:36:15] Work [00:36:15] You [00:36:15] Know [00:36:16] Makes [00:36:16] You [00:36:16] Special [00:36:16] You [00:36:17] Design [00:36:18] Machines [00:36:18] You [00:36:19] Operate [00:36:20] Factories [00:36:20] You [00:36:22] Discover [00:36:22] Medicine [00:36:23] You [00:36:24] Have [00:36:24] Extraordinary [00:36:25] Craft [00:36:25] That [00:36:27] Craft [00:36:28] That [00:36:28] Specialty [00:36:29] Makes [00:36:30] You [00:36:30] Special [00:36:30] Well [00:36:31] What [00:36:31] Makes [00:36:32] NVIDIA [00:36:32] Special [00:36:32] Are [00:36:34] Really [00:36:34] Two [00:36:34] Things [00:36:35] The [00:36:36] First [00:36:36] Thing [00:36:36] Is [00:36:37] That [00:36:37] Our [00:36:38] Architecture [00:36:39] Is [00:36:39] The [00:36:39] Only [00:36:40] Accelerated [00:36:40] Computing [00:36:41] Architecture [00:36:41] That [00:36:42] Is [00:36:42] General [00:36:42] Purpose [00:36:43] We [00:36:44] Don't [00:36:44] Specialize [00:36:45] On [00:36:45] Just [00:36:46] One [00:36:46] Algorithm [00:36:47] For [00:36:47] Example [00:36:48] Transformers [00:36:49] We [00:36:49] Specialize [00:36:50] Across [00:36:51] Anything [00:36:52] That [00:36:53] Could [00:36:54] Be [00:36:54] Fluids [00:36:54] And [00:36:55] Particles [00:36:55] It [00:36:56] Could [00:36:56] Be [00:36:56] Quantum [00:36:56] It [00:36:57] Could [00:36:57] Be [00:36:58] Quantum [00:36:58] Chemistry [00:36:58] Could [00:36:59] Be [00:37:00] Image [00:37:00] Processing [00:37:01] Could [00:37:01] Be [00:37:01] Data [00:37:01] Processing [00:37:02] Structured [00:37:03] Vector [00:37:03] Data [00:37:03] Processing [00:37:04] Processing [00:37:05] Of [00:37:05] All [00:37:06] Kinds [00:37:06] And [00:37:07] Because [00:37:07] We [00:37:07] Been [00:37:07] Working [00:37:07] At [00:37:08] It [00:37:08] For [00:37:08] So [00:37:08] Long [00:37:08] And [00:37:09] Because [00:37:09] NVIDIA [00:37:09] Architecture [00:37:10] Has [00:37:10] Remained [00:37:11] Consistent [00:37:12] For [00:37:12] Literally [00:37:13] Two [00:37:13] And [00:37:13] A [00:37:13] Half [00:37:13] Decades [00:37:14] The [00:37:14] Number [00:37:15] Of [00:37:15] Algorithms [00:37:17] The [00:37:23] Install [00:37:24] Base [00:37:24] Our [00:37:25] General [00:37:26] Purpose [00:37:26] Architecture [00:37:27] Our [00:37:27] Dedication [00:37:28] To [00:37:28] This [00:37:28] Architecture [00:37:29] And [00:37:30] Our [00:37:30] Collaboration [00:37:31] With [00:37:31] All [00:37:31] Of [00:37:31] The [00:37:31] World's [00:37:32] Industries [00:37:32] Has [00:37:33] Made [00:37:33] NVIDIA's [00:37:34] Architecture [00:37:34] Literally [00:37:35] Everywhere [00:37:36] The [00:37:37] Second [00:37:37] Thing [00:37:39] Is [00:37:39] How [00:37:39] We [00:37:39] Do [00:37:39] It [00:37:40] Of [00:37:41] Course [00:37:41] We [00:37:41] Have [00:37:42] The [00:37:42] Architecture [00:37:42] That [00:37:42] We [00:37:43] Developed [00:37:44] But [00:37:44] We [00:37:44] Also [00:37:44] Build [00:37:45] It [00:37:45] Now [00:37:45] In [00:37:46] A [00:37:46] Full [00:37:47] Stack [00:37:48] Complete [00:37:49] AI [00:37:50] Factory [00:37:50] There [00:37:51] Are [00:37:51] Many [00:37:52] Companies [00:37:52] In [00:37:52] The [00:37:52] World [00:37:52] Who [00:37:53] Want [00:37:54] To [00:37:54] Design [00:37:54] Their [00:37:54] Own [00:37:54] Chips [00:37:55] But [00:37:56] Most [00:37:56] Of [00:37:56] The [00:37:56] Companies [00:37:57] In [00:37:57] The [00:37:57] World [00:37:57] Just [00:37:58] Wants [00:37:58] To [00:37:58] Deploy [00:37:59] The [00:37:59] AI [00:37:59] Infrastructure [00:38:00] So [00:38:00] That [00:38:00] They [00:38:01] Can [00:38:01] Run [00:38:01] Their [00:38:01] Business [00:38:01] On [00:38:02] Top [00:38:02] Of [00:38:02] It [00:38:02] The [00:38:03] Ability [00:38:03] To [00:38:04] Design [00:38:04] A [00:38:04] Chip [00:38:04] And [00:38:05] Build [00:38:05] A [00:38:05] Super [00:38:05] Computer [00:38:06] Write [00:38:06] All [00:38:06] The [00:38:06] Software [00:38:07] Is [00:38:08] Really [00:38:08] Completely [00:38:09] Out [00:38:10] Of [00:38:10] Mind [00:38:10] For [00:38:10] Most [00:38:11] Of [00:38:11] The [00:38:11] Companies [00:38:11] Around [00:38:11] The [00:38:12] World [00:38:12] And [00:38:12] So [00:38:13] We [00:38:13] Now [00:38:13] Have [00:38:13] The [00:38:14] Combination [00:38:14] Of [00:38:14] These [00:38:15] Two [00:38:15] Ideas [00:38:15] Cuda [00:38:16] Full [00:38:17] Stack [00:38:17] AI [00:38:18] Factories [00:38:18] And [00:38:19] It [00:38:19] Exposes [00:38:19] Us [00:38:20] To [00:38:20] Markets [00:38:21] That [00:38:21] Otherwise [00:38:22] Is [00:38:22] Impossible [00:38:23] We [00:38:24] Are [00:38:24] Able [00:38:24] To [00:38:24] Partner [00:38:25] With [00:38:25] So [00:38:25] Many [00:38:25] Companies [00:38:26] Here [00:38:26] For [00:38:26] This [00:38:27] Very [00:38:27] Reason [00:38:27] If [00:38:28] I [00:38:28] Were [00:38:28] To [00:38:28] Go [00:38:28] Around [00:38:29] This [00:38:30] Starting [00:38:30] With [00:38:31] CSP [00:38:31] This [00:38:32] Is [00:38:32] Where [00:38:32] Most [00:38:33] People [00:38:33] Think [00:38:34] The [00:38:34] Vast [00:38:34] Majority [00:38:35] Of [00:38:35] AI [00:38:35] Is [00:38:35] Today [00:38:36] But [00:38:36] As [00:38:37] You [00:38:37] Can [00:38:37] See [00:38:37] In [00:38:38] The [00:38:38] Beginning [00:38:38] Of [00:38:38] Our [00:38:38] Keynote [00:38:39] I [00:38:39] Mentioned [00:38:39] A [00:38:40] Whole [00:38:40] Bunch [00:38:40] Of [00:38:40] Companies [00:38:40] Who [00:38:41] Are [00:38:41] Not [00:38:41] Those [00:38:42] So [00:38:42] CSP [00:38:43] Remain [00:38:44] Very [00:38:44] Large [00:38:45] Businesses [00:38:45] For [00:38:46] The [00:38:46] World's [00:38:46] Computing [00:38:47] But [00:38:47] There [00:38:47] Are [00:38:48] AI [00:38:48] Natives [00:38:48] Companies [00:38:49] That [00:38:50] Are [00:38:50] Starting [00:38:50] All [00:38:50] Over [00:38:51] The [00:38:51] World [00:38:51] There [00:38:51] Are [00:38:52] So [00:38:52] Many [00:38:52] That [00:38:52] Are [00:38:52] Here [00:38:53] Neo [00:38:54] Clouds [00:38:55] Because [00:38:55] Every [00:38:56] Region [00:38:57] Wants [00:38:57] To [00:38:57] Participate [00:38:58] In [00:38:58] The [00:38:58] AI [00:38:59] Economy [00:38:59] Every [00:39:00] Region [00:39:00] Wants [00:39:01] To [00:39:01] Ensure [00:39:01] That [00:39:02] The [00:39:02] Valuable [00:39:03] Resources [00:39:03] That [00:39:03] They [00:39:03] Have [00:39:04] The [00:39:04] Energy [00:39:05] That [00:39:05] They [00:39:05] Have [00:39:05] Is [00:39:06] Dedicated [00:39:06] To [00:39:06] Companies [00:39:07] In [00:39:07] Their [00:39:07] Region [00:39:08] For [00:39:09] Example [00:39:09] Mistral [00:39:10] Here [00:39:10] In [00:39:10] Europe [00:39:11] Mistral [00:39:12] In [00:39:12] France [00:39:12] Nibias [00:39:13] Here [00:39:13] In [00:39:13] Europe [00:39:14] So [00:39:14] Many [00:39:15] Companies [00:39:15] Around [00:39:16] Europe [00:39:16] That [00:39:16] Are [00:39:17] Building [00:39:17] Up [00:39:17] Neo [00:39:18] Clouds [00:39:18] OEMs [00:39:20] And [00:39:20] ODMs [00:39:21] Who [00:39:21] Are [00:39:21] Building [00:39:21] Computers [00:39:22] For [00:39:22] The [00:39:22] World's [00:39:23] Enterprise [00:39:23] AI [00:39:24] Labs [00:39:25] Of [00:39:26] Course [00:39:26] We [00:39:26] Know [00:39:26] About [00:39:27] Last [00:39:28] Year [00:39:28] All [00:39:29] We [00:39:29] Spoke [00:39:29] About [00:39:30] Was [00:39:30] Open [00:39:30] AI [00:39:31] Today [00:39:32] We [00:39:32] Spoke [00:39:32] About [00:39:33] Open [00:39:33] AI [00:39:34] Anthropic [00:39:35] But [00:39:35] We [00:39:36] Also [00:39:36] Talk [00:39:36] About [00:39:36] The [00:39:37] Incredible [00:39:37] Meta [00:39:38] Muse [00:39:38] Grock [00:39:40] Bot [00:39:41] And [00:39:41] Of [00:39:42] Course [00:39:42] Gemini [00:39:42] So [00:39:43] Many [00:39:43] Different [00:39:44] AI [00:39:44] Labs [00:39:45] Are [00:39:45] Now [00:39:45] Participating [00:39:46] At [00:39:46] The [00:39:47] Frontier [00:39:47] And [00:39:48] Then [00:39:48] Very [00:39:48] Importantly [00:39:49] And [00:39:49] This [00:39:49] Is [00:39:50] Ultimately [00:39:50] The [00:39:50] Most [00:39:51] Important [00:39:51] For [00:39:52] AI [00:39:52] To [00:39:52] Succeed [00:39:53] For [00:39:54] Europe [00:39:54] To [00:39:55] Benefit [00:39:55] From [00:39:55] AI [00:39:56] And [00:39:56] To [00:39:56] Have [00:39:56] Really [00:39:57] Created [00:39:57] An [00:39:57] AI [00:39:57] Ecosystem [00:39:58] And [00:39:58] AI [00:39:59] Economy [00:39:59] You [00:40:00] Need [00:40:00] All [00:40:01] Of [00:40:01] The [00:40:01] End [00:40:02] Markets [00:40:02] The [00:40:02] End [00:40:02] Industries [00:40:03] To [00:40:04] Adopt [00:40:04] AI [00:40:04] And [00:40:05] To [00:40:05] Succeed [00:40:05] With [00:40:06] AI [00:40:06] Whether [00:40:07] It's [00:40:07] In [00:40:07] The [00:40:07] Automotive [00:40:07] Industry [00:40:08] Because [00:40:08] We're [00:40:08] Building [00:40:09] Self-Driving [00:40:09] Cars [00:40:09] Or [00:40:10] The [00:40:11] Way [00:40:11] You [00:40:11] Manufacture [00:40:12] With [00:40:12] Robotics [00:40:13] The [00:40:13] Way [00:40:13] You [00:40:13] Discover [00:40:14] New [00:40:14] Drugs [00:40:15] In [00:40:15] Healthcare [00:40:16] Your [00:40:16] Financial [00:40:17] Services [00:40:18] Industry [00:40:20] All [00:40:20] All [00:40:20] All [00:40:20] All [00:40:20] Of [00:40:21] These [00:40:21] Different [00:40:21] Industries [00:40:22] Have [00:40:22] To [00:40:22] Become [00:40:23] AI [00:40:23] Companies [00:40:24] Every [00:40:25] Single [00:40:25] One [00:40:25] Of [00:40:25] Them [00:40:26] Well [00:40:26] We [00:40:26] Have [00:40:27] A [00:40:27] Unique [00:40:27] Ability [00:40:28] Through [00:40:28] What [00:40:29] I [00:40:29] Just [00:40:29] Mentioned [00:40:30] About [00:40:30] Cuda [00:40:30] And [00:40:31] The [00:40:31] Way [00:40:31] That [00:40:32] We [00:40:32] Build [00:40:32] Computers [00:40:32] We [00:40:33] Have [00:40:33] A [00:40:33] Unique [00:40:33] Ability [00:40:34] To [00:40:34] Serve [00:40:34] Literally [00:40:35] Every [00:40:35] Single [00:40:36] Segment [00:40:36] Of [00:40:37] AI [00:40:37] And [00:40:42] So [00:40:42] One [00:40:43] Architecture [00:40:43] Many [00:40:44] Markets [00:40:45] And [00:40:46] Because [00:40:46] These [00:40:46] Infrastructure [00:40:47] Are [00:40:47] So [00:40:48] Significant [00:40:48] And [00:40:49] So [00:40:49] Expensive [00:40:49] We [00:40:50] Need [00:40:50] To [00:40:50] Make [00:40:51] Sure [00:40:51] That [00:40:52] They [00:40:52] Are [00:40:52] As [00:40:52] Productive [00:40:53] As [00:40:53] Possible [00:40:54] As [00:40:54] Soon [00:40:54] As [00:40:55] Possible [00:40:55] That [00:40:56] They [00:40:56] Are [00:40:56] Fungible [00:40:56] That [00:40:58] You [00:40:58] Can [00:40:58] Use [00:40:58] It [00:40:58] For [00:40:58] Every [00:40:59] Workload [00:40:59] From [00:41:00] The [00:41:00] Beginning [00:41:00] Of [00:41:01] Your [00:41:01] AI [00:41:02] Journey [00:41:02] Data [00:41:03] Processing [00:41:04] To [00:41:04] Training [00:41:04] To [00:41:04] Post [00:41:05] Training [00:41:05] Inference [00:41:07] And [00:41:07] Agentic [00:41:07] Processing [00:41:08] Every [00:41:08] Single [00:41:09] Step [00:41:09] Of [00:41:09] The [00:41:09] AI [00:41:10] Journey [00:41:10] To [00:41:11] Every [00:41:11] Type [00:41:11] Of [00:41:12] AI [00:41:12] Chemicals [00:41:13] And [00:41:13] Physics [00:41:13] To [00:41:14] Languages [00:41:15] To [00:41:15] Every [00:41:15] Single [00:41:16] Industry [00:41:16] And [00:41:17] Algorithms [00:41:17] That [00:41:18] They [00:41:18] Use [00:41:18] And [00:41:19] So [00:41:19] NVIDIA's [00:41:20] Architecture [00:41:20] Is [00:41:20] Fungible [00:41:21] And [00:41:21] As [00:41:21] Result [00:41:21] Of [00:41:22] That [00:41:22] Continuous [00:41:23] Innovation [00:41:24] Continuous [00:41:25] Updating [00:41:26] Of [00:41:26] Software [00:41:26] Because [00:41:28] The [00:41:29] Install [00:41:29] Base [00:41:29] Is [00:41:29] So [00:41:29] Great [00:41:30] We [00:41:30] Have [00:41:30] Such [00:41:31] An [00:41:31] Incredibly [00:41:31] Rich [00:41:32] Ecosystem [00:41:32] Of [00:41:32] Software [00:41:33] Then [00:41:33] These [00:41:34] Architectures [00:41:34] Also [00:41:35] Durable [00:41:35] Fungible [00:41:36] Durable [00:41:37] Productive [00:41:38] Well [00:41:39] That [00:41:40] Lays [00:41:41] The [00:41:41] Foundation [00:41:41] That [00:41:43] Is [00:41:43] The [00:41:43] Infrastructure [00:41:44] Now [00:41:44] What [00:41:45] Do [00:41:45] We [00:41:45] Do [00:41:45] With [00:41:45] It [00:41:45] This [00:41:47] Last [00:41:47] Year [00:41:47] We [00:41:48] Saw [00:41:48] A [00:41:48] Transition [00:41:48] In [00:41:49] AI [00:41:49] AI [00:41:50] Is [00:41:50] Now [00:41:50] Productive [00:41:51] And [00:41:51] It's [00:41:52] One [00:41:53] Of [00:41:53] The [00:41:53] Most [00:41:53] Important [00:41:53] Times [00:41:54] Was [00:41:54] In [00:41:55] The [00:41:55] Beginning [00:41:55] Of [00:41:55] The [00:41:55] Year [00:41:56] Spoke [00:41:57] About [00:41:57] Open [00:41:58] Claw [00:41:58] Which [00:41:59] Introduced [00:42:00] The [00:42:00] Concept [00:42:00] Of [00:42:01] An [00:42:01] Open [00:42:01] Source [00:42:01] Agent [00:42:02] And [00:42:02] Really [00:42:02] Started [00:42:03] What [00:42:04] We [00:42:04] All [00:42:04] See [00:42:04] Across [00:42:05] The [00:42:05] World [00:42:05] Today [00:42:05] Agents [00:42:07] Think [00:42:08] Reason [00:42:10] Iterate [00:42:11] Use [00:42:11] Tools [00:42:12] And [00:42:12] As [00:42:12] Result [00:42:13] Consume [00:42:14] An [00:42:14] Enormous [00:42:15] Amount [00:42:15] Of [00:42:15] Tokens [00:42:15] But [00:42:16] As [00:42:16] Result [00:42:17] They [00:42:17] Become [00:42:17] Useful [00:42:18] So [00:42:19] More [00:42:19] Agents [00:42:20] Are [00:42:20] Being [00:42:20] Developed [00:42:21] They're [00:42:21] More [00:42:21] Useful [00:42:22] And [00:42:22] Therefore [00:42:23] They're [00:42:23] Used [00:42:23] By [00:42:24] More [00:42:24] People [00:42:25] And [00:42:25] Because [00:42:25] They have [00:42:26] To [00:42:26] Use [00:42:26] A [00:42:26] Lot [00:42:26] Of [00:42:26] Tokens [00:42:27] To [00:42:27] Think [00:42:27] The [00:42:28] Amount [00:42:28] Of [00:42:28] Token [00:42:29] Consumption [00:42:29] Has [00:42:30] Gone [00:42:30] Through [00:42:30] The [00:42:30] Roof [00:42:30] GPU [00:42:31] Man [00:42:32] To [00:42:32] The [00:42:33] Roof [00:42:33] Now [00:42:34] What's [00:42:35] Processing [00:42:35] It [00:42:35] Is [00:42:35] This [00:42:36] This [00:42:36] Is [00:42:37] An [00:42:37] Agent [00:42:37] These [00:42:38] Are [00:42:38] The [00:42:39] Five [00:42:39] Elements [00:42:39] Of [00:42:40] An [00:42:40] Agent [00:42:40] The [00:42:40] First [00:42:41] Part [00:42:41] Are [00:42:42] The [00:42:42] Language [00:42:43] Models [00:42:43] These [00:42:44] Are [00:42:44] The [00:42:44] Tokens [00:42:45] That [00:42:45] Are [00:42:45] Running [00:42:45] On [00:42:46] Top [00:42:46] Of [00:42:46] The [00:42:46] GPUs [00:42:47] And [00:42:47] They're [00:42:47] Generating [00:42:48] Tokens [00:42:48] That [00:42:48] Are [00:42:49] Consumed [00:42:49] By [00:42:50] The [00:42:50] Agent [00:42:50] What [00:42:51] Is [00:42:51] I [00:42:51] Called [00:42:52] The [00:42:52] Exoskeleton [00:42:53] That [00:42:54] Open [00:42:54] Claw [00:42:55] Harness [00:42:55] Sits [00:42:56] Inside [00:42:57] A [00:42:57] Sandbox [00:42:58] And [00:42:58] That [00:42:58] Sandbox [00:42:59] Isolates [00:43:00] It [00:43:00] From [00:43:00] The [00:43:00] Rest [00:43:00] Of [00:43:00] The [00:43:00] System [00:43:01] This [00:43:01] Is [00:43:01] A [00:43:02] Very [00:43:02] Important [00:43:02] Concept [00:43:02] That [00:43:03] We're [00:43:03] Going [00:43:03] To [00:43:03] Come [00:43:03] Back [00:43:03] To [00:43:03] There [00:43:05] Are [00:43:05] Five [00:43:05] Things [00:43:05] There's [00:43:06] The [00:43:06] Model [00:43:06] There's [00:43:08] The [00:43:08] Harness [00:43:08] The [00:43:09] Sandbox [00:43:10] That [00:43:10] The [00:43:10] Harness [00:43:10] Sits [00:43:11] In [00:43:11] And [00:43:12] Two [00:43:12] Other [00:43:12] Things [00:43:12] Knowledge [00:43:14] And [00:43:14] Tools [00:43:15] Models [00:43:16] Sandbox [00:43:17] Harness [00:43:19] Knowledge [00:43:20] And [00:43:20] Tools [00:43:20] These [00:43:21] Are [00:43:21] The [00:43:21] Five [00:43:21] Elements [00:43:22] Of [00:43:23] An [00:43:23] Agent [00:43:24] Now [00:43:24] Before [00:43:25] I [00:43:25] Talk [00:43:26] About [00:43:26] How [00:43:26] To [00:43:26] Make [00:43:26] The [00:43:26] Agents [00:43:27] Useful [00:43:28] The [00:43:29] Most [00:43:29] Important [00:43:29] Thing [00:43:29] Is [00:43:29] We [00:43:30] Have [00:43:30] To [00:43:30] Make [00:43:30] Sure [00:43:30] That [00:43:30] The [00:43:30] Agent [00:43:31] Is [00:43:31] Secure [00:43:31] The [00:43:32] Agent [00:43:33] Is [00:43:33] Autonomous [00:43:33] And [00:43:34] It's [00:43:34] Autonomous [00:43:35] System [00:43:35] That [00:43:36] Has [00:43:36] Access [00:43:36] To [00:43:37] Files [00:43:37] Has [00:43:37] Access [00:43:38] To [00:43:38] Resources [00:43:38] Networks [00:43:39] Access [00:43:39] To [00:43:40] Tools [00:43:40] And [00:43:41] We [00:43:41] Need [00:43:41] To [00:43:41] Make [00:43:41] Sure [00:43:42] That [00:43:42] The [00:43:42] Privileges [00:43:44] That [00:43:44] We [00:43:44] Provide [00:43:45] This [00:43:45] Autonomous [00:43:46] System [00:43:46] Is [00:43:47] Minimal [00:43:48] And [00:43:52] He [00:43:52] Introduced [00:43:53] A [00:43:54] Fantastic [00:43:55] New [00:43:55] Product [00:43:55] A [00:43:56] New [00:43:56] Platform [00:43:57] And [00:43:57] It's [00:43:57] Called [00:43:58] Open [00:43:58] Shell [00:43:58] Open [00:43:59] Shell [00:44:00] Came [00:44:00] Just [00:44:00] In [00:44:00] Time [00:44:01] This [00:44:02] Is [00:44:02] A [00:44:02] Secure [00:44:02] Runtime [00:44:03] That [00:44:04] Provides [00:44:04] For [00:44:04] The [00:44:05] Things [00:44:05] That [00:44:05] I [00:44:05] Just [00:44:05] Mentioned [00:44:06] And [00:44:07] Has [00:44:08] An [00:44:08] Additional [00:44:09] Hardware [00:44:10] Based [00:44:10] Monitor [00:44:10] Before [00:44:11] I [00:44:11] Explain [00:44:12] It [00:44:12] To [00:44:12] You [00:44:12] Let's [00:44:13] Watch [00:44:13] A [00:44:13] Video [00:44:13] Prises [00:44:18] Are [00:44:18] Preparing [00:44:19] To [00:44:19] Deploy [00:44:19] Billions [00:44:20] Of [00:44:20] AI [00:44:20] Agents [00:44:21] To [00:44:22] Be [00:44:22] Useful [00:44:22] They'll [00:44:23] Reason [00:44:23] Plan [00:44:24] Use [00:44:25] Tools [00:44:25] Access [00:44:26] Data [00:44:26] And [00:44:27] Work [00:44:27] Together [00:44:28] Today [00:44:29] The [00:44:30] Only [00:44:30] Thing [00:44:30] Stopping [00:44:30] An [00:44:31] Agent [00:44:31] From [00:44:31] Misusing [00:44:32] All [00:44:32] That [00:44:32] Is [00:44:33] Asking [00:44:33] It [00:44:33] Nicely [00:44:34] Please [00:44:34] Don't [00:44:35] Delete [00:44:35] Anything [00:44:35] Please [00:44:36] Don't [00:44:36] Push [00:44:37] To [00:44:37] Main [00:44:37] Please [00:44:38] Don't [00:44:38] Touch [00:44:39] My [00:44:39] SSH [00:44:39] Keys [00:44:40] That's [00:44:41] The [00:44:41] Please [00:44:41] Don't [00:44:42] Approach [00:44:42] But [00:44:43] Hope [00:44:43] Isn't [00:44:44] Security [00:44:44] Safety [00:44:45] Isn't [00:44:46] An [00:44:46] Aspiration [00:44:46] It's [00:44:47] Technology [00:44:48] And [00:44:48] Engineering [00:44:49] NVIDIA [00:44:50] Open [00:44:50] Shell [00:44:50] And [00:44:50] Sentry [00:44:51] Put [00:44:51] That [00:44:51] Philosophy [00:44:52] Into [00:44:52] Practice [00:44:53] You [00:44:54] Write [00:44:54] A [00:44:54] Policy [00:44:55] For [00:44:55] What [00:44:55] Each [00:44:55] Agent [00:44:55] Can [00:44:56] Touch [00:44:56] And [00:44:57] Open [00:44:57] Shell [00:44:57] Enforces [00:44:58] It [00:44:58] At [00:44:58] The [00:44:58] Granularity [00:44:59] Of [00:44:59] System [00:45:00] Calls [00:45:00] File [00:45:01] Access [00:45:01] And [00:45:02] Network [00:45:02] Connections [00:45:03] Formal [00:45:04] Verification [00:45:05] Every [00:45:06] Policy [00:45:06] Change [00:45:07] Against [00:45:07] Your [00:45:07] Permission [00:45:08] Ceiling [00:45:08] So [00:45:09] Agents [00:45:09] Never [00:45:10] Gain [00:45:10] More [00:45:10] Access [00:45:10] Than [00:45:11] Intended [00:45:11] It's [00:45:12] Like [00:45:13] The [00:45:13] Early [00:45:13] Internet [00:45:13] Viruses [00:45:15] Were [00:45:15] Everywhere [00:45:15] Until [00:45:16] Browsers [00:45:17] Enforced [00:45:17] Strict [00:45:18] Limits [00:45:18] On [00:45:18] What [00:45:19] A [00:45:19] Web [00:45:19] Page [00:45:19] Could [00:45:19] Do [00:45:20] Open [00:45:21] Shell [00:45:21] Does [00:45:21] The [00:45:22] Same [00:45:22] For [00:45:22] Agents [00:45:23] Enforcement [00:45:24] Happens [00:45:24] Outside [00:45:25] The [00:45:25] Agent [00:45:25] Not [00:45:26] Just [00:45:26] In [00:45:26] LLM [00:45:27] Guardrails [00:45:27] Bringing [00:45:28] Agent [00:45:28] Native [00:45:29] Security [00:45:29] Into [00:45:30] Your [00:45:35] With [00:45:36] No [00:45:36] Direct [00:45:36] Route [00:45:36] To [00:45:37] The [00:45:37] Network [00:45:37] Every [00:45:38] Outbound [00:45:39] Network [00:45:39] Request [00:45:40] Goes [00:45:40] Through [00:45:40] The [00:45:40] Supervisor [00:45:41] Outside [00:45:42] The [00:45:42] Sandbox [00:45:42] It [00:45:43] Checks [00:45:43] The [00:45:44] Action [00:45:44] Against [00:45:44] Your [00:45:45] Policy [00:45:45] If [00:45:46] It's [00:45:46] Not [00:45:46] Permitted [00:45:46] It's [00:45:47] Denied [00:45:47] Keys [00:45:48] Stay [00:45:49] Outside [00:45:49] With [00:45:50] The [00:45:50] Supervisor [00:45:51] And [00:45:51] Are [00:45:51] Added [00:45:51] Only [00:45:52] On [00:45:52] The [00:45:52] Way [00:45:52] To [00:45:52] An [00:45:53] Approved [00:45:53] End [00:45:53] Point [00:45:54] A [00:45:54] Compromised [00:45:55] Agent [00:45:55] Can't [00:45:56] Leak [00:45:56] A [00:45:56] Key [00:45:56] It [00:45:56] Never [00:45:56] Had [00:45:57] The [00:45:58] Gateway [00:45:58] Is [00:45:58] The [00:45:59] Control [00:45:59] Plane [00:45:59] The [00:46:00] Supervisor [00:46:01] Makes [00:46:01] The [00:46:01] Decisions [00:46:02] The [00:46:03] Sandbox [00:46:03] Holds [00:46:03] The [00:46:04] Agent [00:46:04] And [00:46:05] The [00:46:05] Policy [00:46:05] Prover [00:46:05] Checks [00:46:06] Proposed [00:46:06] Policies [00:46:07] Against [00:46:07] Your [00:46:08] Permission [00:46:08] Ceilings [00:46:09] Open [00:46:10] Shell [00:46:10] Combines [00:46:11] Linux [00:46:11] Kernel [00:46:11] Isolation [00:46:12] With [00:46:12] Runtime [00:46:13] Policy [00:46:13] Enforcement [00:46:14] NVIDIA [00:46:15] Sentry [00:46:16] Adds [00:46:16] An [00:46:16] Optional [00:46:17] Layer [00:46:17] Of [00:46:17] Protection [00:46:18] Outside [00:46:18] The [00:46:19] Host [00:46:19] Itself [00:46:20] Away [00:46:20] From [00:46:20] The [00:46:21] Agent [00:46:21] Running [00:46:22] On [00:46:22] NVIDIA [00:46:23] Bluefield [00:46:23] 4 [00:46:24] Sentry [00:46:24] Monitors [00:46:25] Agent [00:46:25] Activity [00:46:26] And [00:46:26] Enforces [00:46:27] Security [00:46:27] Policies [00:46:28] In [00:46:28] A [00:46:28] Hardware [00:46:29] Isolated [00:46:29] Trust [00:46:30] Domain [00:46:30] Because [00:46:31] Agent [00:46:31] Reasoning [00:46:32] Traffic [00:46:32] Passes [00:46:33] Through [00:46:33] Sentry [00:46:33] If [00:46:34] The [00:46:34] Agent [00:46:34] Is [00:46:35] Planning [00:46:35] To [00:46:35] Do [00:46:35] Something [00:46:36] Unintended [00:46:36] It [00:46:37] Will [00:46:37] Update [00:46:37] Policies [00:46:38] Immediately [00:46:38] And [00:46:39] Cut [00:46:39] The [00:46:39] Connection [00:46:40] To [00:46:40] The [00:46:40] Model [00:46:40] Quarantining [00:46:41] It [00:46:42] In [00:46:42] Milliseconds [00:46:43] And [00:46:44] It [00:46:44] Keeps [00:46:44] Tamper [00:46:45] Evident [00:46:45] Records [00:46:45] That [00:46:46] Operators [00:46:46] Can [00:46:46] Verify [00:46:47] For [00:46:47] Detection [00:46:48] Investigation [00:46:49] And [00:46:49] Response [00:46:50] Open [00:46:51] Shell [00:46:51] Is [00:46:51] Open [00:46:52] Source [00:46:52] Under [00:46:52] Apache [00:46:53] 2.0 [00:46:54] Runs [00:46:54] Everywhere [00:46:55] Your [00:46:55] Agents [00:46:55] Do [00:46:56] Edge [00:46:56] To [00:46:56] Cloud [00:46:57] And [00:46:57] Takes [00:46:58] Two [00:46:58] Commands [00:46:59] To [00:46:59] Adopt [00:46:59] Install [00:47:00] The [00:47:00] CLI [00:47:01] Create [00:47:02] A [00:47:02] Sandbox [00:47:02] And [00:47:03] Run [00:47:03] Your [00:47:03] Favorite [00:47:03] Agent [00:47:04] Agents [00:47:05] Are [00:47:05] Only [00:47:05] As [00:47:06] Useful [00:47:06] As [00:47:06] The [00:47:06] Access [00:47:07] You [00:47:07] Give [00:47:07] Them [00:47:07] Open [00:47:08] Shell [00:47:08] And [00:47:09] Sentry [00:47:09] Put [00:47:09] That [00:47:10] Access [00:47:10] Under [00:47:10] Your [00:47:10] Control [00:47:11] That's [00:47:12] How [00:47:12] We [00:47:12] Advance [00:47:12] AI [00:47:13] Capabilities [00:47:13] And [00:47:14] Safety [00:47:14] Together [00:47:15] Agents [00:47:26] Agents [00:47:26] Are [00:47:26] Autonomous [00:47:26] Systems [00:47:27] We [00:47:27] Have [00:47:27] To [00:47:28] Create [00:47:28] Extremely [00:47:30] Secure [00:47:31] Technology [00:47:32] To [00:47:32] Make [00:47:33] Sure [00:47:33] That [00:47:33] We [00:47:33] Could [00:47:33] Develop [00:47:34] These [00:47:34] Agents [00:47:34] And [00:47:35] Deploy [00:47:35] These [00:47:35] Agents [00:47:36] Safely [00:47:36] Now [00:47:37] It's [00:47:38] Incredible [00:47:38] After [00:47:39] Announcing [00:47:39] Open [00:47:40] Shell [00:47:40] Just [00:47:40] About [00:47:41] Every [00:47:41] Major [00:47:42] Company [00:47:42] In [00:47:42] The [00:47:42] World [00:47:43] From [00:47:43] The [00:47:43] Frontier [00:47:44] Labs [00:47:44] To [00:47:45] The [00:47:45] Operating [00:47:46] System [00:47:46] Companies [00:47:46] Cyber [00:47:47] Security [00:47:48] Companies [00:47:48] Platform [00:47:49] Companies [00:47:50] Have [00:47:50] Come [00:47:50] Out [00:47:51] To [00:47:51] Adopt [00:47:51] Open [00:47:52] Shell [00:47:52] It's [00:47:52] Really [00:47:53] Fantastic [00:47:53] To [00:47:53] See [00:47:54] And [00:47:54] The [00:47:54] Reason [00:47:54] For [00:47:54] That [00:47:54] Is [00:47:55] Because [00:47:55] This [00:47:55] Technology [00:47:55] Has [00:47:56] Arrived [00:47:56] At [00:47:56] Exactly [00:47:57] The [00:47:57] Right [00:47:57] Time [00:47:57] At [00:47:58] A [00:47:58] Time [00:47:58] When [00:47:58] We [00:47:58] Want [00:47:58] To [00:47:58] Deploy [00:47:59] Agents [00:47:59] All [00:48:00] Over [00:48:00] The [00:48:00] World [00:48:00] Now [00:48:01] Remember [00:48:01] This [00:48:02] Isn't [00:48:02] Just [00:48:02] About [00:48:03] The [00:48:03] Services [00:48:04] Using [00:48:05] Open [00:48:05] Shell [00:48:06] As [00:48:06] They [00:48:06] Deploy [00:48:07] Agents [00:48:07] This [00:48:08] Is [00:48:08] Really [00:48:08] About [00:48:08] All [00:48:09] Of [00:48:09] The [00:48:09] World [00:48:09] Companies [00:48:10] Also [00:48:11] Adopting [00:48:12] Open [00:48:12] Shell [00:48:12] So [00:48:13] That [00:48:13] You [00:48:13] Could [00:48:13] Receive [00:48:14] These [00:48:14] Agents [00:48:14] And [00:48:15] Have [00:48:15] Them [00:48:15] Run [00:48:15] Autonomously [00:48:16] Inside [00:48:16] Your [00:48:16] Company [00:48:17] Every [00:48:18] Single [00:48:18] Level [00:48:19] Of [00:48:20] Capability [00:48:20] Every [00:48:21] Single [00:48:22] Task [00:48:22] That [00:48:23] You [00:48:23] Give [00:48:23] The [00:48:24] Agent [00:48:24] For [00:48:25] Every [00:48:25] Person [00:48:25] That [00:48:26] Uses [00:48:26] The [00:48:26] Agent [00:48:27] Every [00:48:27] One [00:48:27] Of [00:48:27] Those [00:48:27] Combinations [00:48:28] Requires [00:48:29] A [00:48:29] Different [00:48:29] Level [00:48:30] Of [00:48:30] Privilege [00:48:31] That [00:48:31] You [00:48:31] Would [00:48:31] Provide [00:48:32] To [00:48:32] The [00:48:32] Agent [00:48:32] And [00:48:33] Every [00:48:34] Single [00:48:34] Time [00:48:34] An [00:48:35] Agent [00:48:35] Is [00:48:35] Launched [00:48:36] The [00:48:36] Privileges [00:48:37] Are [00:48:37] Different [00:48:37] And [00:48:38] So [00:48:38] Privileges [00:48:39] The [00:48:40] Policy [00:48:40] Enforcement [00:48:41] The [00:48:42] Monitoring [00:48:43] All [00:48:43] Has [00:48:44] To [00:48:44] Be [00:48:44] Done [00:48:44] Dynamically [00:48:45] And [00:48:46] So [00:48:46] Open [00:48:46] Shell [00:48:47] Is [00:48:47] Really [00:48:47] Revolutionary [00:48:48] I [00:48:49] Delight [00:48:49] To [00:48:49] Say [00:48:49] That [00:48:50] Anthropic [00:48:51] And [00:48:51] NVIDIA [00:48:51] Are [00:48:51] Partnering [00:48:52] To [00:48:52] Create [00:48:53] Clyde [00:48:53] Open [00:48:54] Shell [00:48:54] Agents [00:48:55] And [00:48:56] So [00:48:56] You [00:48:56] Can [00:48:56] Go [00:48:56] To [00:48:56] The [00:48:56] Website [00:48:57] There [00:48:57] Is [00:48:57] A [00:48:58] Really [00:48:58] Great [00:48:58] Write [00:48:58] Up [00:48:58] About [00:48:58] That [00:48:59] So [00:48:59] The [00:48:59] First [00:49:00] Part [00:49:00] Is [00:49:02] How [00:49:03] Do [00:49:03] You [00:49:04] Use [00:49:04] These [00:49:04] Harnesses [00:49:05] Open [00:49:05] Claw [00:49:06] Hermes [00:49:07] Open [00:49:08] Hand [00:49:08] You [00:49:09] Could [00:49:09] Create [00:49:09] Your [00:49:09] Own [00:49:10] Harnesses [00:49:11] From [00:49:11] Lang [00:49:11] Chain [00:49:12] Deep [00:49:13] Agents [00:49:13] Or [00:49:14] Just [00:49:14] Write [00:49:14] It [00:49:14] From [00:49:14] Python [00:49:15] There [00:49:17] Agent [00:49:18] Harnesses [00:49:19] That [00:49:19] Sits [00:49:20] Inside [00:49:20] Now [00:49:20] An [00:49:21] Open [00:49:21] Shell [00:49:21] Secure [00:49:23] Harness [00:49:23] Okay [00:49:24] And [00:49:25] That's [00:49:25] The [00:49:25] Sandbox [00:49:26] That [00:49:26] It [00:49:27] Sits [00:49:27] In [00:49:27] The [00:49:28] Secure [00:49:28] Sandbox [00:49:29] And [00:49:30] Then [00:49:30] Next [00:49:31] Let's [00:49:31] Talk [00:49:31] About [00:49:32] Models [00:49:33] There [00:49:34] You [00:49:35] Should [00:49:35] Use [00:49:35] Models [00:49:36] Of [00:49:36] All [00:49:36] Kind [00:49:37] In [00:49:37] The [00:49:37] Future [00:49:37] It [00:49:37] Won't [00:49:37] Just [00:49:38] Be [00:49:38] One [00:49:38] Model [00:49:38] That [00:49:39] We [00:49:39] Run [00:49:39] Everything [00:49:40] On [00:49:40] We're [00:49:41] Going [00:49:41] To [00:49:41] Run [00:49:41] That [00:49:42] We're [00:49:42] Going [00:49:42] To [00:49:42] Use [00:49:42] The [00:49:42] Best [00:49:43] Models [00:49:43] Whether [00:49:43] They're [00:49:44] Closed [00:49:44] Open [00:49:45] Or [00:49:46] Custom [00:49:46] We're [00:49:47] Going [00:49:47] To [00:49:48] Use [00:49:48] All [00:49:48] The [00:49:49] Models [00:49:49] To [00:49:49] Do [00:49:49] Your [00:49:50] Work [00:49:50] Sometimes [00:49:51] The [00:49:52] Work [00:49:52] That [00:49:52] You [00:49:52] Do [00:49:52] Is [00:49:52] Fairly [00:49:53] Generic [00:49:53] Artificial [00:49:55] General [00:49:55] Intelligence [00:49:56] Is [00:49:56] Going [00:49:56] To [00:49:57] Do [00:49:57] An [00:49:57] Incredibly [00:49:57] Good [00:49:57] Job [00:49:58] However [00:49:59] Sometimes [00:50:00] The [00:50:00] Work [00:50:00] That [00:50:01] You [00:50:01] Want [00:50:01] To [00:50:01] Do [00:50:01] Captures [00:50:02] Too [00:50:02] Much [00:50:03] Of [00:50:03] Your [00:50:03] Proprietary [00:50:03] Knowledge [00:50:04] Too [00:50:04] Much [00:50:05] Of [00:50:05] Your [00:50:05] Specialty [00:50:06] And [00:50:06] You [00:50:06] Want [00:50:06] To [00:50:07] Continuously [00:50:07] Enhance [00:50:08] That [00:50:08] Intelligence [00:50:09] Every [00:50:10] Company [00:50:11] In [00:50:11] The [00:50:11] Future [00:50:11] Will [00:50:12] Use [00:50:12] Both [00:50:13] Closed [00:50:14] Models [00:50:14] As [00:50:14] Well [00:50:15] As [00:50:15] Open [00:50:16] Models [00:50:16] That [00:50:16] You [00:50:16] Customize [00:50:17] By [00:50:18] The [00:50:18] End [00:50:18] Of [00:50:18] The [00:50:18] Session [00:50:18] You [00:50:19] Will [00:50:19] See [00:50:19] That [00:50:19] In [00:50:19] Fact [00:50:20] Building [00:50:20] Your [00:50:21] Own [00:50:21] AIs [00:50:21] Is [00:50:22] Not [00:50:22] That [00:50:22] Hard [00:50:22] It [00:50:24] Starts [00:50:24] With [00:50:24] Of [00:50:25] Course [00:50:25] The [00:50:25] Model [00:50:26] We [00:50:27] Been [00:50:27] Dedicated [00:50:27] To [00:50:27] Advancing [00:50:28] Open [00:50:28] Models [00:50:29] To [00:50:29] Open [00:50:30] Up [00:50:30] New [00:50:30] Opportunities [00:50:31] And [00:50:32] Enable [00:50:32] People [00:50:32] To [00:50:32] Build [00:50:33] Their [00:50:33] Own [00:50:33] Custom [00:50:33] AIs [00:50:34] Nemo [00:50:35] Tron [00:50:36] Is [00:50:36] For [00:50:37] Languages [00:50:37] Cosmos [00:50:38] Is [00:50:38] Our [00:50:39] World [00:50:40] Foundation [00:50:41] Model [00:50:41] We [00:50:42] Are [00:50:42] The [00:50:42] Frontier [00:50:42] Groot [00:50:44] For [00:50:45] Human [00:50:45] Robotics [00:50:46] We [00:50:46] Are [00:50:46] The [00:50:47] Frontier [00:50:47] Alpha [00:50:48] Mayo [00:50:48] For [00:50:49] Autonomous [00:50:50] Navigation [00:50:50] We [00:50:51] Are [00:50:51] The [00:50:51] Frontier [00:50:52] I [00:50:52] Sing [00:50:53] For [00:50:53] Quantum [00:50:53] Classical [00:50:54] Computing [00:50:55] We [00:50:55] Are [00:50:55] The [00:50:56] Frontier [00:50:56] And [00:50:57] Of [00:50:57] Course [00:50:57] Earth [00:50:57] Two [00:50:58] Physics [00:50:59] AI [00:50:59] We [00:50:59] Are [00:51:00] The [00:51:00] Frontier [00:51:00] So [00:51:01] In [00:51:01] All [00:51:02] Of [00:51:02] These [00:51:02] Different [00:51:02] Areas [00:51:03] We [00:51:03] Offer [00:51:04] You [00:51:04] All [00:51:05] Of [00:51:05] The [00:51:05] Models [00:51:06] And [00:51:06] We [00:51:06] Been [00:51:08] Working [00:51:08] On [00:51:08] It [00:51:08] For [00:51:08] A [00:51:08] Decade [00:51:09] And [00:51:09] We [00:51:09] Continue [00:51:09] To [00:51:10] Work [00:51:10] On [00:51:10] This [00:51:10] Forever [00:51:10] We [00:51:11] Offer [00:51:11] You [00:51:18] Created [00:51:19] As [00:51:19] Well [00:51:20] As [00:51:20] The [00:51:20] Scripts [00:51:21] Necessary [00:51:21] For [00:51:21] You [00:51:22] To [00:51:22] Reproduce [00:51:23] It [00:51:23] So [00:51:24] Not [00:51:24] Only [00:51:24] Do [00:51:24] You [00:51:24] Have [00:51:24] The [00:51:25] Open [00:51:25] Weights [00:51:25] You [00:51:26] Have [00:51:26] From [00:51:27] Us [00:51:27] The [00:51:27] Entire [00:51:28] Recipe [00:51:28] And [00:51:28] The [00:51:28] Data [00:51:29] To [00:51:29] Recreate [00:51:30] The [00:51:30] Open [00:51:31] Models [00:51:31] So [00:51:32] That [00:51:32] You [00:51:32] Could [00:51:32] Know [00:51:33] Where [00:51:33] It [00:51:33] Came [00:51:33] From [00:51:34] And [00:51:34] Have [00:51:34] Complete [00:51:35] Trust [00:51:35] Into [00:51:35] It [00:51:36] Well [00:51:37] These [00:51:37] Models [00:51:38] Are [00:51:39] Not [00:51:40] Necessarily [00:51:41] The [00:51:42] Frontier [00:51:42] Of [00:51:42] Frontiers [00:51:43] Especially [00:51:43] For [00:51:44] The [00:51:44] Language [00:51:44] Model [00:51:44] However [00:51:45] With [00:51:46] Customization [00:51:47] And [00:51:48] With [00:51:48] Partners [00:51:48] That [00:51:49] The [00:51:49] Ecosystem [00:51:50] That [00:51:50] We [00:51:50] Work [00:51:50] With [00:51:51] You [00:51:51] Could [00:51:51] Turn [00:51:51] These [00:51:52] Models [00:51:52] Into [00:51:53] Frontier [00:51:53] And [00:51:54] As [00:51:54] Result [00:51:54] Turn [00:51:55] Your [00:51:55] Propriate [00:51:56] Data [00:51:56] Into [00:51:57] Sovereign [00:51:57] Intelligence [00:51:58] Here's [00:51:59] An [00:51:59] Example [00:51:59] Of [00:51:59] A21 [00:52:00] Helping [00:52:01] Alice [00:52:01] Turn [00:52:03] Nemotron [00:52:03] Into [00:52:04] A [00:52:05] Model [00:52:05] That [00:52:05] Exceeds [00:52:06] Frontier [00:52:07] Capabilities [00:52:07] Here's [00:52:08] Base [00:52:08] 10 [00:52:08] Helping [00:52:09] Code [00:52:09] Rabbit [00:52:09] Turning [00:52:10] A [00:52:11] Nemotron [00:52:11] Model [00:52:11] After [00:52:12] Fine [00:52:13] Tuning [00:52:13] On [00:52:13] Your [00:52:13] Data [00:52:14] Turned [00:52:15] Into [00:52:15] Frontier [00:52:17] Capability [00:52:17] Base [00:52:18] 10 [00:52:18] And [00:52:18] Code [00:52:18] Rapid [00:52:19] Here's [00:52:19] Trajectory [00:52:19] And [00:52:20] Harvey [00:52:20] Again [00:52:22] Using [00:52:22] Nemotron [00:52:23] Post [00:52:24] Trained [00:52:24] It [00:52:24] And [00:52:25] Turned [00:52:25] And [00:52:26] Created [00:52:27] Their [00:52:27] Own [00:52:27] Frontier [00:52:28] Level [00:52:28] Capability [00:52:29] Here's [00:52:29] Scale [00:52:30] Working [00:52:31] With [00:52:31] Skydance [00:52:31] Again [00:52:32] Doing [00:52:32] The [00:52:32] Same [00:52:33] Thing [00:52:33] It [00:52:33] Goes [00:52:33] On [00:52:34] And [00:52:34] On [00:52:34] And [00:52:34] On [00:52:34] The [00:52:35] Point [00:52:35] Being [00:52:36] That [00:52:36] You [00:52:36] Could [00:52:37] Start [00:52:37] With [00:52:37] An [00:52:37] Open [00:52:37] Model [00:52:38] Something [00:52:38] That's [00:52:39] Incredibly [00:52:39] Good [00:52:40] Like [00:52:40] Nemotron [00:52:40] And [00:52:42] And [00:52:42] And [00:52:42] With [00:52:42] Your [00:52:42] Own [00:52:43] Data [00:52:43] Post [00:52:44] Train [00:52:45] It [00:52:45] For [00:52:45] Your [00:52:45] Use [00:52:46] Case [00:52:46] And [00:52:46] Achieve [00:52:47] Of [00:52:47] Course [00:52:48] Super [00:52:48] Intelligence [00:52:49] That [00:52:50] Is [00:52:50] A [00:52:50] Very [00:52:50] Sensible [00:52:51] Thing [00:52:51] To [00:52:51] Do [00:52:51] And [00:52:51] All [00:52:51] The [00:52:52] Technology [00:52:52] Exists [00:52:53] Today [00:52:53] Here's [00:52:54] An [00:52:54] Example [00:52:54] Of [00:52:54] Palantir [00:52:55] With [00:52:56] Tampa [00:52:56] General [00:52:56] Hospital [00:52:57] The [00:52:58] List [00:52:58] Goes [00:52:58] On [00:52:59] And [00:52:59] On [00:52:59] We [00:53:00] Have [00:53:00] A [00:53:00] Rich [00:53:00] Ecosystem [00:53:01] Of [00:53:01] Partners [00:53:01] That [00:53:01] Can [00:53:01] Help [00:53:02] You [00:53:02] And [00:53:03] They're [00:53:03] Incredible [00:53:03] Companies [00:53:04] And [00:53:04] They [00:53:04] Can [00:53:04] Work [00:53:05] With [00:53:05] You [00:53:05] You [00:53:06] Know [00:53:06] From [00:53:06] Around [00:53:06] The [00:53:06] World [00:53:07] Okay [00:53:07] So [00:53:08] We [00:53:09] It's [00:53:09] Just [00:53:10] Anyways [00:53:10] I'll [00:53:10] Mention [00:53:11] One [00:53:11] And [00:53:11] I'll [00:53:11] Offend [00:53:12] Everybody [00:53:12] Else [00:53:12] I'll [00:53:14] Have [00:53:14] To [00:53:15] Sit [00:53:15] Here [00:53:15] And [00:53:15] Mention [00:53:16] All [00:53:16] Of [00:53:16] Them [00:53:16] And [00:53:17] I [00:53:17] Love [00:53:17] Them [00:53:18] All [00:53:18] And [00:53:19] So [00:53:19] First [00:53:21] Agent [00:53:22] Harness [00:53:23] Running [00:53:24] In [00:53:24] A [00:53:24] Secure [00:53:25] Sandbox [00:53:25] Open [00:53:26] Shell [00:53:26] Second [00:53:27] Nemotron [00:53:28] Large [00:53:29] Language [00:53:29] Models [00:53:30] The [00:53:30] First [00:53:30] Two [00:53:31] Components [00:53:31] Of [00:53:31] What [00:53:32] You [00:53:32] Need [00:53:32] But [00:53:33] Ultimately [00:53:34] What [00:53:34] Turns [00:53:35] These [00:53:36] Super [00:53:36] Agents [00:53:37] These [00:53:37] Agents [00:53:37] Into [00:53:37] Super [00:53:38] Agents [00:53:38] Is [00:53:38] That [00:53:39] You [00:53:39] Give [00:53:39] Them [00:53:39] Access [00:53:40] To [00:53:40] Your [00:53:41] Company's [00:53:41] Knowledge [00:53:42] And [00:53:43] This [00:53:43] Is [00:53:44] Really [00:53:45] Special [00:53:45] This [00:53:45] Is [00:53:45] Called [00:53:46] The [00:53:46] Agent [00:53:46] Knowledge [00:53:47] Foundation [00:53:48] That [00:53:48] We [00:53:48] Put [00:53:48] Together [00:53:49] Libraries [00:53:50] And [00:53:50] Models [00:53:50] That [00:53:51] Are [00:53:51] Completely [00:53:52] World [00:53:52] Class [00:53:52] And [00:53:53] Allows [00:53:53] You [00:53:53] To [00:53:53] Turn [00:53:54] Your [00:53:54] Company's [00:53:54] Data [00:53:55] The [00:53:56] Data [00:53:57] Could [00:53:57] Be [00:53:57] In [00:53:57] The [00:53:58] Assets [00:53:58] Could [00:53:58] Be [00:53:58] In [00:53:59] Unstructured [00:54:01] Format [00:54:01] And [00:54:01] Documents [00:54:02] So [00:54:03] That [00:54:03] You [00:54:04] Understand [00:54:04] How [00:54:04] To [00:54:04] Read [00:54:05] These [00:54:05] PDFs [00:54:05] And [00:54:06] Read [00:54:06] All [00:54:06] These [00:54:06] Unstructured [00:54:07] Documents [00:54:07] Or [00:54:08] It [00:54:08] Could [00:54:08] Be [00:54:08] In [00:54:08] Structured [00:54:09] Data [00:54:09] Inside [00:54:10] Databases [00:54:10] We [00:54:11] Have [00:54:12] The [00:54:12] Libraries [00:54:12] And [00:54:13] The [00:54:13] Tools [00:54:13] Necessary [00:54:14] To [00:54:14] Help [00:54:15] You [00:54:15] Extract [00:54:16] Meaning [00:54:16] And [00:54:17] Understanding [00:54:17] From [00:54:18] Your [00:54:18] Company's [00:54:19] Data [00:54:19] One [00:54:21] At [00:54:21] A [00:54:21] Time [00:54:22] Nemo [00:54:22] Nemo [00:54:22] Nemo [00:54:22] Tron [00:54:22] Parse [00:54:23] Completely [00:54:24] World [00:54:24] Class [00:54:24] Top [00:54:25] Of [00:54:25] Leader [00:54:25] Board [00:54:25] Nemo [00:54:26] Tron [00:54:27] Embed [00:54:27] Taking [00:54:28] A [00:54:28] Query [00:54:28] And [00:54:28] Associating [00:54:29] With [00:54:30] The [00:54:30] Knowledge [00:54:31] That's [00:54:31] Inside [00:54:31] The [00:54:32] Vector [00:54:32] Database [00:54:32] Top [00:54:33] Of [00:54:33] Leader [00:54:34] Board [00:54:34] Nemo [00:54:35] Tron [00:54:35] Rerank [00:54:36] It [00:54:37] Might [00:54:37] Find [00:54:37] Many [00:54:37] Different [00:54:38] Sources [00:54:38] Where [00:54:39] The [00:54:39] Query [00:54:39] Is [00:54:40] Associated [00:54:40] With [00:54:40] Different [00:54:41] Pieces [00:54:42] Of [00:54:42] Data [00:54:42] It [00:54:43] Brings [00:54:43] It [00:54:43] Together [00:54:43] Ranks [00:54:44] Ranks [00:54:45] Them [00:54:45] Based [00:54:45] On [00:54:45] Which [00:54:45] One [00:54:45] Is [00:54:45] The [00:54:46] Most [00:54:46] Relevant [00:54:46] To [00:54:47] You [00:54:47] Kumo [00:54:49] To [00:54:50] Be [00:54:50] Able [00:54:50] To [00:54:50] Predict [00:54:51] From [00:54:52] Your [00:54:53] Structured [00:54:53] Database [00:54:54] What [00:54:54] Is [00:54:54] Likely [00:54:55] To [00:54:55] Happen [00:54:55] And [00:54:56] So [00:54:56] This [00:54:56] Is [00:54:56] For [00:54:57] Example [00:54:57] If [00:54:58] You [00:54:58] Were [00:54:58] To [00:54:58] Receive [00:54:59] A [00:55:01] Notice [00:55:01] That [00:55:02] One [00:55:03] Of [00:55:03] Your [00:55:03] Suppliers [00:55:03] Will [00:55:04] Be [00:55:04] Short [00:55:05] On [00:55:05] Supply [00:55:06] Or [00:55:06] Delay [00:55:06] On [00:55:07] Supply [00:55:07] That [00:55:08] Piece [00:55:09] Of [00:55:09] Information [00:55:09] Is [00:55:10] Likely [00:55:10] Unstructured [00:55:12] Data [00:55:12] It's [00:55:12] A [00:55:13] Document [00:55:13] The [00:55:14] Information [00:55:16] Your [00:55:16] Company [00:55:16] Are [00:55:16] Likely [00:55:17] Inside [00:55:17] A [00:55:17] Structured [00:55:18] Database [00:55:18] Your [00:55:19] SQL [00:55:20] Database [00:55:20] For [00:55:20] Example [00:55:21] And [00:55:22] In [00:55:22] Order [00:55:23] To [00:55:23] Predict [00:55:23] What [00:55:24] To [00:55:25] What's [00:55:25] Going [00:55:26] To [00:55:26] Happen [00:55:26] To [00:55:26] All [00:55:26] Of [00:55:26] These [00:55:27] Orders [00:55:27] Who [00:55:27] Are [00:55:27] Going [00:55:27] To [00:55:27] Be [00:55:28] Short [00:55:28] The [00:55:29] Amount [00:55:30] Of [00:55:30] Delay [00:55:30] You [00:55:30] Use [00:55:31] Kuma [00:55:31] To [00:55:31] Make [00:55:31] That [00:55:32] Prediction [00:55:32] And [00:55:33] In [00:55:33] Order [00:55:33] To [00:55:34] Figure [00:55:35] Out [00:55:35] What [00:55:35] To [00:55:35] Replace [00:55:36] It [00:55:36] With [00:55:36] You [00:55:37] Might [00:55:37] Use [00:55:37] Vector [00:55:37] Database [00:55:38] Search [00:55:38] And [00:55:39] To [00:55:39] Go [00:55:39] Find [00:55:39] An [00:55:46] Libraries [00:55:46] And [00:55:47] Data [00:55:47] Processing [00:55:47] Capabilities [00:55:48] Now [00:55:49] Makes [00:55:49] For [00:55:50] You [00:55:50] Your [00:55:50] Knowledge [00:55:51] Foundation [00:55:51] You've [00:55:52] Now [00:55:52] Armed [00:55:53] Your [00:55:53] AI [00:55:54] With [00:55:54] Proprietary [00:55:55] Data [00:55:55] Proprietary [00:55:56] Knowledge [00:55:57] Now [00:55:58] The [00:55:59] Last [00:55:59] Part [00:55:59] They'll [00:56:00] Talk [00:56:01] About [00:56:01] Is [00:56:01] Ultimately [00:56:01] Tools [00:56:02] Tools [00:56:03] Of [00:56:03] The [00:56:03] Trade [00:56:03] NVIDIA [00:56:05] Has [00:56:05] Been [00:56:05] Working [00:56:05] On [00:56:05] Cuda [00:56:06] X [00:56:06] And [00:56:06] For [00:56:06] Most [00:56:07] Of [00:56:07] You [00:56:07] In [00:56:08] The [00:56:08] Last [00:56:08] Before [00:56:08] AI [00:56:09] All [00:56:10] Of [00:56:10] Our [00:56:10] Conversations [00:56:11] Have [00:56:11] Been [00:56:11] Around [00:56:11] How [00:56:12] Do [00:56:12] We [00:56:12] Accelerate [00:56:13] Simulators [00:56:14] And [00:56:14] Solvers [00:56:15] Data [00:56:16] Processing [00:56:16] Engines [00:56:17] Domain [00:56:18] Specific [00:56:18] Libraries [00:56:19] Of [00:56:19] All [00:56:19] Kinds [00:56:20] They're [00:56:20] Used [00:56:20] In [00:56:21] Scientific [00:56:21] Discovery [00:56:22] In [00:56:23] All [00:56:23] The [00:56:23] Different [00:56:23] Fields [00:56:23] Of [00:56:24] Science [00:56:24] They're [00:56:25] Used [00:56:25] In [00:56:25] Engineering [00:56:26] They're [00:56:26] Used [00:56:27] In [00:56:27] Research [00:56:27] They're [00:56:27] Used [00:56:28] In [00:56:28] Designing [00:56:29] Building [00:56:30] Creative [00:56:31] Arts [00:56:31] At [00:56:32] This [00:56:32] Core [00:56:33] Are [00:56:34] Cuda [00:56:34] X [00:56:34] Libraries [00:56:35] And [00:56:36] These [00:56:36] Cuda [00:56:36] X [00:56:36] Libraries [00:56:37] Are [00:56:37] Algorithms [00:56:37] And [00:56:38] Algorithms [00:56:38] As [00:56:39] You [00:56:39] Know [00:56:39] Are [00:56:39] Beautiful [00:56:40] Let's [00:56:40] Take [00:56:40] A [00:56:40] Look [00:56:41] At [00:56:41] It [00:56:55] I'm sorry. [00:57:25] I'm sorry. [00:57:55] I'm sorry. [00:58:25] I'm sorry. [00:58:55] I'm sorry. [00:59:25] When you're building an AI assistant, it could be a lab assistant, it could be a research assistant, it could be a co-designer, somebody who's co-creator with you, building something with you. [00:59:37] Ultimately, what makes those agents useful, what makes them able to do real work are the tools. [00:59:47] And so we now have models that allow these agents to reason, understand information, reason, to be able to plan and iteratively act and use tools until it could get the job done. [01:00:02] But it still comes down to the tools. [01:00:05] And these tools we've been working on for so many years. [01:00:08] And there was one particular computer I showed you on the right-hand side that runs every single one of these tools. [01:00:14] And it could be fluid dynamics and particle physics. [01:00:16] It could be quantum chemistry. [01:00:18] It could be data processing. [01:00:19] It could be ray tracing. [01:00:22] All of these tools run on that one computer. [01:00:25] And that one computer speeds up the processing of it compared to just running it natively on a CPU by 10, 50, 100x. [01:00:35] And CUDAX tools is what give the agents the tools of the trade. [01:00:39] And it's so cool to see how other people are using it. [01:00:42] Let's take a look. [01:01:05] So this is the magic, magic behind the agents. [01:01:20] Ultimately, the agents have to use tools. [01:01:22] Now, of course, in one of the first fields of agentic use is software programming. [01:01:26] And therefore, a compiler or a Python is all really at needs. [01:01:29] But in the future, as we have these tools do more and more and more and more useful work, we're going to need to give them more tools. [01:01:37] And here's some examples. [01:01:38] So here's Lida Sciences. [01:01:39] They're creating a lab assistant. [01:01:41] And they're using NVIDIA's Bionemo models that are designed for understanding proteins and understanding cells and generating proteins and chemicals in drug discovery. [01:01:52] They're using Bionemo and CUDAX tools. [01:01:54] This is Siemens' factory, building a factory agent that uses NVIDIA Cosmos, our world foundation model that understands the physical world and CUDAX tools. [01:02:04] And here's an example of Dassault creating a virtual companion that uses NVIDIA's NemoTron, our language model, and CUDAX tools. [01:02:13] Let's see them in use. [01:02:18] The most intelligent scientist still needs a laboratory. [01:02:23] The best engineer still needs simulators and solvers. [01:02:26] An AI agent is no different. [01:02:30] An agent brings together five building blocks. [01:02:33] The model perceives reasons and plans. [01:02:37] The harness coordinates the work. [01:02:39] Knowledge Foundation gives the agent context from your company's structured and unstructured data. [01:02:44] Tools let the agent simulate, experiment, and act. [01:02:49] NVIDIA OpenShell controls what agents can access and execute. [01:02:53] Start with a frontier-closed model or customize an open model. [01:02:58] With NVIDIA's agent toolkit, any company can turn its knowledge and expertise into a customized superagent. [01:03:05] Let's look at three examples. [01:03:08] For Lila Sciences, the goal is to design a therapy for B-cell lymphoma. [01:03:13] The Lila Iris agent, built on NVIDIA NemoTron, turns that goal into an experimental plan. [01:03:20] In the computational dry lab, Lila agents use NVIDIA BioNemo agent toolkit to call models for protein design and RNA optimization, running millions of virtual experiments. [01:03:32] The most promising candidates move into Lila's AI science factory, where autonomous robots and instruments developed with NVIDIA Omniverse and Cosmos make and test them. [01:03:44] The results return to Iris, shaping the next designs and experiments. [01:03:48] The loop continues. [01:03:50] At Siemens Electronic Factory Erlangen, the goal is to bring new products into production. [01:03:56] Process Preparation X agents use NVIDIA Cosmos to generate virtual prototypes and synthetic data, then train and validate inspection models before production starts. [01:04:08] Intelligence Center X agents, built with NVIDIA's factory operations blueprint, use orders, supplies, and line setups to recommend production plans. [01:04:19] Digital Twin Composer uses NVIDIA Omniverse libraries to validate those plans with physics simulation and real-time factory data before changes reach the factory floor. [01:04:31] With the Dasso Systems Platform, an engineer needs to adapt a Volkswagen Golf door design for the Tiguan. [01:04:38] Aura and Lila use NVIDIA Nematron to reason over engineering requirements. [01:04:44] The harness coordinates the specialized sub-agents and preserves context across planning, design, and simulation. [01:04:52] Aura uses Inovia to plan the work around the vehicle's dimensions, weight limits, and other engineering requirements. [01:04:59] Leo Design Agent selects a compatible golf door design and uses Katia to reshape it for the Tiguan. [01:05:07] Leo Simulation Agent uses Simulia, accelerated by NVIDIA Cuda X libraries, to test the variants, revealing a vibration problem. [01:05:16] The results return to the Design Agent for refinement. [01:05:19] The engineer reviews the refined designs and selects the final one. [01:05:24] Different goals. [01:05:25] Different goals. [01:05:25] Different expertise. [01:05:27] The same building blocks. [01:05:29] Open models. [01:05:30] Cuda X accelerated tools. [01:05:32] And the open shell secure runtime. [01:05:35] Brought together with the NVIDIA Agent toolkit. [01:05:40] Isn't that great? [01:05:44] Some people said that AI is here, it's going to be the end of software. [01:05:50] It's going to be end of tools company. [01:05:52] It's exactly the opposite. [01:05:54] AI is here. [01:05:55] There'll be more users of these tools that we've created over time than ever before. [01:06:02] Every single tools company will now become an AI company. [01:06:05] They'll be an agent's company. [01:06:07] Not only will they offer tools, but they'll offer you specially optimized agents that are incredibly good at using their tools. [01:06:15] Caden's and CrowdStrike and DeSau, IQvius, and just about every enterprise company out there are all building agents. [01:06:22] And we're just delighted to be working with them. [01:06:25] Building agents is not that hard. [01:06:27] It's got five components. [01:06:29] It's got the model, the harness, both of them have closed and open versions, the open shell secure runtime, and then most importantly, the knowledge and tools. [01:06:40] And those knowledge and tools belong to you. [01:06:42] That's your expertise. [01:06:45] And so all of this is now possible to do, and I'll show you more in just a second. [01:06:50] But before I do that, let's talk about physical AI. [01:06:53] That's really the next wave of AI. [01:06:56] AI that understands the physical world can reason about it because the physical world is diverse. [01:07:03] It is complicated. [01:07:05] It's got extremely long tails. [01:07:07] And to reason about it and use appropriate tools to interact with the physical world. [01:07:15] Whenever people think about physical worlds, they think robotics. [01:07:17] In fact, one of my favorite physical world applications is the radio cell phone cell towers. [01:07:26] And the reason for that is this. [01:07:28] It uses, instead of cameras, instead of speakers and microphones, it uses radio waves to sense the world. [01:07:38] And to, you know, it reasons about it, it processes information, reasons about it, and interacts with the world using beamforming. [01:07:45] So, in fact, this is an application, a technology, that is just perfectly suited for AI. [01:07:53] AI will revolutionize the way we do telecommunications. [01:07:57] It will enhance our spectrum efficiency tremendously, just as we've done so with ray tracing. [01:08:03] All of the graphics that you've seen, completely 100% ray traced. [01:08:06] The amount of computation we thought necessary in the past, and now that we apply AI to, is a million to one difference. [01:08:14] And so, we're going to bring to the telecommunications network, an enormous step up in spectrum efficiency. [01:08:21] But beyond that, we're also going to revolutionize all of these cell towers. [01:08:26] There's 100,000 sites around the world, around the Western world, about 100 gigawatts of distributed data centers. [01:08:33] That will, in the future, become a distributed AI infrastructure. [01:08:38] We're delighted to partner with many companies here in Europe. [01:08:42] And just this last week, after working on this for nearly eight years, NVIDIA made the first commercial phone call on the T-Mobile network with Nokia. [01:08:53] And so, now, CUDA makes phone calls, the traditional way. [01:08:58] Then once you can do that, then we can layer AI on top of it. [01:09:02] And so, the revolution is going to begin. [01:09:04] This is a really exciting opportunity for us. [01:09:06] The other physical AI application that I'm super excited about is, essentially, robotic radiotherapy. [01:09:14] The tumor is changing in shape. [01:09:17] The physical body is changing while you're in therapy. [01:09:22] The robot, this adaptive radiotherapy system, has to perceive the changing environment, understand where the tumor is, segment it just perfectly, and direct, precisely direct the radiotherapy accordingly. [01:09:36] Essentially, robotic, that's right. [01:09:39] Go ahead. [01:09:45] This is a form of robotic surgery. [01:09:49] Now, of course, the highest volume, the most near-term robotics application or physical AI application that everybody recognized, has been something we've been working on for over a decade. [01:09:59] And now, we recognize that every single car company will want to make their cars part of the ride-hailing fleet. [01:10:09] Today, we have ride-hailing, and it's part of Uber, and here in Europe, we have Bolt, and we have Grab, and Lyft, and so on and so forth, and all these different ride-hailing services. [01:10:22] In the future, there will be robo-taxis that will be connected to those ride-hailing services as well. [01:10:28] But it won't be from just one brand. [01:10:31] And so, we recognize that building these systems are incredibly expensive, particularly the software that goes on top of it. [01:10:38] We decided to create one standard architecture, 14 different cameras, 9 radar systems, 3 lidars, umpteen ultrasonics, this cocoon of a shell of sensors powered by dual NVIDIA GPU systems, [01:10:58] for the AV computing system. [01:11:01] We turned this into an industry standard, and now it's being designed into one car fleet after another. [01:11:08] On top of it, our AV software companies, and we are, in fact, one of them as well, to build an autonomous robo-taxi capability, which then is connected into the ride-hailing services. [01:11:22] This ecosystem is being built out. [01:11:24] The world travels about a trillion miles per year around the world. [01:11:28] Many of those miles in the future, you know, a dollar or two per mile, are going to be on robo-taxis. [01:11:35] So, this is one giant opportunity, but it's incredibly fragmented. [01:11:41] And so, creating an industry standard, Hyperion, and working with an entire ecosystem, we can create, essentially, a giant virtual fleet that everybody can connect into. [01:11:52] Today, I'm super proud to announce that four more brands, four more cars are going to, lines are going to come in. [01:12:00] The VW van, I just can't imagine anything more iconic. [01:12:05] The JLR brand new car, the Volvo fleet, as well as Mercedes next generation, are going to be also Hyperion compatible, [01:12:15] and be part of this network, this ride-hailing network, robo-taxi fleet. [01:12:20] So, welcome to Hyperion. [01:12:28] The next major industry, the next major industry, the next great industry, and this one is likely to be extremely large, is human or robotics. [01:12:38] There's a reason why robotics is hard to do, because all the different work is rather diverse. [01:12:44] And so, building a special configuration for each one of the diverse tasks segments the robotics industry into all of these different configurations, [01:12:55] which prevents the technology flywheel to happen. [01:12:58] The reason why the cell phone has advanced so quickly, the reason why personal computers have advanced so quickly, [01:13:03] The reason why NVIDIA's NVLink 72 has advanced so quickly, is because those standardized configurations allows many customers to use it, [01:13:12] the volume to therefore be derived, that volume drives the technology flywheel, and it gets faster, better, incredibly fast. [01:13:22] And so, we need a robotics configuration that is as standard as possible. [01:13:29] Now, human or robotics obviously has the ability, it probably isn't perfect for everything, but it's likely to work for many things. [01:13:37] As a result, this is likely going to be the largest volume robotics configuration. [01:13:43] Now, if we could just create the AI that allows it to learn very quickly, then for small-medium businesses, [01:13:50] who don't have an entire robotics team to teach and program the robot, can now show the robot what needs to be done. [01:13:58] The robot, because it has general capabilities and general intelligence, would be able to learn and adapt quickly. [01:14:04] And so, this is likely to be the next major platform, technology platform. [01:14:10] We're working with so many different companies, you will recognize many of them here in Europe. [01:14:14] This is your great opportunity, no country, no region has dedicated itself to making megatronics, [01:14:23] electronics, mechanical systems at the level of precision and craft as the European region. [01:14:30] So many companies here are incredibly good at manufacturing. [01:14:34] This is going to be the next high-volume industry, and you could see there are so many companies that are now part of it, [01:14:40] and we're delighted to be partnered with you. [01:14:51] Now, in order to create a physical AI system, you need three computers. [01:14:59] The first computer is the one we've been talking about, and everybody knows quite well, which is training the AI model. [01:15:05] The second computer is where we have the AI learn how to do its skills. [01:15:11] You can't just have the AI, program the AI, and have the AI come out into the physical world and try to do whatever it wants to do. [01:15:18] You do it inside a digital twin. [01:15:21] We create a digital twin called Omniverse, and you've seen us talk about that repeatedly. [01:15:26] And in fact, just now, when you saw all the CUDAX libraries, they're all being done inside a digital twin world. [01:15:33] That digital twin world is called Omniverse. [01:15:36] Omniverse is a library that's connected into all kinds of platforms and tools, Siemens, and Dassault, and Cadence, and so on and so forth. [01:15:43] And so you could see these digital twin systems being created. [01:15:47] The second computer. [01:15:49] The third computer is the robotics computer. [01:15:53] That computer integrates, in the case of the self-driving car, the robo-taxi, we call it Hyperion. [01:15:59] Hyperion consists of the computer itself, the functional safety system, the active safety system, an operating system designed for physical AI safety. [01:16:13] Well, we need to do exactly the same for robotic systems so that every single time you start to develop research and start to work on your own robotic systems, [01:16:22] you don't have to start from scratch to cobble up all of the software and computing systems and the sensors and all the mechanical systems. [01:16:30] That alone would take several years. [01:16:33] And so we decided to build a reference robot. [01:16:36] This is the NVIDIA Agile One in partnership with Agile Robotics. [01:16:41] This robot is going to be available in the first quarter of next year, and it will come pre-assembled with computer inside, software integrated, [01:16:59] and it should be able to perform quite well right out of the bat. [01:17:03] And so this is a great starting point, and it will be the flywheel by which we can accelerate development of robotics. [01:17:09] Well, let me start, let me end today's speech from where we started. [01:17:15] NVIDIA started in the PC revolution. [01:17:18] It was in 1993, and the PC revolution was just about to start. [01:17:23] Remember, Windows 95 kicked it into turbo charge. [01:17:26] We came since then, invented the GPU, programmable shading, real-time ray tracing. [01:17:32] We added CUDA to the computing model and helped ignite the modern AI revolution. [01:17:40] And so four years ago, we started to think, what can we do to reimagine the personal computer for this new era? [01:17:48] Where the personal computer is not just your tool, your personal computer tool, but your personal agent or your personal assistant. [01:17:57] And the way you interact with this computer would be essentially a companion, somebody who helps you do work. [01:18:06] And so this computer needs to have exactly the five things that I mentioned today. [01:18:13] It needs to have model. [01:18:15] It could be running locally or it could be running in the cloud. [01:18:18] It needs a harness. [01:18:20] It needs a secure runtime, open shell. [01:18:24] And it needs access, because it's available on your PC, it needs access to knowledge and tools. [01:18:31] And the good news is, because NVIDIA has been accelerating tools for so long, if we could just create a computer that can accelerate OpenGL and DirectX and everything CUDA, everything NVIDIA has ever worked on would run on this computer. [01:18:48] And so this is it. [01:18:49] After four years of partnership with Microsoft, we have created the RTX Spark. [01:18:54] And this is a brand new line of computers, this every, literally thousands of years, thousands of engineering years went into this. [01:19:01] We've been secretly working on it for four years, created a chip called the N1X, and then we just recently announced it with the entire PC industry joining forces. [01:19:12] We created desktops, laptops, and workstations. [01:19:16] This is one petaflops, one petaflops, just to put that in perspective, in the year 2016, I delivered one petaflops and 128 gigabytes of memory, one petaflops and 128 gigabytes to a small company called OpenAI. [01:19:35] This is one petaflops and 128 gigabytes. [01:19:48] The only difference, of course, is that I can hold this like this, and DGX1 took two people to carry upstairs. [01:19:56] And so this is the brand new RTX Spark PC. [01:19:59] And if you would like to have a companion workstation that runs Katia and everything, the cell and everything Autodesk and, you know, Adobe and Cadence and Synopsys, you can always get yourself a DGX station that runs Windows. [01:20:15] And this is 20 petaflops and 754 gigabytes of memory. [01:20:23] And so incredible, and of course, there's also the desktop version. [01:20:27] All of these are now, these are now available, and I think the DGX station with RTX Spark, with DGX Spark, DGX station with Windows will be available at the end of this year. [01:20:39] Now, the question is, so far I just talked about the five components of an agent, and now the question is, how do you build it? [01:20:48] How do you build an agent that runs, let's say, on one of those computers, or runs in a cloud, or runs on-prem somewhere? [01:20:55] Well, the answer is, it's really easy, and you do it this way. [01:20:59] The way you build an agent is you ask an agent to build it for you. [01:21:11] So this, this is the prompt that is going to run the next application, and what you're about to see is this. [01:21:22] You're about to see a robotics application. [01:21:25] And while you're watching this demo, just remember, there's an agent running on this personal computer. [01:21:32] This demo runs across three computers. [01:21:36] The first computer is to train. [01:21:38] The second is a digital twin. [01:21:41] And the third is to put that AI into the computer itself, okay? [01:21:48] So two very, two ideas. [01:21:50] The first idea is agents, how do you build it and run it, and what you can use it to do? [01:21:56] And the second is physical AI computing model, which consists of three computers. [01:22:02] Let's take a look. [01:22:14] Meet MicroDuck, hugging face's robot. [01:22:17] Let's teach it to see and walk with NVIDIA RTX Spark. [01:22:23] I use Hermes as my agent harness. [01:22:26] Nemotron and OpenAI's GPT-6 Astra as my models. [01:22:30] And CUDA-X accelerated tools FreeCAD, Blender, and Isaac Sim, Newton Physics Engine, and Isaac Lab. [01:22:37] I'll run everything on my RTX Spark, except when I need to burst larger workloads to cloud. [01:22:43] First, I describe the course and what my MicroDuck needs to learn. [01:22:48] On RTX Spark, Hermes and Nemotron plan the course, listing every object required. [01:22:57] Astra takes it from there. [01:22:59] Using FreeCAD to model 3D objects. [01:23:03] Blender to apply materials. [01:23:06] And Omniverse SimReady libraries to add physical properties. [01:23:11] In Isaac Sim, Astra composes the scene. [01:23:18] Astra rigs MicroDuck's joints, actuators, and links. [01:23:25] And simulates how MicroDuck moves with Newton Physics Engine. [01:23:29] Now, training begins. [01:23:33] In the cloud, thousands of MicroDucks train in parallel in Isaac Lab. [01:23:43] Every fall, a lesson. [01:23:48] Improving the policy every time. [01:23:50] Omniverse Libraries simulate MicroDuck's cameras, teaching it to see. [01:23:56] With perception and physics working together, my MicroDuck can avoid obstacles, navigate tight spaces, and adapt to its surroundings. [01:24:02] I download the policy to RTX Spark, and use Nemotron and Isaac Sim to evaluate my MicroDuck with multi-robot orchestration. [01:24:22] I deploy the validated policy to the real duck, and he's on his way. [01:24:30] NVIDIA RTX Spark, build and run AI agents and their accelerated tools, right on your laptop. [01:24:40] Hey, guys. [01:24:49] Oh, yeah? [01:24:52] Okay, come on, here, come over here, come over here, come over here. [01:24:57] Okay, come on, here, come over here, show people you can walk, and you have a camera, oh, yeah? [01:25:09] Don't trip, can you guys sit down? [01:25:17] Okay, stand up, take a nap, take a nap. [01:25:31] Take a nap. [01:25:33] You need to take a nap. [01:25:36] Okay, stand up. [01:25:39] Did you know that you were created in here? [01:25:42] You were created on the RTX Spark, I know, and trained you in the cloud on a whole bunch of NVIDIA supercomputers. [01:25:51] And you learned your skills right here on RTX Spark in a digital twin, the Omniverse digital twin. [01:25:57] Isn't that incredible? [01:26:00] I know, I know, you guys are so cute. [01:26:05] And your parents are called Polen, P-O-L-L-E-N, is that right? [01:26:14] Uh-huh. [01:26:15] Polen Robotics. [01:26:17] And apparently, everybody loves you guys so much, you're sold out. [01:26:22] I know, so if you go to the Polen website, everybody, everybody watching this, you know, there's going to be like 20 million viewers, everybody watching this, go to the Polen website and place an order. [01:26:35] I think the startup company is like 10 people, they might blow up, poof, it's going to be so exciting. [01:26:43] Everybody should have one. [01:26:45] You guys are so cute. [01:26:47] And so we'll just have an RTX Spark and we'll teach you new skills. [01:26:49] Isn't that right? [01:26:50] Oi, oi, oi, oi, oi, oh, look at that. [01:26:54] That's incredible. [01:26:55] Can you do a somersault? [01:26:57] That's so good. [01:27:09] That's so good. [01:27:12] Wow. [01:27:13] Okay. [01:27:14] Well, guess what? [01:27:16] Okay. [01:27:17] So today, what did we talk about? [01:27:19] We talked about Europe building its AI economy. [01:27:24] And it starts with infrastructure. [01:27:26] And we announced a brand new, brand new system called DGX Pod. [01:27:31] It makes it easy for you to build AI factories. [01:27:33] And these AI factories don't just run models anymore. [01:27:38] Pay attention. [01:27:40] Stay focused. [01:27:41] Come here. [01:27:42] Come here. [01:27:43] Come over here. [01:27:43] Stay focused. [01:27:45] And these AI factories don't just run. [01:27:47] There you go. [01:27:47] Sit down. [01:27:48] Yeah. [01:27:49] They don't just run AI models anymore. [01:27:52] They run entire agents. [01:27:55] And these agents are consistent of five things. [01:27:57] There's the large language model that does the perception, reasoning, planning, and acting. [01:28:02] And it runs in conjunction with a harness that turns it, it's essentially an exoskeleton. [01:28:09] And that harness that harness runs inside a secure runtime we call open shell, which has just so many people that have supported it. [01:28:18] And you give it access to knowledge and tools. [01:28:23] And this agent runs on top of these computers, but we're also extending into the next generation, which is where you guys come in, physical AI. [01:28:32] And you give it access to AI. [01:28:33] So that's it. [01:28:34] That's what we talked about. [01:28:35] Everybody can build agents. [01:28:37] Every company in Europe will become an AI company. [01:28:41] There's no question everybody will become an AI company. [01:28:45] And we're going to have a thriving AI economy here in Europe. [01:28:49] So after 33 years, what an incredible journey, and we made a really cool video. [01:28:57] You guys want to watch it together? [01:28:58] Okay. [01:28:59] Come on. [01:28:59] Come over here. [01:29:00] Just come over here. [01:29:01] Just a little bit. [01:29:02] We'll get out of the way. [01:29:03] Come here. [01:29:04] Top speed. [01:29:06] Top speed. [01:29:11] How the heck are we supposed to follow that? [01:29:14] Well, there is one thing that can make this crowd go even crazier. [01:29:20] What is it? [01:29:26] You won't believe this. [01:29:28] GTC has just gone full Eurovision. [01:29:34] The keynotes done in Berlin, AI's next chapter begins. [01:29:39] Agents working by your side, Europe's moment has arrived. [01:29:44] But if you missed what we said at GTC, we'll break it down for you, Germany. [01:29:53] This is a story all about how the world's building new in for now. [01:29:59] First power plants let energy scale, then came data centers. [01:30:03] Now we've got mail. [01:30:04] Now we're way past turning coal into watts. [01:30:07] AI factories turn sand into thawks. [01:30:10] Move over oil, the S&P too. [01:30:13] There's something more investable, our GPU. [01:30:16] It works like an asset producing green. [01:30:19] Looks like a chip, acts like a cash machine. [01:30:21] They make revenue while you relax. [01:30:24] The cash flow smoother than it says. [01:30:42] Agents code and make decisions. [01:30:44] Give 'em freedom but control their missions. [01:30:47] Open shell on bearers. [01:30:48] Checking the list, no credentials. [01:30:51] Access dismissed. [01:30:52] Sentry on Bluefield never clocks off. [01:30:55] Protecting the sandbox like Hasselhoff. [01:30:58] We advance AI but limit risks. [01:31:00] We're responsibly optimistic. [01:31:03] RTX spark keeps agents safe and bright. [01:31:06] Our spot rates keep the price just right. [01:31:09] We do quantum rex, this game ain't all in one. [01:31:12] 'Cause you can't spell fungible without fun. [01:31:17] AI's got a body, now it's learning the rules of gravity friction. [01:31:22] And using real tools. [01:31:23] With Isaac Lab and Cosmos, they train on their own. [01:31:25] It's like college for robots without student loans. [01:31:29] DSX pods keep AI in sync. [01:31:31] Less power, more brains like Legos that think. [01:31:34] QVS name, Atron and Kumo are here. [01:31:37] Throwing knowledge faster than a Berlin cashier. [01:31:40] Europe's full stack, it's all on it's checklist. [01:31:43] AI's five layer cake, it eats it for breakfast. [01:31:46] The Euroracle system is leading the race. [01:31:48] In love's open source, just ask, hug and fade. [01:31:51] The continent is rising and the best is yet to come. [01:31:57] Come on, all of Europe, let's sing as one! [01:32:07] We are Europe, we are Europe, we are Europe, we are Europe, we are Europe, we are Europe, we are Europe, we are Europe! [01:32:27] You and me, welcome all to GTC! [01:32:50] Welcome to GTC! [01:32:53] I want to thank all of you, all the sponsors, you can see all the names of companies that are here, all the partners we've been working with for 33 years, MicroDuck, my new friends. [01:33:08] Let's work together to build Europe's AI economy and put AI to work. [01:33:15] All right, have a great GTC! [01:33:17] Thank you! [01:33:20] Thank you! [01:33:21] Thank you! [01:33:23] Thank you! [01:33:25] Thank you! [01:33:27] Thank you! [01:33:28] Thank you! [01:33:54] Thank you! [01:33:56] Thank you!