# NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang

## Abstract

NVIDIA founder and CEO Jensen Huang takes the stage to share his vision for the technologies and breakthroughs that will define what comes next.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:09 – 00:00:13]** This is how intelligence is made

**[00:00:14 – 00:00:22]** A new kind of factory, generator of tokens, building blocks of AI.

**[00:00:26 – 00:00:29]** Tokens have opened a new frontier,

**[00:00:29 – 00:00:32]** turning data into knowledge

**[00:00:32 – 00:00:35]** and knowledge into action.

**[00:00:38 – 00:00:44]** They decipher nature's code to reveal hidden cures

**[00:00:48 – 00:00:52]** Mirror our cities

**[00:00:57 – 00:01:01]** And tirelessly navigate our world.

**[00:01:01 – 00:01:04]** To keep us moving

**[00:01:06 – 00:01:10]** Tokens learn from the way we move

**[00:01:18 – 00:01:21]** Deliver the care we need.

**[00:01:24 – 00:01:28]** We walk into danger

**[00:01:30 – 00:01:34]** To protect the ones we love.

**[00:01:34 – 00:01:38]** They see what we cannot

**[00:01:41 – 00:01:48]** So the steadiest hands never work alone.

**[00:01:48 – 00:01:51]** Every idea tokens spark

**[00:01:53 – 00:02:00]** Every success they share puts what matters within reach.

**[00:02:09 – 00:02:14]** Tokens are helping us break new ground

**[00:02:15 – 00:02:19]** On a scale never seen before

**[00:02:24 – 00:02:26]** So we can reach

**[00:02:29 – 00:02:33]** Of the sky and beyond.

**[00:02:40 – 00:02:45]** Together, we take the next great leap into a bright,

**[00:02:45 – 00:02:52]** new future built for all mankind.

**[00:02:58 – 00:03:03]** And here is where it all begins.

**[00:03:15 – 00:03:21]** Welcome to the stage, NVIDIA founder and CEO, Jensen Wang.

**[00:03:28 – 00:03:31]** Hello, Berlin!

**[00:03:37 – 00:03:39]** Welcome to GTC.

**[00:03:39 – 00:03:41]** It's great to be here.

**[00:03:41 – 00:03:48]** Before we start, let's give it up for our pre-game show hosts.

**[00:03:58 – 00:04:01]** Smart and adorable, smart and adorable.

**[00:04:01 – 00:04:05]** They spoke about AI infrastructure, sovereign

**[00:04:05 – 00:04:09]** AI, AI safety, and physical AI.

**[00:04:09 – 00:04:13]** They covered an enormous amount of ground this morning, and we

**[00:04:13 – 00:04:15]** are gonna take it one step further.

**[00:04:16 – 00:04:18]** It's really interesting.

**[00:04:18 – 00:04:26]** What do L'Oréal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche,

**[00:04:26 – 00:04:31]** Nestle, and HSBC have in common?

**[00:04:32 – 00:04:35]** Well, they're all European companies, and they're all

**[00:04:35 – 00:04:39]** working with NVIDIA on AI.

**[00:04:39 – 00:04:46]** Everybody, everybody here is all activated on the AI ecosystem

**[00:04:46 – 00:04:50]** and working towards the AI economy.

**[00:04:50 – 00:04:53]** We've been working here in Europe now for 33 years.

**[00:04:53 – 00:04:56]** You probably don't realize

**[00:04:56 – 00:05:03]** that NV1, NVIDIA's first chip, MV2, And NV3, which became

**[00:05:03 – 00:05:07]** Riva 128, the chip that saved our company, really, and launched our

**[00:05:07 – 00:05:11]** company, were all built in Europe.

**[00:05:11 – 00:05:14]** Our first partner was SGS Thompson.

**[00:05:14 – 00:05:19]** We manufactured a chip in Creaux, France, assembled it in Malta,

**[00:05:19 – 00:05:21]** and shipped it around the world.

**[00:05:21 – 00:05:26]** Our first largest customer, the largest customer of GeForce

**[00:05:26 – 00:05:31]** days in the beginning, was a company called MediaMark.

**[00:05:31 – 00:05:34]** Do you know them?

**[00:05:36 – 00:05:41]** MediaMark had a spectacular idea about building PCs, as

**[00:05:41 – 00:05:45]** you remember, and they were our largest customer.

**[00:05:45 – 00:05:49]** It was ELSA, a German company, that brought NVIDIA into the workstation

**[00:05:49 – 00:05:50]** industry in the year 2000.

**[00:05:50 – 00:05:55]** Now NVIDIA is in every single workstation all

**[00:05:56 – 00:05:58]** over Europe, designing cars.

**[00:05:58 – 00:06:03]** Everything from styling, design, engineering, simulation, all the

**[00:06:03 – 00:06:09]** way through to even inside the car, which a European company brought

**[00:06:09 – 00:06:10]** us into, Audi, in the year 2005.

**[00:06:11 – 00:06:16]** Then in year 2020, we started working on autonomous

**[00:06:17 – 00:06:18]** vehicles with Mercedes.

**[00:06:19 – 00:06:23]** As you could see, NVIDIA's life journey, the entire 33 years, I've

**[00:06:23 – 00:06:26]** spent a lot of time in Germany.

**[00:06:27 – 00:06:28]** So I'm delighted to be back.

**[00:06:28 – 00:06:30]** Thank you.

**[00:06:33 – 00:06:36]** I'm delighted to do GTC here.

**[00:06:36 – 00:06:41]** This year, this last year, we started the new journey to building

**[00:06:41 – 00:06:46]** a new industry here in Europe, the beginnings of the AI economy.

**[00:06:47 – 00:06:50]** In the last year alone, we've invested nearly $10

**[00:06:50 – 00:06:55]** billion in Europe across startups, big partnerships, small

**[00:06:55 – 00:06:59]** partnerships, all over Europe.

**[00:06:59 – 00:07:05]** This last week alone, as I journeyed from France and UK

**[00:07:05 – 00:07:09]** to Sweden and now here in Berlin, we announced the investments

**[00:07:09 – 00:07:12]** of 14 start-up companies.

**[00:07:12 – 00:07:18]** Overall, since the beginning, we now have 12,000-plus partnerships

**[00:07:18 – 00:07:24]** across Europe, and we're all working towards one goal now.

**[00:07:25 – 00:07:29]** Decades after decades of advancement in the computer

**[00:07:29 – 00:07:33]** industry, we are now at the beginnings of a new

**[00:07:33 – 00:07:35]** industrial revolution.

**[00:07:35 – 00:07:37]** This image was generated by a computer.

**[00:07:37 – 00:07:41]** It was prompted by me; nice simple prompt about

**[00:07:41 – 00:07:46]** ai powering a new industrial revolution generated this.

**[00:07:46 – 00:07:50]** I gave it a few partners that i wanted to highlight, companies

**[00:07:50 – 00:07:53]** that we're working with, and this is an extraordinary thing.

**[00:07:53 – 00:07:57]** In the mind of the ai, this is what it means to build

**[00:07:58 – 00:08:03]** a new industry, generation of tokens, the produce production of

**[00:08:03 – 00:08:08]** intelligence, which integrates into every single industry, powering

**[00:08:08 – 00:08:10]** a new industrial revolution.

**[00:08:10 – 00:08:18]** A new industrial revolution happens when the new technology

**[00:08:18 – 00:08:22]** affects how we build something

**[00:08:22 – 00:08:27]** and what we can build, how we build something and what we can build.

**[00:08:27 – 00:08:32]** I think in every single way, artificial intelligence will affect

**[00:08:32 – 00:08:36]** everything we build and, in fact, enable us to build new things.

**[00:08:36 – 00:08:40]** In this way, because everyone will use it, every company

**[00:08:40 – 00:08:44]** will be powered by it, because it's digital intelligence,

**[00:08:44 – 00:08:47]** the foundation of everything.

**[00:08:47 – 00:08:49]** Every single person will use it, every company will be

**[00:08:49 – 00:08:53]** powered by it, every industry will be transformed, and new

**[00:08:53 – 00:08:55]** ones will be created.

**[00:08:55 – 00:09:01]** In this way, like the electricity, it is a new industrial revolution.

**[00:09:01 – 00:09:04]** And we're at the very beginnings of that.

**[00:09:04 – 00:09:08]** The question is, what is AI?

**[00:09:08 – 00:09:11]** And this is a big deal.

**[00:09:11 – 00:09:15]** If you want to be part of creating this future, understanding

**[00:09:15 – 00:09:18]** what is AI is vitally important.

**[00:09:18 – 00:09:22]** And this is really a simple framework.

**[00:09:22 – 00:09:25]** We were early to observe that, in fact, AI is a

**[00:09:25 – 00:09:27]** manufacturing industry.

**[00:09:27 – 00:09:30]** AI is not just software technology.

**[00:09:30 – 00:09:32]** It is software technology, but not just.

**[00:09:32 – 00:09:36]** AI, of course, runs on a computer, but the way that

**[00:09:36 – 00:09:39]** that computer operates is fundamentally different.

**[00:09:39 – 00:09:46]** Rather than a tool that's used by us, it is now part of a machinery,

**[00:09:46 – 00:09:51]** an infrastructure, factories that are producing intelligence,

**[00:09:51 – 00:09:56]** that are integrated into every single industry and application.

**[00:09:56 – 00:10:00]** A simple model is to think of AI as a five-layer cake.

**[00:10:00 – 00:10:07]** Energy powers advanced computers that are part of an AI factory

**[00:10:07 – 00:10:11]** that generates tokens, digital intelligence from the models

**[00:10:11 – 00:10:16]** that run on it, that are integrated into every application.

**[00:10:16 – 00:10:18]** A five-layer cake.

**[00:10:18 – 00:10:21]** To be part of this AI economy, you don't want to just be

**[00:10:21 – 00:10:24]** consuming the AI.

**[00:10:24 – 00:10:27]** You want to be producing the AI.

**[00:10:27 – 00:10:29]** We could do all of that here in Europe.

**[00:10:29 – 00:10:32]** And in fact, we've already started.

**[00:10:32 – 00:10:37]** Europe is building its AI economy and its five-layer

**[00:10:37 – 00:10:38]** cake represented here.

**[00:10:39 – 00:10:40]** We've got great partnerships.

**[00:10:40 – 00:10:42]** And when you look across this five-layer cake, you could

**[00:10:42 – 00:10:47]** recognize that, in fact, it is a new industry because There's

**[00:10:47 – 00:10:49]** growth at every single layer.

**[00:10:50 – 00:10:54]** We're creating jobs, creating growth at extraordinary rate.

**[00:10:54 – 00:10:59]** Companies here in Europe, Schneider, Siemens Electric,

**[00:10:59 – 00:11:03]** incredible Siemens Energy, doing incredibly well around the world.

**[00:11:03 – 00:11:05]** We're the computing layer.

**[00:11:05 – 00:11:08]** We integrate into AI factories that are now

**[00:11:08 – 00:11:10]** flourishing all over Europe.

**[00:11:10 – 00:11:14]** Startup companies like Nibia is doing incredibly well.

**[00:11:14 – 00:11:18]** A new company called Firebird, which is setting up a new

**[00:11:18 – 00:11:25]** cloud in the Central Asia region, and even Cassava in Africa.

**[00:11:25 – 00:11:29]** Every single telco in the future, and I predict this,

**[00:11:29 – 00:11:34]** every single telco in the future, with 100,000 sites representing

**[00:11:34 – 00:11:37]** some 100 gigawatts of distributed

**[00:11:37 – 00:11:41]** Telco base stations and data centers, will in the future evolve

**[00:11:41 – 00:11:45]** into a distributed AI network.

**[00:11:45 – 00:11:49]** So the AI infrastructure layer is being built out, which

**[00:11:49 – 00:11:51]** is then, above that, the AI models.

**[00:11:51 – 00:11:54]** It's really important to recognize that long before

**[00:11:54 – 00:11:59]** the AI models emerge, the infrastructure has to exist.

**[00:11:59 – 00:12:01]** Just as before you couldn't have an internet, you have

**[00:12:01 – 00:12:02]** to have the networking.

**[00:12:02 – 00:12:06]** Just as before you could really enjoy the capabilities

**[00:12:06 – 00:12:10]** of electricity, you have to have electrified houses,

**[00:12:10 – 00:12:12]** electrified power plants.

**[00:12:12 – 00:12:15]** And so the models are the next layer.

**[00:12:15 – 00:12:19]** It's also interesting to observe that the model layer is not

**[00:12:19 – 00:12:21]** just language models.

**[00:12:21 – 00:12:27]** It represents language, biology, chemicals, physics, robotic

**[00:12:27 – 00:12:32]** actions, animation, and it represents models of all sizes.

**[00:12:32 – 00:12:35]** These different industries are going to be powered by

**[00:12:35 – 00:12:39]** different modalities of AI models, which then integrate

**[00:12:39 – 00:12:42]** into The AI natives, the startups.

**[00:12:42 – 00:12:44]** And there are so many here in Europe.

**[00:12:45 – 00:12:48]** So the Europe's AI economy is starting, and it's

**[00:12:48 – 00:12:50]** really accelerating.

**[00:12:50 – 00:12:52]** It's great to see.

**[00:12:52 – 00:12:56]** At the foundation of it, where we contribute, most

**[00:12:56 – 00:12:58]** people think is GPUs.

**[00:12:58 – 00:12:59]** It is true.

**[00:12:59 – 00:13:03]** We started In 1993, building a graphics chip, which

**[00:13:03 – 00:13:07]** emerged into the GPU, the programmable shading GPU.

**[00:13:07 – 00:13:11]** Over time, we added CUDA acceleration to that, reinvented

**[00:13:12 – 00:13:15]** computing as we know it, which then brought forth

**[00:13:15 – 00:13:17]** artificial intelligence.

**[00:13:17 – 00:13:23]** But NVIDIA today is really a full-stack, end-to-end,

**[00:13:23 – 00:13:27]** complete AI factory platform.

**[00:13:27 – 00:13:31]** The breadth of the technology of our company over the years

**[00:13:31 – 00:13:32]** has really expanded.

**[00:13:32 – 00:13:37]** Everything from GPU computing, Vera Rubin is our latest generation,

**[00:13:37 – 00:13:41]** to the NVLink switches that connect them into one giant

**[00:13:41 – 00:13:46]** GPU, the Vera BlueField 4 compute that we recently introduced,

**[00:13:46 – 00:13:51]** and I'll say a few words about that later, to Spectrum X, which

**[00:13:52 – 00:13:55]** is our ethernet solution, making NVIDIA the largest networking

**[00:13:55 – 00:14:02]** company in the world today, to Groq3, a decode accelerator for

**[00:14:02 – 00:14:07]** generation, for token generation, to Vera4, Vera BlueField 4

**[00:14:07 – 00:14:11]** storage acceleration, I'll say a few words in that in just a second,

**[00:14:11 – 00:14:15]** and then Vera RTX compute, and I'll mention that in just a second.

**[00:14:15 – 00:14:21]** But this now represents the breadth of technology necessary

**[00:14:21 – 00:14:24]** to create the AI factory.

**[00:14:24 – 00:14:29]** But nothing works unless you have software that activates it all.

**[00:14:29 – 00:14:34]** The first part is the system software that runs the AI factory.

**[00:14:34 – 00:14:36]** These systems are gigantic.

**[00:14:36 – 00:14:41]** A one gigawatt data center, one gigawatt AI factory, costs

**[00:14:41 – 00:14:43]** about $50 billion.

**[00:14:43 – 00:14:49]** Land, power, shell, all of the compute, networking, storage,

**[00:14:49 – 00:14:55]** power, cooling, all connected together, multiple acres.

**[00:14:56 – 00:15:00]** Just turning it on and operating it is a marvel in itself.

**[00:15:00 – 00:15:07]** The magnificent amount of software necessary to energize this factory,

**[00:15:07 – 00:15:12]** to put it to use and produce tokens, is an extraordinary feat.

**[00:15:12 – 00:15:15]** NVIDIA offers the entire full-stack software.

**[00:15:15 – 00:15:18]** On top of that are what we call the CUDA X libraries,

**[00:15:18 – 00:15:20]** the algorithms of our company.

**[00:15:20 – 00:15:25]** What NVIDIA truly is, is an algorithm company that designs

**[00:15:25 – 00:15:27]** silicon to accelerate them.

**[00:15:27 – 00:15:31]** Of course, one of them is cuDNN, what many researchers

**[00:15:31 – 00:15:37]** recognize as the library that brought forth That started

**[00:15:37 – 00:15:40]** the big bang of modern AI, QDNN.

**[00:15:40 – 00:15:46]** It's QDNN is to AI what SQL is to the world's database industry.

**[00:15:46 – 00:15:51]** It's a domain-specific language, a domain-specific library.

**[00:15:51 – 00:15:57]** Just as OpenGL and DirectX enabled real-time computer graphics, QDNN

**[00:15:57 – 00:15:59]** enabled the AI computing industry.

**[00:15:59 – 00:16:03]** 350 other libraries.

**[00:16:04 – 00:16:08]** Those libraries are, in fact, the crown jewels of our company.

**[00:16:08 – 00:16:12]** Everything from computational lithography to signal processing,

**[00:16:12 – 00:16:19]** image processing, data processing, particles, fluids, sparse solvers.

**[00:16:19 – 00:16:24]** Incredible breadth of libraries that many of you in the audience

**[00:16:24 – 00:16:26]** have connected into your company.

**[00:16:26 – 00:16:28]** This is how we open new markets.

**[00:16:28 – 00:16:34]** This is how we connect NVIDIA into algorithms that enable you to do

**[00:16:34 – 00:16:37]** computation at incredible speeds.

**[00:16:37 – 00:16:44]** And then finally, on the right-hand side, on that side, the models.

**[00:16:44 – 00:16:50]** This is the new software, the new engine of new applications.

**[00:16:50 – 00:16:52]** And it sits on top of and it collaborates with the

**[00:16:52 – 00:16:53]** CUDA X libraries.

**[00:16:53 – 00:16:58]** These models, NVIDIA had been working on across multiple domains.

**[00:16:58 – 00:17:00]** With these models, it opens up new opportunities.

**[00:17:00 – 00:17:03]** And I have lots to say about that later on.

**[00:17:03 – 00:17:10]** Well, every single generation, NVIDIA started with Pascal.

**[00:17:10 – 00:17:18]** The AI era started with the Pascal new What is called DGX-1 system.

**[00:17:19 – 00:17:24]** DGX-1 brought to the world the world's first AI supercomputer.

**[00:17:24 – 00:17:31]** I introduced it 10 years ago, delivered the first one to OpenAI,

**[00:17:31 – 00:17:34]** started their computation journey.

**[00:17:34 – 00:17:41]** From Pascal to Volta to Ampere, then to many of you who heard

**[00:17:41 – 00:17:44]** about us for the very first time, Hopper, the current

**[00:17:44 – 00:17:50]** generation, Blackwell, and then now, Vera Rubin in full production.

**[00:17:50 – 00:17:53]** Notice six generations of architectures.

**[00:17:53 – 00:17:58]** With each generation, we advance the technology.

**[00:17:58 – 00:18:01]** In the course of that last 10 years, we've improved AI

**[00:18:01 – 00:18:04]** performance by one million times.

**[00:18:04 – 00:18:06]** If you would take a step back and think about what would

**[00:18:06 – 00:18:11]** Moore's Law have done in 10 years' time, the first five

**[00:18:11 – 00:18:15]** years would have improved by a factor of 10, the second five

**[00:18:15 – 00:18:20]** years by another factor of 10, a 100x improvement in 10 years' time.

**[00:18:21 – 00:18:24]** In the same period of time, because of the way NVIDIA extreme

**[00:18:24 – 00:18:29]** co-designs, we're innovating across every single computer, across

**[00:18:29 – 00:18:33]** every single stack of the software, optimizing new algorithms, creating

**[00:18:33 – 00:18:36]** new algorithms, while creating architectures to accelerate it.

**[00:18:36 – 00:18:40]** Working across that entire domain, what we call extreme

**[00:18:41 – 00:18:45]** co-design, we were able to accelerate AI by a million times.

**[00:18:45 – 00:18:48]** And we're gonna continue to do so.

**[00:18:48 – 00:18:51]** This is what NVIDIA builds today.

**[00:18:51 – 00:18:57]** And this is the foundation by which we can enable Europe to be part of

**[00:18:57 – 00:19:00]** the AI ecosystem, the AI economy.

**[00:19:00 – 00:19:05]** Now, each generation, In the beginning, Pascal was GPUs,

**[00:19:05 – 00:19:11]** Volta was GPUs, Ampere was GPUs, and Hopper was GPUs.

**[00:19:11 – 00:19:14]** We invented the new DGX form factor.

**[00:19:15 – 00:19:18]** Then it was Blackwell that introduced to the world the first

**[00:19:18 – 00:19:22]** rack-scale computing, and I'll show you that in just a moment.

**[00:19:22 – 00:19:29]** So computing turned one entire system into one giant

**[00:19:29 – 00:19:31]** computer, one GPU.

**[00:19:31 – 00:19:35]** And now with Vera Rubin, we're extending it to processing

**[00:19:35 – 00:19:39]** beyond models, beyond large language models, to now processing

**[00:19:39 – 00:19:41]** the entire agent.

**[00:19:41 – 00:19:48]** Now, we'll refer back to this slide later, but remember this.

**[00:19:48 – 00:19:52]** The agent has to think.

**[00:19:52 – 00:19:55]** The agent has to think.

**[00:19:55 – 00:19:56]** And that's where the model comes in.

**[00:19:57 – 00:20:02]** Perception, reasoning, planning, and using tools.

**[00:20:02 – 00:20:05]** All of that runs on the large language model, and the models

**[00:20:05 – 00:20:07]** are designed to do that.

**[00:20:07 – 00:20:12]** But the agent also has what is called a harness, basically this

**[00:20:12 – 00:20:17]** exoskeleton of The digital agent.

**[00:20:17 – 00:20:19]** Just last, we have the brain.

**[00:20:19 – 00:20:22]** The rest of our body represents our exoskeleton.

**[00:20:22 – 00:20:26]** The agent is the exoskeleton that orchestrates the work, that

**[00:20:26 – 00:20:32]** allows the agent to actually engage memory, active memory, working

**[00:20:32 – 00:20:36]** memory, as well as long-term memory, and activating tools.

**[00:20:36 – 00:20:41]** Then the third part is a computer that processes knowledge,

**[00:20:41 – 00:20:43]** designed for data processing.

**[00:20:43 – 00:20:48]** And then lastly, a computer dedicated to CUDAx libraries,

**[00:20:48 – 00:20:50]** and its job is to run tools.

**[00:20:50 – 00:20:56]** And these computers now represent the entire agentic system.

**[00:20:56 – 00:20:58]** So Vera Rubin is in full production.

**[00:20:58 – 00:21:00]** It's nice to kind of see it.

**[00:21:00 – 00:21:04]** This is the first computer in the world that we've created.

**[00:21:04 – 00:21:06]** No cables.

**[00:21:06 – 00:21:10]** No wires, completely 100% liquid cooled, and it's just

**[00:21:10 – 00:21:11]** really delightful.

**[00:21:11 – 00:21:15]** Whoops, hang on.

**[00:21:15 – 00:21:18]** Really delightful to be able to pull it out.

**[00:21:18 – 00:21:22]** And this is one Vera Rubin computer.

**[00:21:22 – 00:21:27]** Okay, so this has two GPUs, yep.

**[00:21:32 – 00:21:36]** What used to take, in the case of Grace Blackwell, two

**[00:21:36 – 00:21:39]** hours to assemble one node, and when you see it, you'll

**[00:21:39 – 00:21:44]** know what I mean, cables and wires everywhere, 80% liquid-cooled.

**[00:21:44 – 00:21:47]** Now, Vera Rubin is 100% liquid-cooled, and it takes less

**[00:21:47 – 00:21:49]** than five minutes to assemble it.

**[00:21:49 – 00:21:52]** And so, incredible, incredible achievement.

**[00:21:52 – 00:21:55]** Of course, this translates to higher cycle times in

**[00:21:55 – 00:21:59]** manufacturing, much, much better reliability, and of

**[00:21:59 – 00:22:03]** course, a lot better performance.

**[00:22:03 – 00:22:08]** This is the NVLink switch tray, and what this does is

**[00:22:08 – 00:22:12]** it connects 72 of these GPUs.

**[00:22:12 – 00:22:14]** There are four of them in here right now.

**[00:22:14 – 00:22:19]** It connects 72 GPUs into one gigantic GPU.

**[00:22:19 – 00:22:24]** Okay, so each GPU is 20 petaflops, connecting 72 of them into

**[00:22:24 – 00:22:27]** one gigantic system.

**[00:22:29 – 00:22:33]** And this is our new Vera computer.

**[00:22:34 – 00:22:37]** So this auditorium is packed.

**[00:22:37 – 00:22:42]** We have 100 other watch parties around the world.

**[00:22:42 – 00:22:46]** 100 watch parties, 15,000 people are watching us right now.

**[00:22:46 – 00:22:47]** And so I'll get through the keynote.

**[00:22:47 – 00:22:50]** There's some really, really fun things I wanna show you

**[00:22:50 – 00:22:51]** in the end, okay?

**[00:22:51 – 00:22:54]** So you just have to, you have to, I have to torture you

**[00:22:54 – 00:22:57]** through some education first.

**[00:22:57 – 00:22:59]** And then the entertainment is coming.

**[00:22:59 – 00:23:03]** This, this is our brand new, this is our brand new Vera computer.

**[00:23:03 – 00:23:08]** Vera CPU is an extraordinary CPU beast.

**[00:23:08 – 00:23:11]** Eight instructions per clock.

**[00:23:11 – 00:23:15]** It has the world's first LPDDR memory subs, memory

**[00:23:15 – 00:23:18]** system, so that we can have high performance, very high

**[00:23:18 – 00:23:23]** bandwidth, as well as extremely excellent energy efficiency.

**[00:23:24 – 00:23:28]** The on-chip fabric is the fastest of any chip in the world.

**[00:23:28 – 00:23:31]** As a result, data processing is off the charts.

**[00:23:31 – 00:23:37]** Agentic sessions starting and processing is off the charts.

**[00:23:37 – 00:23:39]** And this is designed for doing exactly that.

**[00:23:39 – 00:23:41]** This is our Vera computer.

**[00:23:41 – 00:23:47]** And this computer here, this computer here is the Spectrum X.

**[00:23:47 – 00:23:52]** This is the world's first co-packaged optics.

**[00:23:52 – 00:23:53]** And you could see it here.

**[00:23:53 – 00:23:56]** I'll show you.

**[00:23:57 – 00:24:02]** Some of this area falls into the ground, and some of this

**[00:24:02 – 00:24:07]** area you can walk on, I'm just not sure which.

**[00:24:09 – 00:24:17]** All right, so if I accidentally disappear, you guys know us,

**[00:24:18 – 00:24:20]** it's not a stunt, it's an accident.

**[00:24:20 – 00:24:27]** And so you could see here, these are the lasers, and these

**[00:24:27 – 00:24:33]** lasers come up through optics, and it directly connects to our chip.

**[00:24:33 – 00:24:36]** And it connects using a process technology we invented with

**[00:24:36 – 00:24:38]** TSMC called COUPE.

**[00:24:38 – 00:24:42]** And so silicon photonics are directly attached

**[00:24:42 – 00:24:44]** right onto our chip.

**[00:24:44 – 00:24:48]** Otherwise, what you would see is all of this would be PC

**[00:24:48 – 00:24:54]** board, and PC boards Transmission capability is much worse than

**[00:24:54 – 00:24:58]** optical cables, and so we now have extremely high data rate,

**[00:24:58 – 00:25:03]** extremely high quality data that comes directly onto the chip, and

**[00:25:03 – 00:25:09]** this goes directly into otherwise transceivers without going

**[00:25:09 – 00:25:13]** through a PCB, and a transceiver goes right into the optics.

**[00:25:13 – 00:25:16]** Okay, so this is the world's first CPO.

**[00:25:16 – 00:25:22]** We're in complete volume production, and a 512 port

**[00:25:22 – 00:25:24]** at 200 gigabits per second.

**[00:25:24 – 00:25:27]** The world's most advanced ethernet switch.

**[00:25:27 – 00:25:30]** And we're really proud of it.

**[00:25:35 – 00:25:39]** And so what that gives us, what that gives us, is the

**[00:25:39 – 00:25:44]** ability to process the thoughts, the large language model,

**[00:25:44 – 00:25:50]** process the agent harness, run data processing for its

**[00:25:50 – 00:25:55]** knowledge, and then run accelerated computing for all of the tools.

**[00:25:55 – 00:25:58]** That represents essentially the agent workload, and I'll

**[00:25:58 – 00:26:00]** show you more in a second.

**[00:26:00 – 00:26:03]** While Vera Rubin is in full production, and the performance

**[00:26:03 – 00:26:05]** is off the charts.

**[00:26:05 – 00:26:08]** Every single generation, as I mentioned, through extreme

**[00:26:08 – 00:26:11]** co-design, we're able to achieve performances that are well

**[00:26:11 – 00:26:15]** beyond reason, well beyond what Moore's Law would have dictated.

**[00:26:16 – 00:26:18]** And as you know, Moore's Law has really largely ended.

**[00:26:18 – 00:26:23]** And so we need these extreme co-design ideas, working from

**[00:26:23 – 00:26:27]** chips to systems, systems to system software, to algorithm,

**[00:26:27 – 00:26:29]** to the entire AI factory.

**[00:26:29 – 00:26:35]** Well, Vera Rubin is another giant step on top of Blackwell.

**[00:26:35 – 00:26:41]** Blackwell was about 30 times faster than Hopper in doing

**[00:26:41 – 00:26:43]** large language model processing.

**[00:26:43 – 00:26:50]** And now, Vera Rubin is another 30 times greater than Blackwell

**[00:26:50 – 00:26:52]** in just two years' time.

**[00:26:52 – 00:26:56]** Let's take a look at how Vera Rubin turns into what is a

**[00:26:56 – 00:27:01]** DSX full-stack AI factory.

**[00:27:04 – 00:27:07]** AI is creating the largest infrastructure

**[00:27:07 – 00:27:08]** build-out in history.

**[00:27:08 – 00:27:11]** The factories are massive.

**[00:27:11 – 00:27:13]** The investments are significant.

**[00:27:13 – 00:27:15]** We have to build it right.

**[00:27:15 – 00:27:19]** But great chips alone don't make a great AI factory.

**[00:27:19 – 00:27:23]** NVIDIA DSX is a full-stack AI factory reference architecture,

**[00:27:23 – 00:27:27]** co-innovated with the world's leading infrastructure companies;

**[00:27:27 – 00:27:31]** compute, networking, storage, software, power and cooling,

**[00:27:32 – 00:27:34]** co-designed as one system.

**[00:27:34 – 00:27:40]** Optimized from design to operation, DSX starts in a digital twin.

**[00:27:40 – 00:27:44]** Performance, power, water, and scale, validated virtually

**[00:27:44 – 00:27:47]** before construction begins.

**[00:27:47 – 00:27:51]** PTC turns equipment designs into simulation-ready assets.

**[00:27:52 – 00:27:55]** Dassault Systems connects rack and facility design through

**[00:27:55 – 00:27:57]** model-based systems engineering.

**[00:27:58 – 00:28:02]** Schneider Electric simulates power distribution and liquid cooling.

**[00:28:02 – 00:28:06]** Jacobs brings it all together in the Omniverse Digital Twin.

**[00:28:07 – 00:28:11]** Siemens evaluates thermal, acoustic, and grid impacts.

**[00:28:12 – 00:28:15]** And Procore supports construction phase planning.

**[00:28:16 – 00:28:20]** Power limits how much an AI factory can produce.

**[00:28:20 – 00:28:23]** DSX Max LPS is an end-to-end control

**[00:28:23 – 00:28:27]** loop that dynamically steers power to match workload demand.

**[00:28:27 – 00:28:31]** It uses 30% less power for the same output.

**[00:28:31 – 00:28:34]** DSX rethinks cooling.

**[00:28:34 – 00:28:37]** NVIDIA Vera Rubin systems are designed for closed-loop liquid

**[00:28:37 – 00:28:42]** cooling, with coolant temperatures up to 45 degrees Celsius.

**[00:28:43 – 00:28:46]** In suitable climates, outdoor dry coolers release heat into

**[00:28:46 – 00:28:49]** the air without evaporating water.

**[00:28:49 – 00:28:55]** At 2 gigawatts, DSX can save 54 million liters of water a day.

**[00:28:55 – 00:28:59]** That's the daily water use of 180,000 households.

**[00:29:00 – 00:29:04]** But an AI factory can take several months to build.

**[00:29:04 – 00:29:08]** Until it is operational, it produces no tokens.

**[00:29:08 – 00:29:13]** Today, we are announcing NVIDIA DSX Pod, bringing this full-stack

**[00:29:13 – 00:29:17]** platform into a factory-built, integrated system.

**[00:29:17 – 00:29:19]** It starts with Vera Rubin NVL72.

**[00:29:19 – 00:29:25]** Seven liquid-cooled compute racks connected with 17 kilometers

**[00:29:25 – 00:29:30]** of Spectrum X cables form a Vera Rubin DSX Pod.

**[00:29:31 – 00:29:35]** Vera Rubin DSX pods are integrated, fully cabled, and factory

**[00:29:35 – 00:29:39]** tested, then delivered in transportable containers.

**[00:29:39 – 00:29:44]** Dry coolers replace traditional evaporative cooling.

**[00:29:44 – 00:29:48]** 800 volts DC connects to the racks for efficient power delivery.

**[00:29:48 – 00:29:52]** Once site utilities are ready, estimated on-site deployment

**[00:29:52 – 00:29:54]** goes from several months to days.

**[00:29:55 – 00:30:00]** NVIDIA DSX AI factory is responsible by design.

**[00:30:00 – 00:30:06]** And now, with DSX Pod, AI factories can be built at the speed of light.

**[00:30:06 – 00:30:08]** More intelligence per watt.

**[00:30:08 – 00:30:11]** Less time to first token.

**[00:30:16 – 00:30:18]** What do you think?

**[00:30:27 – 00:30:32]** So in 2016, we introduced the world's first AI supercomputer.

**[00:30:32 – 00:30:36]** It was in one node, one GPU node.

**[00:30:36 – 00:30:42]** Then three years ago, we introduced the Grace Blackwell NVLink

**[00:30:42 – 00:30:46]** 72 rack-scale computer.

**[00:30:46 – 00:30:50]** Both times, we had to reinvent not just the computer, the

**[00:30:50 – 00:30:55]** software, and how it works, but also the supply chain.

**[00:30:55 – 00:31:01]** In the case of NVLink72, we took the entire rack, which

**[00:31:01 – 00:31:05]** would have otherwise been assembled on site, and we assembled

**[00:31:05 – 00:31:08]** it inside the supply chain.

**[00:31:08 – 00:31:13]** We essentially manufactured supercomputers at scale.

**[00:31:13 – 00:31:17]** It is the reason why we're able to move so fast and be able

**[00:31:17 – 00:31:21]** to provide for literally several hundred data center build-outs

**[00:31:22 – 00:31:24]** around the world at the same time.

**[00:31:24 – 00:31:26]** Every single computer manufacturer in the world

**[00:31:26 – 00:31:30]** now manufactures NVLink 72.

**[00:31:30 – 00:31:34]** The scale at which we are now producing computing is

**[00:31:34 – 00:31:36]** unheard of ever, unimaginable.

**[00:31:37 – 00:31:42]** Well, even then, when you bring these systems onto your

**[00:31:42 – 00:31:47]** sites; there are still months and months of assembly, and

**[00:31:47 – 00:31:55]** you just heard 17,000 kilometers of wiring to connect just Spectrum

**[00:31:55 – 00:31:59]** X for the scale-out computing.

**[00:32:00 – 00:32:02]** 17 kilometers, 17,000 kilometers.

**[00:32:02 – 00:32:04]** It's crazy number.

**[00:32:04 – 00:32:07]** And so 17 kilometers, incredible numbers.

**[00:32:07 – 00:32:12]** Now suppose we were to do that all inside the supply chain, run

**[00:32:12 – 00:32:16]** everything up, build everything up, make sure that it runs completely,

**[00:32:16 – 00:32:18]** get it qualified, burnt in.

**[00:32:18 – 00:32:22]** And by the time we ship it, it is fully functional.

**[00:32:22 – 00:32:26]** The amount of cabling and on-site assembly is going

**[00:32:26 – 00:32:28]** to be greatly reduced.

**[00:32:28 – 00:32:31]** But there's something much, much more.

**[00:32:31 – 00:32:35]** In fact, what's gonna happen here is that by moving this

**[00:32:35 – 00:32:40]** into the supply chain, the amount of idle computing, the amount

**[00:32:40 – 00:32:43]** of idle computing that's gonna be sitting inside at your data

**[00:32:43 – 00:32:46]** centers will reduce dramatically.

**[00:32:46 – 00:32:48]** Remember this.

**[00:32:48 – 00:32:54]** Each gigawatt data center is $50 billion.

**[00:32:55 – 00:33:00]** The revenue generation you expect is tens of billion

**[00:33:00 – 00:33:02]** dollars per year.

**[00:33:02 – 00:33:08]** And so if we just said it's$ 40 billion per year of revenue

**[00:33:08 – 00:33:12]** generation that comes out, let's get out of that infrastructure.$

**[00:33:12 – 00:33:18]** 40 billion, if delayed by just three months, That's $5

**[00:33:18 – 00:33:20]** billion, excuse me, what is that?$

**[00:33:20 – 00:33:23]** 10 billion worth of revenues.$

**[00:33:23 – 00:33:28]** 10 billion worth of revenues will make anybody anxious.

**[00:33:28 – 00:33:34]** And so by moving this into the supply chain, the speed by which we

**[00:33:34 – 00:33:39]** can turn on, go from installation, arrival of computing, to production

**[00:33:39 – 00:33:44]** of tokens, will go from many months, we think something

**[00:33:44 – 00:33:47]** along the lines of four to five months, down to about a month.

**[00:33:47 – 00:33:50]** Okay, so this is a complete transformation of the world

**[00:33:50 – 00:33:54]** supply chain, and this will change the way that people install and

**[00:33:54 – 00:33:58]** build data centers, and hopefully, as a result, velocity goes

**[00:33:58 – 00:34:05]** up, the wasted idle compute, the wasted idle capex will reduce, and

**[00:34:06 – 00:34:08]** everybody becomes more profitable.

**[00:34:08 – 00:34:15]** So this is NVLink 72, now into a large cabinet system.

**[00:34:15 – 00:34:17]** We call it DSX pod.

**[00:34:18 – 00:34:23]** Okay, so this is the next click of NVIDIA computing.

**[00:34:30 – 00:34:37]** Now, what you make and how you make it makes you special.

**[00:34:37 – 00:34:43]** Just as here in Europe, the work you know makes you special.

**[00:34:43 – 00:34:45]** You design machines.

**[00:34:45 – 00:34:47]** You operate factories.

**[00:34:47 – 00:34:50]** You discover medicine.

**[00:34:50 – 00:34:53]** You have extraordinary craft.

**[00:34:53 – 00:34:57]** That craft, that specialty, makes you special.

**[00:34:57 – 00:35:01]** Well, what makes NVIDIA special are really two things.

**[00:35:02 – 00:35:06]** The first thing is that our architecture is the only

**[00:35:06 – 00:35:10]** accelerated computing architecture that is general purpose.

**[00:35:10 – 00:35:13]** We don't specialize on just one algorithm, for

**[00:35:13 – 00:35:15]** example, transformers.

**[00:35:15 – 00:35:19]** We specialize across anything that could be paralyzed.

**[00:35:19 – 00:35:21]** Could be fluids and particles.

**[00:35:21 – 00:35:23]** It could be quantum.

**[00:35:23 – 00:35:25]** It could be quantum chemistry.

**[00:35:25 – 00:35:27]** It could be image processing.

**[00:35:27 – 00:35:30]** It could be data processing, structured vector data processing,

**[00:35:31 – 00:35:32]** processing of all kinds.

**[00:35:32 – 00:35:35]** And because we've been working at it for so long, and because

**[00:35:35 – 00:35:38]** NVIDIA's architecture has remained consistent for literally

**[00:35:38 – 00:35:43]** two and a half decades, the number of algorithms, the

**[00:35:43 – 00:35:46]** number of domains, the number of industries we serve, we

**[00:35:46 – 00:35:49]** accelerate, has grown tremendously.

**[00:35:49 – 00:35:53]** Our installed base, our general purpose architecture, our

**[00:35:53 – 00:35:57]** dedication to this architecture, and our collaboration with

**[00:35:57 – 00:36:00]** all of the world's industries has made NVIDIA's architecture

**[00:36:00 – 00:36:03]** literally everywhere.

**[00:36:03 – 00:36:06]** The second thing is how we do it.

**[00:36:06 – 00:36:10]** Of course, we have the architecture that we developed, but we

**[00:36:10 – 00:36:17]** also build it now in a full stack, complete AI factory.

**[00:36:17 – 00:36:20]** There are many companies in the world who want to design their own

**[00:36:20 – 00:36:25]** chips, but most of the companies in the world just wants to deploy the

**[00:36:25 – 00:36:29]** AI infrastructure so that they can run their business on top of it.

**[00:36:29 – 00:36:32]** The ability to design a chip and build a supercomputer,

**[00:36:32 – 00:36:37]** write all the software, is really completely out of mind for most

**[00:36:37 – 00:36:38]** of the companies around the world.

**[00:36:38 – 00:36:43]** And so we now have the combination of these two ideas, CUDA,

**[00:36:43 – 00:36:47]** full-stack AI factories, and it exposes us to markets that

**[00:36:47 – 00:36:49]** otherwise is impossible.

**[00:36:50 – 00:36:52]** We're able to partner with so many companies here for

**[00:36:52 – 00:36:53]** this very reason.

**[00:36:53 – 00:36:58]** If I were to go around this, starting with CSPs, this is

**[00:36:58 – 00:37:02]** where most people think the vast majority of AI is today.

**[00:37:02 – 00:37:05]** But as you can see in the beginning of our keynote,

**[00:37:05 – 00:37:08]** I mentioned a whole bunch of companies who are not those.

**[00:37:08 – 00:37:11]** So CSPs remain very large businesses for

**[00:37:11 – 00:37:13]** the world's computing.

**[00:37:13 – 00:37:16]** But there are AI natives, companies that are starting

**[00:37:16 – 00:37:17]** all over the world.

**[00:37:17 – 00:37:19]** There are so many that are here.

**[00:37:19 – 00:37:24]** Neoclouds, because every region wants to participate

**[00:37:24 – 00:37:25]** in the AI economy.

**[00:37:25 – 00:37:29]** Every region wants to ensure that the valuable resources

**[00:37:29 – 00:37:32]** that they have, the energy that they have, is dedicated

**[00:37:32 – 00:37:34]** to companies in their region.

**[00:37:34 – 00:37:38]** For example, Mistral here in Europe, Mistral in France,

**[00:37:38 – 00:37:40]** Nebia is here in Europe.

**[00:37:40 – 00:37:45]** So many companies around Europe that are building up neoclouds.

**[00:37:45 – 00:37:48]** OEMs and ODMs who are building computers for

**[00:37:48 – 00:37:50]** the world's enterprise.

**[00:37:50 – 00:37:51]** AI labs.

**[00:37:52 – 00:37:57]** Of course, we know about last year, all we spoke about was OpenAI.

**[00:37:57 – 00:38:02]** Today, we speak about OpenAI, Anthropic, but we also talk

**[00:38:02 – 00:38:09]** about the incredible Meta's Muse, Groq, Bot, and of course, Gemini.

**[00:38:09 – 00:38:13]** So many different AI labs are now participating at the frontier.

**[00:38:13 – 00:38:16]** And then very importantly, and this is ultimately the most

**[00:38:16 – 00:38:22]** important, for AI to succeed, for Europe to benefit from AI and to

**[00:38:22 – 00:38:27]** have really created an AI ecosystem and AI economy, you need all of the

**[00:38:27 – 00:38:32]** end markets, the end industries, to adopt AI and to succeed with AI.

**[00:38:32 – 00:38:34]** Whether it's in the automotive industry, because we're building

**[00:38:34 – 00:38:39]** self-driving cars, or the way you manufacture with robotics,

**[00:38:39 – 00:38:43]** the way you discover new drugs in healthcare, your financial

**[00:38:43 – 00:38:46]** services industry.

**[00:38:46 – 00:38:50]** All of these different industries have to become AI companies.

**[00:38:50 – 00:38:52]** Every single one of them.

**[00:38:52 – 00:38:56]** Well, we have a unique ability through what I just mentioned

**[00:38:56 – 00:38:59]** about CUDA and the way that we build computers.

**[00:38:59 – 00:39:01]** We have a unique ability to serve literally every

**[00:39:01 – 00:39:05]** single segment of AI.

**[00:39:08 – 00:39:12]** And so one architecture, many markets, and because these

**[00:39:12 – 00:39:16]** infrastructure are so significant and so expensive, we need to make

**[00:39:16 – 00:39:21]** sure that they're as productive as possible, as soon as possible,

**[00:39:21 – 00:39:26]** that they're fungible, that you can use it for every workload.

**[00:39:26 – 00:39:29]** From the beginning of your AI journey, data processing

**[00:39:29 – 00:39:34]** to training to post-training, inference and agentic processing,

**[00:39:34 – 00:39:38]** every single step of the AI journey, to every type of

**[00:39:38 – 00:39:42]** AI, chemicals and physics, to languages, to every single industry

**[00:39:42 – 00:39:44]** and the algorithms that they use.

**[00:39:44 – 00:39:47]** And so, NVIDIA's architecture is fungible and as a result

**[00:39:47 – 00:39:54]** of that, continuous innovation, continuous updating of software.

**[00:39:54 – 00:39:57]** Because the installed base is so great, we have such an incredibly

**[00:39:57 – 00:40:01]** rich ecosystem of software, then this architecture is also durable.

**[00:40:02 – 00:40:05]** Fungible, durable, productive.

**[00:40:05 – 00:40:08]** Well, that lays the foundation.

**[00:40:08 – 00:40:10]** That is the infrastructure.

**[00:40:10 – 00:40:12]** Now what do we do with it?

**[00:40:12 – 00:40:15]** This last year, we saw a transition in AI.

**[00:40:15 – 00:40:17]** AI is now productive.

**[00:40:17 – 00:40:21]** And it's one of the most important times was in the beginning

**[00:40:21 – 00:40:26]** of the year when we spoke about OpenClaw, which introduced the

**[00:40:26 – 00:40:28]** concept of an open source agent.

**[00:40:28 – 00:40:32]** It really started what we all see across the world today.

**[00:40:32 – 00:40:40]** Agents think, reason, iterate, use tools, and as a result, consume

**[00:40:40 – 00:40:42]** an enormous amount of tokens.

**[00:40:42 – 00:40:44]** But as a result, they become useful.

**[00:40:44 – 00:40:49]** So more agents are being developed, they're more useful, and therefore

**[00:40:49 – 00:40:51]** they're used by more people.

**[00:40:51 – 00:40:54]** And because they have to use a lot of tokens to think,

**[00:40:54 – 00:40:57]** the amount of token consumption has gone through the roof.

**[00:40:57 – 00:41:00]** GPU demand has really gone through the roof.

**[00:41:00 – 00:41:02]** Now, what's processing it is this.

**[00:41:02 – 00:41:04]** This is an agent.

**[00:41:04 – 00:41:06]** These are the five elements of an agent.

**[00:41:06 – 00:41:09]** The first part are the language models.

**[00:41:10 – 00:41:12]** These are the tokens that are running on top of the

**[00:41:12 – 00:41:16]** GPUs, and they're generating tokens that are consumed by the

**[00:41:16 – 00:41:19]** agent, what I call the exoskeleton.

**[00:41:20 – 00:41:26]** That open-claw harness sits inside a sandbox and that sandbox isolates

**[00:41:26 – 00:41:27]** it from the rest of the system.

**[00:41:27 – 00:41:30]** This is a very important concept that I'm going to come back to.

**[00:41:30 – 00:41:32]** There are five things.

**[00:41:32 – 00:41:36]** There's the model, there's the harness, the sandbox that

**[00:41:36 – 00:41:41]** the harness sits in, and two other things, knowledge and tools.

**[00:41:42 – 00:41:46]** Models, sandbox, harness, knowledge, and tools.

**[00:41:46 – 00:41:50]** These are the five elements of an agent.

**[00:41:50 – 00:41:54]** Now, before I talk about how to make the agents useful, the

**[00:41:54 – 00:41:58]** most important thing is we have to make sure that the agent is secure.

**[00:41:58 – 00:42:01]** The agent is autonomous, and it's an autonomous system

**[00:42:01 – 00:42:04]** that has access to files, has access to resources,

**[00:42:04 – 00:42:06]** networks, access to tools.

**[00:42:07 – 00:42:11]** And we need to make sure that the privileges that we provide

**[00:42:11 – 00:42:18]** this autonomous system is minimal, and that we constantly monitor it.

**[00:42:18 – 00:42:23]** We introduced a fantastic new product, a new platform,

**[00:42:23 – 00:42:25]** and it's called OpenShell.

**[00:42:25 – 00:42:27]** OpenShell came just in time.

**[00:42:27 – 00:42:31]** This is a secure runtime that provides for the things that

**[00:42:31 – 00:42:37]** I just mentioned and has an additional hardware-based monitor.

**[00:42:37 – 00:42:41]** Before I explain it to you, let's watch a video.

**[00:42:44 – 00:42:47]** Enterprises are preparing to deploy billions of AI agents.

**[00:42:48 – 00:42:53]** To be useful, they'll reason, plan, use tools, access data,

**[00:42:53 – 00:42:54]** and work together.

**[00:42:55 – 00:42:58]** Today, the only thing stopping an agent from misusing all

**[00:42:58 – 00:43:00]** that is asking it nicely.

**[00:43:00 – 00:43:02]** Please don't delete anything.

**[00:43:02 – 00:43:04]** Please don't push to make.

**[00:43:04 – 00:43:06]** Please don't touch my SSH keys.

**[00:43:06 – 00:43:08]** That's the please don't approach.

**[00:43:08 – 00:43:11]** But hope isn't security.

**[00:43:11 – 00:43:13]** Safety isn't an aspiration.

**[00:43:13 – 00:43:15]** It's technology and engineering.

**[00:43:15 – 00:43:19]** NVIDIA OpenShell and Sentry put that philosophy into practice.

**[00:43:19 – 00:43:22]** You write a policy for what each agent can touch, and

**[00:43:22 – 00:43:27]** OpenShell enforces it at the granularity of system calls, file

**[00:43:27 – 00:43:30]** access, and network connections.

**[00:43:30 – 00:43:33]** Formal verification checks every policy change against your

**[00:43:33 – 00:43:38]** permission ceiling, so agents never gain more access than intended.

**[00:43:38 – 00:43:40]** It's like the early Internet.

**[00:43:40 – 00:43:44]** Viruses were everywhere until browsers enforced strict limits

**[00:43:44 – 00:43:46]** on what a web page could do.

**[00:43:47 – 00:43:49]** OpenShell does the same for agents.

**[00:43:49 – 00:43:53]** Enforcement happens outside the agent, not just in LLM

**[00:43:53 – 00:43:57]** guardrails, bringing agent native security into your infrastructure.

**[00:43:58 – 00:44:01]** Every agent runs inside its own isolated sandbox, with

**[00:44:01 – 00:44:04]** no direct route to the network.

**[00:44:04 – 00:44:07]** Every outbound network request goes through the supervisor,

**[00:44:07 – 00:44:09]** outside the sandbox.

**[00:44:09 – 00:44:11]** It checks the action against your policy.

**[00:44:11 – 00:44:14]** If it's not permitted, it's denied.

**[00:44:14 – 00:44:18]** Keys stay outside, with the supervisor, and are added only

**[00:44:18 – 00:44:20]** on the way to an approved endpoint.

**[00:44:20 – 00:44:23]** A compromised agent can't leak a key it never had.

**[00:44:23 – 00:44:26]** The gateway is the control plane.

**[00:44:26 – 00:44:30]** The supervisor makes the decisions, the sandbox holds the agent,

**[00:44:30 – 00:44:33]** and the policy prover checks proposed policies against

**[00:44:33 – 00:44:35]** your permission ceilings.

**[00:44:35 – 00:44:39]** OpenShell combines Linux kernel isolation with runtime

**[00:44:39 – 00:44:41]** policy enforcement.

**[00:44:41 – 00:44:45]** NVIDIA Sentry adds an optional layer of protection outside the

**[00:44:45 – 00:44:48]** host itself, away from the agent.

**[00:44:48 – 00:44:52]** Running on NVIDIA BlueField 4, Sentry monitors agent activity

**[00:44:52 – 00:44:57]** and enforces security policies in a hardware-isolated trust domain.

**[00:44:57 – 00:45:00]** Because agent reasoning traffic passes through Sentry, if

**[00:45:00 – 00:45:03]** the agent is planning to do something unintended, it will

**[00:45:03 – 00:45:07]** update policies immediately and cut the connection to the model,

**[00:45:07 – 00:45:09]** quarantining it in milliseconds.

**[00:45:09 – 00:45:12]** And it keeps tamper-evident records that operators

**[00:45:12 – 00:45:16]** can verify for detection, investigation, and response.

**[00:45:17 – 00:45:21]** OpenShell is open source under Apache 2.0, runs everywhere

**[00:45:21 – 00:45:26]** your agents do, edge to cloud, and takes two commands to adopt.

**[00:45:26 – 00:45:31]** Install the CLI, create a sandbox, and run your favorite agent.

**[00:45:31 – 00:45:34]** Agents are only as useful as the access you give them.

**[00:45:34 – 00:45:37]** OpenShell and Sentry put that access under your control.

**[00:45:37 – 00:45:43]** That's how we advance AI capabilities and safety together.

**[00:45:51 – 00:45:53]** Agents are autonomous systems.

**[00:45:53 – 00:45:59]** We have to create extremely secure technology to make sure

**[00:45:59 – 00:46:03]** that we could develop these agents and deploy these agents safely.

**[00:46:03 – 00:46:04]** Now, it's incredible.

**[00:46:05 – 00:46:08]** After announcing OpenShell, just about every major company

**[00:46:08 – 00:46:12]** in the world, from the Frontier Labs to the operating system

**[00:46:12 – 00:46:16]** companies, cybersecurity companies, platform companies, have come

**[00:46:16 – 00:46:18]** out to adopt OpenShell.

**[00:46:18 – 00:46:20]** It's really fantastic to see.

**[00:46:20 – 00:46:21]** And the reason for that is because this technology has

**[00:46:22 – 00:46:23]** arrived at exactly the right time.

**[00:46:23 – 00:46:26]** At a time when we want to deploy agents all over the world.

**[00:46:26 – 00:46:32]** Now remember, this isn't just about the services using OpenShell

**[00:46:32 – 00:46:35]** as they deploy agents, this is really about all of the

**[00:46:35 – 00:46:40]** world's companies also adopting OpenShell so that you could receive

**[00:46:40 – 00:46:43]** these agents and have them run autonomously inside your company.

**[00:46:43 – 00:46:49]** Every single level of capability, every single task that you

**[00:46:49 – 00:46:53]** give the agent, for every person that uses the agent, every

**[00:46:53 – 00:46:57]** one of those combinations requires a different level of privilege that

**[00:46:57 – 00:46:59]** you would provide to the agent.

**[00:46:59 – 00:47:02]** And every single time an agent is launched,

**[00:47:02 – 00:47:04]** the privileges are different.

**[00:47:04 – 00:47:09]** And so privileges, the policy enforcement, the monitoring,

**[00:47:09 – 00:47:11]** all has to be done dynamically.

**[00:47:11 – 00:47:14]** And so OpenShell is really revolutionary.

**[00:47:14 – 00:47:17]** I'm delighted to say that Anthropic and NVIDIA are partnering

**[00:47:17 – 00:47:21]** to create cloud OpenShell agents.

**[00:47:21 – 00:47:23]** And so you can go to the website.

**[00:47:23 – 00:47:25]** There's a really great write-up about that.

**[00:47:25 – 00:47:31]** So the first part, the first part is how do you use these harnesses?

**[00:47:31 – 00:47:34]** OpenClaw, Hermes, OpenHand.

**[00:47:34 – 00:47:39]** You could create your own harnesses from Langchain, DeepAgents,

**[00:47:39 – 00:47:41]** or just write it from Python.

**[00:47:41 – 00:47:42]** There's a whole bunch of different ways you could

**[00:47:43 – 00:47:45]** create agent harnesses.

**[00:47:45 – 00:47:50]** That sits inside now an OpenShell secure harness.

**[00:47:50 – 00:47:55]** And that's the sandbox that it sits in, the secure sandbox.

**[00:47:55 – 00:48:00]** And then next, let's talk about models.

**[00:48:00 – 00:48:02]** You should use models of all kinds.

**[00:48:02 – 00:48:06]** In the future, it won't just be one model that we run everything on.

**[00:48:06 – 00:48:09]** We're going to use the best models, whether they're

**[00:48:10 – 00:48:13]** closed, open, or custom.

**[00:48:13 – 00:48:16]** We're gonna use all the models to do your work.

**[00:48:16 – 00:48:20]** Sometimes the work that you do is fairly generic.

**[00:48:20 – 00:48:24]** Artificial general intelligence is gonna do an incredibly good job.

**[00:48:24 – 00:48:28]** However, sometimes the work that you wanna do captures too

**[00:48:28 – 00:48:32]** much of your proprietary knowledge, too much of your specialty,

**[00:48:32 – 00:48:36]** and you want to continuously enhance that intelligence.

**[00:48:36 – 00:48:41]** Every company in the future will use both closed models as well

**[00:48:41 – 00:48:43]** as open models that you customize.

**[00:48:43 – 00:48:46]** By the end of this session, you will see that, in fact, building

**[00:48:46 – 00:48:49]** your own AIs is not that hard.

**[00:48:49 – 00:48:52]** It starts with, of course, the model.

**[00:48:52 – 00:48:56]** We've been dedicated to advancing open models to open up new

**[00:48:56 – 00:49:01]** opportunities and enable people to build their own custom AIs.

**[00:49:01 – 00:49:07]** Nemotron is for languages, Cosmos is our world foundation

**[00:49:07 – 00:49:09]** model, we are the frontier.

**[00:49:10 – 00:49:14]** GROOT for Human and Robotics, we are the frontier.

**[00:49:14 – 00:49:18]** Alpamayo for Autonomous Navigation, we are the frontier.

**[00:49:18 – 00:49:22]** iSYNC for Quantum Classical Computing, we are the frontier.

**[00:49:22 – 00:49:27]** And of course, Earth2Physics AI, we are the frontier.

**[00:49:27 – 00:49:32]** So in all of these different areas, we offer you all of the models and

**[00:49:32 – 00:49:33]** we're continuously working on it.

**[00:49:33 – 00:49:35]** We've been working on it for a decade and we'll continue

**[00:49:35 – 00:49:37]** to work on this forever.

**[00:49:37 – 00:49:39]** We offer you the model completely open.

**[00:49:40 – 00:49:43]** The data that goes along with it so that you know the providence

**[00:49:43 – 00:49:47]** of the model that we created, as well as the scripts necessary

**[00:49:47 – 00:49:49]** for you to reproduce it.

**[00:49:49 – 00:49:53]** So not only do you have the open weights, you have from

**[00:49:53 – 00:49:58]** us the entire recipe and the data to recreate the open models so that

**[00:49:58 – 00:50:02]** you could know where it came from and have complete trust into it.

**[00:50:02 – 00:50:08]** Well, these models are not necessarily the frontier of

**[00:50:08 – 00:50:11]** frontiers, especially for the language model.

**[00:50:11 – 00:50:16]** However, with customization and with partners that the ecosystem

**[00:50:16 – 00:50:20]** that we work with, you could turn these models into frontier, and

**[00:50:20 – 00:50:24]** as a result, turn your proprietary data into sovereign intelligence.

**[00:50:24 – 00:50:31]** Here's an example of A21 helping Alice, turned Nemotron into a model

**[00:50:31 – 00:50:33]** that exceeds frontier capabilities.

**[00:50:33 – 00:50:38]** Here's Base10 helping CodeRabbit, turning a Nemotron model,

**[00:50:38 – 00:50:43]** after fine-tuning on your data, turned into frontier capability.

**[00:50:43 – 00:50:44]** Base10 and CodeRabbit.

**[00:50:44 – 00:50:50]** Here's Trajectory and Harvey, again, using Nemotron, post-trained

**[00:50:50 – 00:50:55]** it and turned and created their own frontier level capability.

**[00:50:55 – 00:50:59]** Here's Scale working with Skydance, again doing the same thing.

**[00:50:59 – 00:51:01]** It goes on and on and on.

**[00:51:01 – 00:51:04]** The point being that you could start with an open model,

**[00:51:04 – 00:51:08]** something that's incredibly good like Nemotron, and with

**[00:51:08 – 00:51:13]** your own data, post-train it for your use case and achieve,

**[00:51:13 – 00:51:15]** of course, super intelligence.

**[00:51:15 – 00:51:17]** That is a very sensible thing to do, and all the

**[00:51:17 – 00:51:19]** technology exists today.

**[00:51:19 – 00:51:23]** Here's an example of Palantir with Tampa General Hospital.

**[00:51:23 – 00:51:25]** The list goes on and on.

**[00:51:25 – 00:51:28]** We have a rich ecosystem of partners that can help you,

**[00:51:28 – 00:51:31]** and they're incredible companies, and they can work with you

**[00:51:31 – 00:51:33]** from around the world.

**[00:51:33 – 00:51:37]** Okay, so, we work, it's just, anyways, I'll mention one

**[00:51:37 – 00:51:40]** and I'll offend everybody else.

**[00:51:40 – 00:51:45]** I'll have to sit here and mention all of them, and I love them all.

**[00:51:45 – 00:51:50]** And so, first, Agent harness running in a

**[00:51:50 – 00:51:53]** secure sandbox open shell.

**[00:51:53 – 00:51:57]** Second, Nemotron large language models, the first two components

**[00:51:57 – 00:51:59]** of what you need.

**[00:51:59 – 00:52:03]** But ultimately, what turns these super agents, these agents into

**[00:52:03 – 00:52:08]** super agents, is that you give them access to your company's knowledge,

**[00:52:08 – 00:52:11]** and this, It's really special.

**[00:52:11 – 00:52:15]** This is called the Agent Knowledge Foundation that we put together,

**[00:52:15 – 00:52:18]** libraries and models that are completely world-class.

**[00:52:18 – 00:52:22]** It allows you to turn your company's data.

**[00:52:22 – 00:52:27]** The data could be in, the assets could be in unstructured

**[00:52:27 – 00:52:30]** format in documents, so that you need an AI to understand

**[00:52:30 – 00:52:33]** how to read these PDFs and read all these unstructured

**[00:52:33 – 00:52:37]** documents, or it could be in structured data inside databases.

**[00:52:37 – 00:52:41]** We have the libraries and the tools necessary to help you

**[00:52:41 – 00:52:47]** extract meaning and understanding from your company's data.

**[00:52:47 – 00:52:50]** One at a time, Nemotron parse, completely world-class,

**[00:52:50 – 00:52:52]** top of leaderboard.

**[00:52:52 – 00:52:56]** Nemotron embed, taking a query and associating with the knowledge

**[00:52:56 – 00:53:00]** that's inside the vector database, top of leaderboard.

**[00:53:01 – 00:53:04]** Nemotron re-rank, it might find many different sources

**[00:53:04 – 00:53:08]** where the query is associated with different pieces of data.

**[00:53:08 – 00:53:11]** It brings it together, ranks them based on which one's

**[00:53:11 – 00:53:15]** the most relevant to you.

**[00:53:15 – 00:53:20]** QMO, to be able to predict from your structured database

**[00:53:20 – 00:53:22]** what is likely to happen.

**[00:53:22 – 00:53:28]** And so this is, for example, if you were to receive, I notice

**[00:53:28 – 00:53:33]** that one of your suppliers will be short on supply or delay on supply.

**[00:53:33 – 00:53:39]** That piece of information is likely unstructured data, it's a document.

**[00:53:40 – 00:53:43]** The information about the orders in your company are likely

**[00:53:43 – 00:53:47]** inside a structured database, your SQL database, for example.

**[00:53:47 – 00:53:52]** And in order to predict what's going to happen to all of

**[00:53:52 – 00:53:56]** these orders, who are going to be short, the amount of delay, you

**[00:53:56 – 00:53:58]** use CUMA to make that prediction.

**[00:53:59 – 00:54:03]** And then in order to figure out what to replace it with, you might

**[00:54:03 – 00:54:07]** use vector database search and to go find an engineering document

**[00:54:07 – 00:54:09]** of a replacement component.

**[00:54:09 – 00:54:13]** Okay, so all of those different libraries and data processing

**[00:54:13 – 00:54:18]** capabilities now makes for you your knowledge foundation.

**[00:54:18 – 00:54:22]** You've now armed your AI with proprietary data,

**[00:54:22 – 00:54:24]** proprietary knowledge.

**[00:54:24 – 00:54:28]** Now, the last part that I'll talk about is ultimately tools.

**[00:54:28 – 00:54:30]** Tools of the trade.

**[00:54:30 – 00:54:33]** NVIDIA has been working on CUDA X, and for most of you,

**[00:54:33 – 00:54:37]** in the last, before AI, all of our conversations have

**[00:54:37 – 00:54:42]** been around how do we accelerate simulators and solvers, data

**[00:54:42 – 00:54:46]** processing engines, domain-specific libraries of all kinds.

**[00:54:46 – 00:54:49]** They're used in scientific discovery, in all the different

**[00:54:49 – 00:54:51]** fields of science.

**[00:54:51 – 00:54:52]** They're used in engineering.

**[00:54:52 – 00:54:53]** They're used in research.

**[00:54:53 – 00:54:58]** They're used in designing, building creative arts.

**[00:54:58 – 00:55:02]** At this core are CUDA X libraries, and these CUDA

**[00:55:02 – 00:55:03]** X libraries are algorithms.

**[00:55:03 – 00:55:06]** And algorithms, as you know, are beautiful.

**[00:55:06 – 00:55:09]** Let's take a look at it.

**[00:57:43 – 00:57:47]** Math is beautiful, math is beautiful.

**[00:57:47 – 00:57:50]** CUDA X libraries, these are the crown jewels of our company,

**[00:57:50 – 00:57:51]** and the reason for that is this.

**[00:57:51 – 00:57:54]** When you're building an AI assistant, it could be

**[00:57:54 – 00:57:57]** a lab assistant, it could be a research assistant,

**[00:57:57 – 00:58:00]** it could be a co-designer, a somebody who's co-creator with

**[00:58:00 – 00:58:03]** you, building something with you.

**[00:58:03 – 00:58:09]** Ultimately, what makes those agents useful, what makes them

**[00:58:09 – 00:58:12]** able to do real work are the tools.

**[00:58:13 – 00:58:20]** And so we now have models that allow these agents To

**[00:58:20 – 00:58:24]** reason, understand information, reason, to be able to plan

**[00:58:24 – 00:58:28]** and iteratively act and use tools until it could get the job done.

**[00:58:28 – 00:58:31]** But it still comes down to the tools.

**[00:58:31 – 00:58:34]** And these tools we've been working on for so many years.

**[00:58:34 – 00:58:36]** And there was one particular computer I showed you on the

**[00:58:36 – 00:58:40]** right-hand side that runs every single one of these tools.

**[00:58:40 – 00:58:42]** And it could be fluid dynamics and particle physics.

**[00:58:42 – 00:58:44]** It could be quantum chemistry.

**[00:58:44 – 00:58:45]** It could be data processing.

**[00:58:45 – 00:58:48]** It could be ray tracing.

**[00:58:48 – 00:58:51]** All of these tools run on that one computer.

**[00:58:51 – 00:58:55]** And that one computer speeds up the processing of it compared

**[00:58:55 – 00:58:59]** to just running it natively on a CPU by 10, 50, 100x.

**[00:58:59 – 00:59:05]** And CUDA X tools is what give the agents the tools of the trade.

**[00:59:06 – 00:59:08]** And it's so cool to see how other people are using it.

**[00:59:08 – 00:59:10]** Let's take a look.

**[00:59:42 – 00:59:45]** So this is the magic, magic behind the agents.

**[00:59:45 – 00:59:48]** Ultimately, the agents have to use tools.

**[00:59:48 – 00:59:50]** Now, of course, in one of the first fields of agentic

**[00:59:50 – 00:59:53]** use is software programming, and therefore, a compiler

**[00:59:53 – 00:59:55]** or a Python is all really at needs.

**[00:59:55 – 00:59:59]** But in the future, as we have these tools do more and more and more,

**[00:59:59 – 01:00:02]** and more useful work, we're going to need to give them more tools.

**[01:00:02 – 01:00:05]** And here's some examples. So here's LiDOS Sciences, they're creating

**[01:00:05 – 01:00:10]** a lab assistant, and they're using NVIDIA's BioNeMo models

**[01:00:10 – 01:00:14]** that are designed for understanding proteins and understanding

**[01:00:14 – 01:00:17]** cells and generating proteins and chemicals in drug discovery.

**[01:00:17 – 01:00:20]** They're using BioNeMo and CUDA X tools.

**[01:00:20 – 01:00:23]** This is Siemens Factory building a factory agent that uses

**[01:00:23 – 01:00:27]** NVIDIA Cosmos, our world foundation model that understands the

**[01:00:27 – 01:00:30]** physical world And CUDA X tools.

**[01:00:30 – 01:00:34]** And here's an example of Dassault creating a virtual companion

**[01:00:34 – 01:00:39]** that uses NVIDIA's Nemotron, our language model, and CUDA X tools.

**[01:00:39 – 01:00:42]** Let's see them in use.

**[01:00:44 – 01:00:48]** The most intelligent scientist still needs a laboratory.

**[01:00:48 – 01:00:52]** The best engineer still needs simulators and solvers.

**[01:00:52 – 01:00:56]** An AI agent is no different.

**[01:00:56 – 01:00:59]** An agent brings together five building blocks.

**[01:00:59 – 01:01:02]** The model perceives reasons and plans.

**[01:01:03 – 01:01:05]** The harness coordinates the work.

**[01:01:05 – 01:01:08]** Knowledge Foundation gives the agent context from your company's

**[01:01:08 – 01:01:11]** structured and unstructured data.

**[01:01:11 – 01:01:15]** Tools let the agent simulate, experiment, and act.

**[01:01:15 – 01:01:19]** NVIDIA OpenShell controls what agents can access and execute.

**[01:01:19 – 01:01:24]** Start with a frontier closed model or customize an open model.

**[01:01:24 – 01:01:27]** With NVIDIA's Agent Toolkit, any company can turn its

**[01:01:27 – 01:01:31]** knowledge and expertise into a customized super agent.

**[01:01:31 – 01:01:34]** Let's look at three examples.

**[01:01:34 – 01:01:37]** For Lyla Sciences, the goal is to design a therapy

**[01:01:37 – 01:01:39]** for B-cell lymphoma.

**[01:01:39 – 01:01:43]** The Lyla Iris agent, built on NVIDIA NeMo Tron, turns that

**[01:01:43 – 01:01:46]** goal into an experimental plan.

**[01:01:46 – 01:01:50]** In the computational dry lab, Lila agents use NVIDIA BioNeMo

**[01:01:50 – 01:01:54]** Agent Toolkit to call models for protein design and RNA

**[01:01:54 – 01:01:58]** optimization, running millions of virtual experiments.

**[01:01:58 – 01:02:02]** The most promising candidates move into Lila's AI science

**[01:02:02 – 01:02:06]** factory, where autonomous robots and instruments, developed

**[01:02:06 – 01:02:09]** with NVIDIA Omniverse and Cosmos, make and test them.

**[01:02:10 – 01:02:14]** The results return to IRIS, shaping the next designs and experiments.

**[01:02:14 – 01:02:17]** The loop continues.

**[01:02:17 – 01:02:20]** At Siemens Electronic Factory Erlangen, the goal is to bring

**[01:02:20 – 01:02:22]** new products into production.

**[01:02:22 – 01:02:27]** Process Preparation X Agents use NVIDIA Cosmos to generate virtual

**[01:02:27 – 01:02:31]** prototypes and synthetic data, then train and validate inspection

**[01:02:31 – 01:02:34]** models before production starts.

**[01:02:34 – 01:02:37]** Intelligence Center X Agents, built with NVIDIA's Factory

**[01:02:37 – 01:02:42]** Operations Blueprint, use orders, supplies, and line setups

**[01:02:42 – 01:02:45]** to recommend production plans.

**[01:02:45 – 01:02:49]** Digital Twin Composer uses NVIDIA Omniverse libraries to validate

**[01:02:49 – 01:02:54]** those plans with physics simulation and real-time factory data before

**[01:02:54 – 01:02:57]** changes reach the factory floor.

**[01:02:57 – 01:03:00]** With the Dassault Systems platform, an engineer needs

**[01:03:00 – 01:03:04]** to adapt a Volkswagen Golf door design for the Tiguan.

**[01:03:04 – 01:03:08]** Aura and Leo use NVIDIA Nematron to reason over

**[01:03:08 – 01:03:10]** engineering requirements.

**[01:03:10 – 01:03:13]** The harness coordinates the specialized sub-agents and

**[01:03:13 – 01:03:17]** preserves context across planning, design, and simulation.

**[01:03:18 – 01:03:21]** AURA uses INNOVIA to plan the work around the vehicle's

**[01:03:21 – 01:03:25]** dimensions, weight limits, and other engineering requirements.

**[01:03:25 – 01:03:29]** LEO Design Agent selects a compatible golf door design

**[01:03:29 – 01:03:33]** and uses CATIA to reshape it for the Tiguan.

**[01:03:33 – 01:03:36]** LEO Simulation Agent uses Simulia, accelerated by NVIDIA CUDA

**[01:03:36 – 01:03:41]** X libraries, to test the variants, revealing a vibration problem.

**[01:03:41 – 01:03:45]** The results return to the design agent for refinement.

**[01:03:45 – 01:03:50]** The engineer reviews the refined designs and selects the final one.

**[01:03:50 – 01:03:51]** Different goals.

**[01:03:51 – 01:03:53]** Different expertise.

**[01:03:53 – 01:03:55]** The same building blocks.

**[01:03:55 – 01:04:00]** Open Models, CUDA X Accelerated Tools, and the OpenShell Secure

**[01:04:00 – 01:04:05]** Runtime, brought together with the NVIDIA Agent Toolkit.

**[01:04:06 – 01:04:09]** Isn't that great?

**[01:04:12 – 01:04:15]** Some people said that AI is here, it's going to be

**[01:04:15 – 01:04:16]** the end of software.

**[01:04:16 – 01:04:18]** It's going to be end of tools company.

**[01:04:18 – 01:04:20]** It's exactly the opposite.

**[01:04:20 – 01:04:25]** AI is here, there'll be more users of these tools that we

**[01:04:25 – 01:04:28]** created over time than ever before.

**[01:04:28 – 01:04:31]** Every single tools company will now become an AI company.

**[01:04:31 – 01:04:32]** They'll be an agents company.

**[01:04:32 – 01:04:34]** Not only will they offer tools, but they'll offer

**[01:04:35 – 01:04:39]** you specially optimized agents that are incredibly

**[01:04:39 – 01:04:40]** good at using their tools.

**[01:04:40 – 01:04:44]** Cadence and CrowdStrike and Dassault, IQVIOUS, and just

**[01:04:44 – 01:04:48]** about every enterprise company out there are all building

**[01:04:48 – 01:04:50]** agents, and we're just delighted to be working with them.

**[01:04:50 – 01:04:53]** Building agents is not that hard.

**[01:04:53 – 01:04:55]** It's got five components.

**[01:04:55 – 01:04:59]** It's got the model, the harness, both of them have closed and open

**[01:04:59 – 01:05:05]** versions, but the OpenShell secure runtime, and then most importantly,

**[01:05:05 – 01:05:08]** the knowledge and tools, and those knowledge and tools belong to you.

**[01:05:08 – 01:05:10]** That's your expertise.

**[01:05:10 – 01:05:14]** And so all of this is now possible to do, and I'll show

**[01:05:14 – 01:05:16]** you more in just a second.

**[01:05:16 – 01:05:18]** But before I do that.

**[01:05:18 – 01:05:19]** Let's talk about physical AI.

**[01:05:19 – 01:05:23]** That's really the next wave of AI.

**[01:05:23 – 01:05:27]** AI that understands the physical world, can reason about it because

**[01:05:27 – 01:05:31]** the physical world is diverse, it is complicated, it's got extremely

**[01:05:31 – 01:05:38]** long tails, and to reason about it and use appropriate tools to

**[01:05:38 – 01:05:41]** interact with the physical world.

**[01:05:41 – 01:05:42]** Whenever people think about physical worlds,

**[01:05:42 – 01:05:44]** they think robotics.

**[01:05:44 – 01:05:48]** In fact, one of my favorite physical world applications is

**[01:05:48 – 01:05:52]** the radio cell phone cell towers.

**[01:05:52 – 01:05:54]** And the reason for that is this.

**[01:05:54 – 01:05:58]** It uses, instead of cameras, instead of speakers and

**[01:05:58 – 01:06:05]** microphones, it uses radio waves to sense the world.

**[01:06:05 – 01:06:08]** It processes information, reasons about it, and interacts

**[01:06:08 – 01:06:11]** with the world using beamforming.

**[01:06:11 – 01:06:17]** So in fact, this is an application, a technology that is just

**[01:06:17 – 01:06:19]** perfectly suited for AI.

**[01:06:19 – 01:06:23]** AI will revolutionize the way we do telecommunications.

**[01:06:23 – 01:06:26]** It will enhance our spectrum efficiency tremendously, just

**[01:06:26 – 01:06:29]** as we've done so with ray tracing.

**[01:06:29 – 01:06:32]** All of the graphics that you've seen completely 100% ray traced,

**[01:06:32 – 01:06:35]** the amount of computation we thought necessary in the

**[01:06:35 – 01:06:40]** past and now that we apply AI to is a million to one difference.

**[01:06:40 – 01:06:43]** And so we're going to bring to the telecommunications

**[01:06:43 – 01:06:47]** network an enormous step up in spectrum efficiency.

**[01:06:47 – 01:06:50]** But beyond that, we're also going to revolutionize all

**[01:06:50 – 01:06:51]** of these cell towers.

**[01:06:51 – 01:06:54]** There's 100,000 sites around the world, around the Western

**[01:06:54 – 01:06:59]** world, about 100 gigawatts of distributed data centers

**[01:06:59 – 01:07:04]** that will in the future become a distributed AI infrastructure.

**[01:07:04 – 01:07:07]** We're delighted to partner with many companies here in Europe.

**[01:07:07 – 01:07:11]** And just this last week, after working on this for nearly

**[01:07:11 – 01:07:16]** eight years, NVIDIA made the first commercial phone call

**[01:07:16 – 01:07:19]** on the T-Mobile network with Nokia.

**[01:07:19 – 01:07:24]** And so now CUDA makes phone calls the traditional way.

**[01:07:24 – 01:07:28]** Then once you can do that, then we can layer AI on top of it.

**[01:07:28 – 01:07:30]** And so the revolution is going to begin.

**[01:07:30 – 01:07:32]** This is really exciting opportunity for us.

**[01:07:32 – 01:07:37]** The other physical AI application that I'm super excited about is

**[01:07:37 – 01:07:40]** essentially robotic radiotherapy.

**[01:07:40 – 01:07:43]** The tumor is changing in shape.

**[01:07:43 – 01:07:47]** The physical body is changing while you're in therapy.

**[01:07:47 – 01:07:51]** The robot, this adaptive radiotherapy system, has

**[01:07:51 – 01:07:55]** to perceive the changing environment, understand

**[01:07:55 – 01:08:00]** where the tumor is, segment it just perfectly, and precisely direct

**[01:08:00 – 01:08:02]** the radiotherapy accordingly.

**[01:08:02 – 01:08:06]** Essentially, robotic, that's right.

**[01:08:11 – 01:08:14]** This is a form of robotic surgery.

**[01:08:14 – 01:08:18]** Now, of course, the highest volume, the most near-term

**[01:08:18 – 01:08:21]** robotics application or physical AI application that everybody

**[01:08:21 – 01:08:24]** recognized has been something we've been working on for over

**[01:08:24 – 01:08:30]** a decade, and now, we recognize that every single car company

**[01:08:30 – 01:08:35]** will want to make their cars part of the ride-hailing fleet.

**[01:08:35 – 01:08:40]** Today, we have ride-hailing and it's part of Uber and

**[01:08:40 – 01:08:44]** here in Europe, we have Bolt and we have Grab and Lyft and

**[01:08:44 – 01:08:48]** so on and so forth and all these different ride-hailing services.

**[01:08:48 – 01:08:51]** In the future, there'll be robo taxis that will

**[01:08:51 – 01:08:54]** be connected to those ride-hailing services as well, but it

**[01:08:54 – 01:08:57]** won't be from just one brand.

**[01:08:57 – 01:09:00]** And so we recognized that building these systems are incredibly

**[01:09:00 – 01:09:05]** expensive, particularly the software that goes on top of it.

**[01:09:05 – 01:09:10]** We decided to create one standard architecture.

**[01:09:10 – 01:09:15]** 14 different cameras, nine radar systems, three lidars,

**[01:09:15 – 01:09:20]** umpteen ultrasonics, this cocoon of a shell of sensors

**[01:09:20 – 01:09:27]** powered by dual NVIDIA GPU systems for the AV computing system.

**[01:09:27 – 01:09:29]** We turned this into an industry standard, and

**[01:09:29 – 01:09:34]** now it's being designed into one car fleet after another.

**[01:09:34 – 01:09:38]** On top of it are AV software companies, and we are, in

**[01:09:38 – 01:09:43]** fact, one of them as well, to build an autonomous robotaxi

**[01:09:43 – 01:09:48]** capability, which then is connected into the ride-hailing services.

**[01:09:48 – 01:09:50]** This ecosystem is being built out.

**[01:09:50 – 01:09:54]** The world travels about a trillion miles per year around the world.

**[01:09:54 – 01:09:57]** Many of those miles in the future, you know, a dollar or two per

**[01:09:57 – 01:10:01]** mile, are going to be on robotaxis.

**[01:10:01 – 01:10:07]** So this is one giant, but it's incredibly fragmented, and

**[01:10:07 – 01:10:11]** so creating an industry standard, Hyperion, and working with an

**[01:10:11 – 01:10:16]** entire ecosystem, we could create essentially a giant virtual fleet

**[01:10:16 – 01:10:19]** that everybody can connect into.

**[01:10:19 – 01:10:23]** I'm super proud to announce that four more brands, four

**[01:10:23 – 01:10:26]** more cars are going to, lines are going to come in.

**[01:10:26 – 01:10:30]** The VW van, I just can't imagine anything more iconic.

**[01:10:31 – 01:10:33]** The JLR, brand new car.

**[01:10:33 – 01:10:38]** The Volvo fleet, as well as Mercedes next generation,

**[01:10:38 – 01:10:41]** are going to be also Hyperion compatible and be part of

**[01:10:41 – 01:10:46]** this network, this ride-hailing network, Robotaxi fleet.

**[01:10:46 – 01:10:50]** So, welcome to Hyperion.

**[01:10:54 – 01:10:58]** The next major industry, the next major industry, the next

**[01:10:58 – 01:11:02]** great industry, and this one is likely to be extremely

**[01:11:02 – 01:11:03]** large, is human or robotics.

**[01:11:04 – 01:11:07]** There's a reason why robotics is hard to do, because all the

**[01:11:07 – 01:11:10]** different work is rather diverse.

**[01:11:10 – 01:11:14]** And so building a special configuration for each one

**[01:11:14 – 01:11:19]** of the diverse tasks segments the robotics industry into all

**[01:11:19 – 01:11:22]** of these different configurations, which prevents the technology

**[01:11:22 – 01:11:24]** flywheel to happen.

**[01:11:24 – 01:11:26]** The reason why the cell phone has advanced so quickly, the

**[01:11:26 – 01:11:29]** reason why personal computers have advanced so quickly,

**[01:11:29 – 01:11:33]** the reason why NVIDIA's NVLink 72 has advanced so quickly is because

**[01:11:33 – 01:11:37]** those standardized configurations allows many customers to use

**[01:11:37 – 01:11:42]** it, the volume to therefore be derived, that volume drives

**[01:11:42 – 01:11:47]** the technology flywheel and it gets faster, better, incredibly fast.

**[01:11:47 – 01:11:50]** And so we need a robotics.

**[01:11:50 – 01:11:54]** This is a configuration that is as standard as possible.

**[01:11:55 – 01:11:58]** Now, human robotics obviously has the ability.

**[01:11:58 – 01:12:01]** It probably isn't perfect for everything, but it's likely

**[01:12:01 – 01:12:03]** to work for many things.

**[01:12:03 – 01:12:07]** As a result, this is likely going to be the largest Volume

**[01:12:07 – 01:12:08]** robotics configuration.

**[01:12:08 – 01:12:12]** Now, if we could just create the AI that allows it to learn

**[01:12:12 – 01:12:17]** very quickly, then for small medium businesses who don't have

**[01:12:17 – 01:12:22]** an entire robotics team to teach and program the robot, can now show

**[01:12:22 – 01:12:23]** the robot what needs to be done.

**[01:12:23 – 01:12:27]** The robot, because it has general capabilities and general

**[01:12:27 – 01:12:30]** intelligence, would be able to learn and adapt quickly.

**[01:12:30 – 01:12:35]** And so this is likely to be the next major platform,

**[01:12:35 – 01:12:36]** technology platform.

**[01:12:36 – 01:12:37]** We're working with so many different companies.

**[01:12:37 – 01:12:40]** You will recognize many of them here in Europe.

**[01:12:40 – 01:12:43]** This is your great opportunity.

**[01:12:43 – 01:12:49]** No country, no region has dedicated itself to making mechatronics,

**[01:12:49 – 01:12:53]** electronics, mechanical systems at the level of precision

**[01:12:53 – 01:12:56]** and craft as the European region.

**[01:12:56 – 01:13:00]** So many companies here are incredibly good at manufacturing.

**[01:13:00 – 01:13:03]** This is going to be the next high volume industry.

**[01:13:03 – 01:13:06]** And you could see there are so many companies that are now part of it.

**[01:13:06 – 01:13:10]** And we're delighted to be partnered with you.

**[01:13:17 – 01:13:23]** Now, in order to create a physical AI system, you

**[01:13:23 – 01:13:25]** need three computers.

**[01:13:25 – 01:13:27]** The first computer is the one we've been talking about

**[01:13:27 – 01:13:31]** and everybody knows quite well, which is training the AI model.

**[01:13:31 – 01:13:35]** The second computer is where we have the AI learn how

**[01:13:35 – 01:13:36]** to do its skills.

**[01:13:36 – 01:13:40]** You can't just have the AI, program the AI, and have the AI come

**[01:13:40 – 01:13:44]** out into the physical world and try to do whatever it wants to do.

**[01:13:44 – 01:13:47]** You do it inside a digital twin.

**[01:13:47 – 01:13:50]** We create a digital twin called Omniverse, and you've seen

**[01:13:50 – 01:13:52]** us talk about that repeatedly.

**[01:13:52 – 01:13:56]** And in fact, just now, when you saw all the CUDAx libraries,

**[01:13:56 – 01:13:59]** they're all being done inside a digital twin world.

**[01:13:59 – 01:14:02]** That digital twin world is called Omniverse.

**[01:14:02 – 01:14:04]** Omniverse is a library that's connected into all kinds of

**[01:14:04 – 01:14:08]** platforms and tools, Siemens and Dassault and Cadence and

**[01:14:08 – 01:14:09]** so on and so forth.

**[01:14:09 – 01:14:12]** And so you could see these digital twin systems being

**[01:14:12 – 01:14:15]** created, the second computer.

**[01:14:15 – 01:14:19]** The third computer is the robotics computer.

**[01:14:19 – 01:14:22]** That computer integrates, in the case of the self-driving car,

**[01:14:22 – 01:14:25]** the robotaxi, we call it Hyperion.

**[01:14:25 – 01:14:31]** Hyperion consists of the computer itself, the functional safety

**[01:14:31 – 01:14:36]** system, the active safety system, an operating system

**[01:14:36 – 01:14:39]** designed for physical AI safety.

**[01:14:39 – 01:14:43]** Well, we need to do exactly the same for robotic systems so

**[01:14:43 – 01:14:47]** that every single time you start to develop research and start to work

**[01:14:47 – 01:14:50]** on your own robotic systems, you don't have to start from scratch to

**[01:14:50 – 01:14:54]** cobble up all of the software and computing systems and the sensors

**[01:14:54 – 01:14:56]** and all the mechanical systems.

**[01:14:56 – 01:15:00]** That alone would take several years, and so we decided to

**[01:15:00 – 01:15:02]** build a reference robot.

**[01:15:03 – 01:15:07]** This is the NVIDIA Agile One in partnership with Agile Robotics.

**[01:15:07 – 01:15:11]** This robot is going...

**[01:15:16 – 01:15:18]** This reference robot will be available in the first

**[01:15:18 – 01:15:22]** quarter of next year, and it will come pre-assembled with computer

**[01:15:22 – 01:15:27]** inside, software integrated, and it should be able to perform

**[01:15:27 – 01:15:29]** quite well right out of the bat.

**[01:15:29 – 01:15:32]** And so this is a great starting point, and it will be the

**[01:15:32 – 01:15:35]** flywheel by which we can accelerate development of robotics.

**[01:15:36 – 01:15:40]** Well, let me start, let me end today's speech from

**[01:15:40 – 01:15:41]** where we started.

**[01:15:41 – 01:15:44]** NVIDIA started in the PC revolution.

**[01:15:44 – 01:15:48]** It was in 1993, and the PC revolution was just about to start.

**[01:15:48 – 01:15:52]** Remember, Windows 95 kicked it into turbocharge.

**[01:15:52 – 01:15:56]** We came since then, invented the GPU, programmable shading,

**[01:15:56 – 01:15:59]** real-time ray tracing.

**[01:15:59 – 01:16:03]** We added CUDA to the computing model and helped ignite the

**[01:16:03 – 01:16:05]** modern AI revolution.

**[01:16:06 – 01:16:09]** And so four years ago, we started to think, what can

**[01:16:09 – 01:16:14]** we do to reimagine the personal computer for this new era,

**[01:16:14 – 01:16:17]** where the personal computer is not just your tool, your personal

**[01:16:17 – 01:16:24]** computer tool, but your personal agent or your personal assistant?

**[01:16:24 – 01:16:27]** And the way you interact with this computer would be essentially

**[01:16:27 – 01:16:33]** a companion, somebody who helps you do work.

**[01:16:33 – 01:16:38]** And so this computer needs to have exactly the five things

**[01:16:38 – 01:16:40]** that I mentioned today.

**[01:16:40 – 01:16:41]** It needs to have model.

**[01:16:41 – 01:16:45]** It could be running locally, or it could be running in the cloud.

**[01:16:45 – 01:16:46]** It needs a harness.

**[01:16:47 – 01:16:50]** It needs a secure runtime, open shell.

**[01:16:50 – 01:16:55]** And it needs access, because it's available on your PC, it needs

**[01:16:55 – 01:16:57]** access to knowledge and tools.

**[01:16:57 – 01:17:01]** And the good news is, because NVIDIA has been accelerating

**[01:17:01 – 01:17:05]** tools for so long, if we could just create a computer that

**[01:17:05 – 01:17:10]** can accelerate OpenGL and DirectX and everything CUDA,

**[01:17:10 – 01:17:14]** everything NVIDIA has ever worked on would run on this computer.

**[01:17:14 – 01:17:15]** And so this is it.

**[01:17:15 – 01:17:18]** After four years of partnership with Microsoft, we have

**[01:17:18 – 01:17:20]** created the RTX Spark.

**[01:17:20 – 01:17:22]** And this is a brand new line of computers.

**[01:17:22 – 01:17:26]** This, every, literally thousands of years, thousands of engineering

**[01:17:26 – 01:17:27]** years went into this.

**[01:17:27 – 01:17:30]** We've been secretly working on it for four years, created

**[01:17:30 – 01:17:35]** a chip called the N1X, and then we just recently announced it with the

**[01:17:35 – 01:17:38]** entire PC industry joining forces.

**[01:17:38 – 01:17:42]** We created desktops, laptops, and workstations.

**[01:17:42 – 01:17:45]** This is one petaflops, one petaflops.

**[01:17:45 – 01:17:50]** Just to put that in perspective, in the year 2016, I delivered

**[01:17:51 – 01:17:57]** one petaflops and 128 gigabytes of memory, one petaflops and

**[01:17:57 – 01:18:01]** 128 gigabytes, to a small company called OpenAI.

**[01:18:01 – 01:18:06]** This is one petaflops and 128 gigabytes.

**[01:18:14 – 01:18:17]** The only difference, of course, is that I can hold this like

**[01:18:17 – 01:18:22]** this, and DGX One took two people to carry upstairs.

**[01:18:22 – 01:18:25]** And so this is the brand new RTX Spark PC.

**[01:18:25 – 01:18:28]** And if you would like to have a companion workstation that runs

**[01:18:28 – 01:18:33]** CATIA and everything, the SO and everything, Autodesk and, you know,

**[01:18:33 – 01:18:38]** Adobe and Cadence and Synopsys, you can always get yourself

**[01:18:38 – 01:18:42]** a DGX Station that runs Windows.

**[01:18:42 – 01:18:49]** And this is 20 petaflops and 754 gigabytes of memory.

**[01:18:49 – 01:18:50]** And so, incredible.

**[01:18:50 – 01:18:53]** And of course, there's also the desktop version.

**[01:18:53 – 01:18:57]** All of these are now, these are now available, and I think

**[01:18:57 – 01:19:02]** the DGX Station with RTX Spark, with DGX Spark, DGX Station

**[01:19:02 – 01:19:05]** with Windows, will be available at the end of this year.

**[01:19:05 – 01:19:09]** Now the question is, so far I just talked about the five

**[01:19:09 – 01:19:14]** components of an agent, and now the question is, how do you build it?

**[01:19:14 – 01:19:17]** How do you build an agent that runs, let's say, on one

**[01:19:17 – 01:19:21]** of those computers, or runs in a cloud, or runs on-prem somewhere?

**[01:19:21 – 01:19:25]** Well, the answer is it's really easy, and you do it this way.

**[01:19:25 – 01:19:31]** The way you build an agent is you ask an agent to build it for you.

**[01:19:37 – 01:19:45]** So this, this is the prompt that is going to run the next

**[01:19:45 – 01:19:48]** application, and what you're about to see is this.

**[01:19:48 – 01:19:52]** You're about to see a robotics application, and while you're

**[01:19:52 – 01:19:55]** watching this demo, just remember, there's an agent running on

**[01:19:55 – 01:19:58]** this personal computer.

**[01:19:58 – 01:20:02]** This demo runs across three computers.

**[01:20:02 – 01:20:04]** The first computer is to train.

**[01:20:04 – 01:20:07]** The second is a digital twin.

**[01:20:07 – 01:20:13]** And the third is to put that AI into the computer itself.

**[01:20:13 – 01:20:16]** Okay, so two ideas.

**[01:20:16 – 01:20:19]** The first idea is agents, how do you build it and run

**[01:20:19 – 01:20:22]** it and what you can use it to do?

**[01:20:22 – 01:20:26]** And the second is Physical AI Computing Model, which

**[01:20:26 – 01:20:28]** consists of three computers.

**[01:20:28 – 01:20:31]** Let's take a look.

**[01:20:40 – 01:20:43]** Meet Microduck, Hugging Face's robot.

**[01:20:43 – 01:20:49]** Let's teach it to see and walk with NVIDIA RTX Spark.

**[01:20:50 – 01:20:54]** I use Hermes as my agent harness, Nemotron and OpenAI's GPT-6

**[01:20:54 – 01:20:59]** Astra as my models, and CUDAx accelerated tools FreeCAD,

**[01:20:59 – 01:21:03]** Blender, and IsaacSim, Newton Physics Engine, and Isaac Lab.

**[01:21:03 – 01:21:06]** I'll run everything on my RTX Spark, except when I need

**[01:21:06 – 01:21:09]** to burst larger workloads to cloud.

**[01:21:09 – 01:21:13]** First, I describe the course and what my microDuck needs to learn.

**[01:21:14 – 01:21:18]** On RTX Spark, Hermes and Nemotron plan the course,

**[01:21:18 – 01:21:21]** listing every object required.

**[01:21:23 – 01:21:29]** Astra takes it from there, using FreeCAD to model 3D objects,

**[01:21:29 – 01:21:32]** Blender to apply materials,

**[01:21:33 – 01:21:37]** and Omniverse SimReady libraries to add physical properties.

**[01:21:38 – 01:21:42]** In Isaac Sim, Astra composes the scene.

**[01:21:45 – 01:21:50]** Astra rigs Microduct's joints, actuators, and links

**[01:21:51 – 01:21:57]** And simulates how Microduck moves with Newton Physics Engine.

**[01:21:57 – 01:22:00]** Now, training begins.

**[01:22:01 – 01:22:03]** In the cloud, thousands of Microducks train in

**[01:22:03 – 01:22:07]** parallel in Isaac Lab.

**[01:22:10 – 01:22:14]** Every fall, a lesson

**[01:22:16 – 01:22:19]** Improving the policy every time.

**[01:22:19 – 01:22:23]** Omniverse libraries simulate Microduck's cameras,

**[01:22:23 – 01:22:25]** teaching it to see.

**[01:22:25 – 01:22:29]** With perception and physics working together, my Microduck can avoid

**[01:22:29 – 01:22:35]** obstacles, navigate tight spaces, and adapt to its surroundings.

**[01:22:42 – 01:22:46]** I download the policy to RTX Spark and use Nemotron and

**[01:22:46 – 01:22:52]** Isaac Sim to evaluate my microduck with multi-robot orchestration.

**[01:22:55 – 01:23:01]** I deploy the validated policy to the real duck, and he's on his way.

**[01:23:01 – 01:23:06]** NVIDIA RTX Spark, build and run AI agents and their accelerated

**[01:23:06 – 01:23:09]** tools right on your laptop.

**[01:23:19 – 01:23:21]** Hey guys!

**[01:23:22 – 01:23:24]** Oh, yeah?

**[01:23:24 – 01:23:25]** Okay, come on.

**[01:23:25 – 01:23:27]** Here, come over here.

**[01:23:27 – 01:23:30]** Show people you can walk.

**[01:23:31 – 01:23:34]** And you have a camera.

**[01:23:39 – 01:23:41]** Don't trip.

**[01:23:42 – 01:23:44]** Can you guys sit down?

**[01:23:52 – 01:23:54]** Okay, stand up.

**[01:23:54 – 01:23:56]** Take a nap.

**[01:23:56 – 01:23:58]** Take a nap.

**[01:23:59 – 01:24:02]** You need to take a nap.

**[01:24:03 – 01:24:05]** Okay, stand up.

**[01:24:05 – 01:24:08]** Did you know that you were created in here?

**[01:24:08 – 01:24:11]** You were created on RTX Spark.

**[01:24:11 – 01:24:16]** I know, and trained you in the cloud on a whole bunch

**[01:24:16 – 01:24:18]** of NVIDIA supercomputers.

**[01:24:18 – 01:24:21]** And you learned your skills right here on RTX Spark in a digital

**[01:24:21 – 01:24:23]** twin, the Omniverse digital twin.

**[01:24:23 – 01:24:26]** Isn't that incredible?

**[01:24:26 – 01:24:29]** I know. I know.

**[01:24:29 – 01:24:31]** You guys are so cute.

**[01:24:31 – 01:24:40]** And your parents are called Pollen, P-O-L-L-E-N, is that right?

**[01:24:40 – 01:24:41]** Uh-huh.

**[01:24:41 – 01:24:43]** Pollen Robotics.

**[01:24:43 – 01:24:49]** And apparently, everybody loves you guys so much, you're sold out.

**[01:24:50 – 01:24:53]** I know, so if you go to DePaulon website, everybody,

**[01:24:53 – 01:24:55]** everybody watching this, you know, there's gonna be

**[01:24:55 – 01:24:59]** like 20 million viewers, everybody watching this, go to DePaulon

**[01:24:59 – 01:25:02]** website and place an order.

**[01:25:02 – 01:25:05]** I think this startup company is like 10 people.

**[01:25:05 – 01:25:07]** They might blow up.

**[01:25:08 – 01:25:09]** It's gonna be so exciting.

**[01:25:10 – 01:25:11]** Everybody should have one.

**[01:25:11 – 01:25:12]** You guys are so cute.

**[01:25:12 – 01:25:14]** And so we'll just have an RTX Spark, and we'll

**[01:25:14 – 01:25:15]** teach you new skills.

**[01:25:15 – 01:25:18]** Isn't that right? Oi, oi.

**[01:25:18 – 01:25:20]** Wait, wait, oh, look at that.

**[01:25:20 – 01:25:21]** That's incredible.

**[01:25:21 – 01:25:24]** Can you do a somersault?

**[01:25:34 – 01:25:35]** That's so good.

**[01:25:35 – 01:25:38]** That's so good.

**[01:25:38 – 01:25:40]** Wow.

**[01:25:40 – 01:25:42]** Okay, well, guess what?

**[01:25:42 – 01:25:45]** Okay, so today, what did we talk about?

**[01:25:45 – 01:25:50]** We talked about Europe building its ai economy, and it starts

**[01:25:50 – 01:25:54]** with infrastructure, and we announced a brand new, brand

**[01:25:54 – 01:25:57]** new system called dgx pod.

**[01:25:57 – 01:25:59]** It makes it easy for you to build ai factories, and these

**[01:25:59 – 01:26:04]** ai The factories don't just run models anymore.

**[01:26:04 – 01:26:07]** Pay attention, stay focused.

**[01:26:07 – 01:26:10]** Come here, come here, come over here, stay focused.

**[01:26:10 – 01:26:15]** And these AI factories don't just run, there you go, sit down, yeah.

**[01:26:15 – 01:26:20]** They don't just run AI models anymore, they run entire agents.

**[01:26:20 – 01:26:23]** And these agents are consistent of five things.

**[01:26:23 – 01:26:25]** There's the large language model that does the perception,

**[01:26:26 – 01:26:31]** reasoning, planning, and acting, and it runs in conjunction

**[01:26:31 – 01:26:36]** with a harness that turns it essentially exoskeleton, and that

**[01:26:36 – 01:26:41]** harness sits inside, runs inside a secure runtime we call OpenShell,

**[01:26:41 – 01:26:45]** which has just so many people that have supported it, and you give

**[01:26:45 – 01:26:48]** it access to knowledge and tools.

**[01:26:48 – 01:26:53]** And this agent runs on top of these computers, but we're

**[01:26:53 – 01:26:56]** also extending into the next generation, which is where

**[01:26:56 – 01:26:59]** you guys come in, physical AI.

**[01:26:59 – 01:27:01]** So that's it. That's what we talked about.

**[01:27:01 – 01:27:03]** Everybody can build agents.

**[01:27:04 – 01:27:08]** Every company in Europe will become an AI company.

**[01:27:08 – 01:27:11]** There's no question everybody will become an AI company,

**[01:27:11 – 01:27:15]** and we're gonna have a thriving AI economy here in Europe.

**[01:27:15 – 01:27:19]** So after 33 years, what an incredible journey.

**[01:27:19 – 01:27:22]** And we made a really cool video.

**[01:27:22 – 01:27:24]** You guys wanna watch it together?

**[01:27:24 – 01:27:26]** Okay, come on, come over here.

**[01:27:26 – 01:27:28]** Just come over here, just a little bit.

**[01:27:28 – 01:27:31]** Well, get out of the way.

**[01:27:32 – 01:27:35]** Top speed, top speed.

**[01:27:37 – 01:27:40]** How the heck are we supposed to follow that?

**[01:27:40 – 01:27:46]** Well, there is one thing that can make this crowd go even crazier.

**[01:27:46 – 01:27:51]** What is it?

**[01:27:52 – 01:27:54]** You won't believe this!

**[01:27:54 – 01:27:59]** GTC has just gone full Eurovision!

**[01:28:00 – 01:28:05]** Keynotes done in Berlin, AI's next chapter begins.

**[01:28:05 – 01:28:10]** Agents working by your side, Europe's moment has arrived.

**[01:28:10 – 01:28:15]** But if you miss what we said at GTC, we'll break it down for you.

**[01:28:15 – 01:28:17]** Germany!

**[01:28:27 – 01:28:30]** Scaled and caved data centers, now we've got mail.

**[01:28:30 – 01:28:33]** Now we're way past turning coal into watts.

**[01:28:33 – 01:28:36]** AI factories turn sand into thoughts.

**[01:28:36 – 01:28:38]** Move over oil, the S&P too.

**[01:28:38 – 01:28:42]** There's something more investable, our GPU.

**[01:28:42 – 01:28:45]** It works like an asset, producing green.

**[01:28:45 – 01:28:47]** Looks like a chip, acts like a cash machine.

**[01:28:47 – 01:28:50]** They make revenue while you relax.

**[01:28:50 – 01:28:55]** The cash flows smoother than this sand.

**[01:29:08 – 01:29:11]** Agents code and make decisions, give them freedom but

**[01:29:11 – 01:29:13]** control their missions.

**[01:29:13 – 01:29:17]** OpenShell on bearers checking the list, no credentials,

**[01:29:17 – 01:29:18]** access dismissed.

**[01:29:18 – 01:29:22]** Sentry on BlueField never clocks off, protecting the

**[01:29:22 – 01:29:23]** sandbox like Hasselhoff.

**[01:29:24 – 01:29:27]** We advance AI but limit risks.

**[01:29:27 – 01:29:30]** We're responsibly optimistic.

**[01:29:30 – 01:29:33]** RTX Spark keeps agents safe and bright, our spark rates

**[01:29:33 – 01:29:35]** keep the price just right.

**[01:29:35 – 01:29:38]** We do quantum rexes, gaming all in one, cause you can't

**[01:29:38 – 01:29:42]** spell fungible without fun.

**[01:29:44 – 01:29:48]** Not a body now, it's learning the rules of gravity, friction,

**[01:29:48 – 01:29:49]** and using real tools.

**[01:29:49 – 01:29:51]** With Isaac Lab and Cosmos, they train on their own.

**[01:29:51 – 01:29:55]** It's like college for robots, without student loans.

**[01:29:55 – 01:29:59]** DSX pods keep AI in sync, less power, more brains, like

**[01:29:59 – 01:30:00]** Legos that think.

**[01:30:00 – 01:30:03]** QVS, Name, Atron, and Kumo are here, throwing

**[01:30:03 – 01:30:05]** knowledge faster than a...

**[01:30:05 – 01:30:11]** This video was made possible in part by...

**[01:30:25 – 01:30:28]** Let's see!

**[01:30:58 – 01:31:01]** Come on to GTC!

**[01:31:16 – 01:31:19]** Welcome to GTC.

**[01:31:19 – 01:31:20]** I want to thank all of you.

**[01:31:20 – 01:31:23]** I want to thank all of you, all the sponsors.

**[01:31:23 – 01:31:25]** You can see all the names of companies that are here,

**[01:31:25 – 01:31:28]** all the partners we've been working with for 33 years,

**[01:31:29 – 01:31:33]** micro duck, my new friends.

**[01:31:33 – 01:31:39]** Um, let's work together to build Europe's AI economy

**[01:31:39 – 01:31:41]** and put AI to work.

**[01:31:41 – 01:31:42]** All right, get ready. Have a great GTC.