Transcript of NVIDIA GTC Berlin 2026: Jensen Huang Full Keynote

Energy powers advanced computers that are part of an AI factory that generates tokens, digital intelligence from the models that run on it, that are integrated into every application.

A five layer cake. To be part of this AI economy, you don’t want to just be consuming the AI. You want to be producing the AI. We could do all of that here in Europe. And in fact, we’ve already started.

Europe is building its AI economy. And it’s five layer cake represented here. We’ve got great partnerships. And when you look across this five layer cake, you can recognize that in fact, it is a new industry because there’s growth at every single layer. We’re creating jobs, creating growth at extraordinary rate.

Companies here in Europe, Schneider, Siemens Electric, Siemens Energy, doing incredibly well around the world. We’re the computing layer. We integrate into AI factories that are now flourishing all over Europe. Start up companies like Nebius is doing incredibly well. A new company called Firebird, which is setting up a new cloud in the Central Asia region.

And even Cassava in Africa. Every single telco in the future, and I predict this, every single telco in the future with a 100,000 sites representing some 100 gigawatts of distributed telco base stations and data centers, will in the future evolve into a distributed AI network.

So the AI infrastructure layer is being built out. Which is then above that, the AI models. It’s really important to recognize that long before the AI models emerge, the infrastructure has to exist.

Just as before you can have an Internet, you have to have the networking. Just as before you could really enjoy the capabilities of electricity, you have to have electrified houses, electrified power plants. And so the models are the next layer.

It’s also interesting to observe that the model layer is not just language models. It represents language, biology, chemicals, physics, robotic actions, animation. And it represents models of all sizes.

These different industries are going to be powered by different modalities of AI models, which then integrate into the AI natives, the startups. And there are so many here in Europe. So the big… The Europe’s AI economy is starting and it’s really accelerating.

It’s great to see.

NVIDIA’s Full-Stack AI Factory Platform

At the foundation of it, where we contribute, most people think is GPUs. It is true. We started in 1993 building a graphics chip, which emerged into the GPU, the programmable shading GPU. Over the time… Over time, we added CUDA acceleration to that, reinvented computing as we know it, which then brought forth artificial intelligence.

But NVIDIA today is really a full stack, end to end, complete AI factory platform. The breadth of the technology of our company over the years has really expanded. Everything from GPU computing, Vera Rubin is our latest generation. To the NVLink switches that connect them into one giant GPU.

The Vera BlueField-4 compute that we recently introduced, and I’ll say a few words about that later. To Spectrum-X, which is our ethernet solution, making NVIDIA the largest networking company in the world today. To Groq 3, a decode accelerator for generation text… For token generation.

To Vera… Vera BlueField-4 storage acceleration, which I’ll say a few words in that in just a second. And then, Vera RTX compute, and I’ll mention that in just a second. But this now represents the breadth of technology necessary to create the AI factory.

But nothing works unless you have software that activates it all. The first part is the system software that runs the AI factory. These systems are gigantic. A one gigawatt data center, one gigawatt AI factory costs about $50 billion.

Land, power, shell, all of the compute, networking, storage, power, cooling, all connected together, multiple acres of AI factory. Just turning it on and operating it is a marvel in itself. The magnificent amount of software necessary to energize this factory, to put it to use and produce tokens is an extraordinary feat. NVIDIA offers the entire full stack software.

On top of that, are what we call the CUDA-X libraries. The algorithms of our company. What NVIDIA truly is, is an algorithm company that designs silicon to accelerate them. Of course, one of them is cuDNN. What many researchers recognize as the library that brought forth, that started the big bang of modern AI, cuDNN.

cuDNN is to AI what SQL is to the world’s database industry. It’s a domain specific language, a domain specific library. Just as OpenGL and DirectX enabled real time computer graphics, cuDNN enabled the AI computing industry.

350 other libraries, those libraries are in fact the crown jewels of our company. Everything from computational lithography to signal processing, image processing, data processing, particles, fluids, sparse solvers.

Incredible breadth of libraries that many of you in the audience have connected into your company. This is how we open new markets. This is how we connect NVIDIA into algorithms that enable you to do computation at incredible speeds.

And then finally, on the right hand side, on that side, the models. This is the new software. The new engine of new applications. And it sits on top of and it collaborates with the CUDA-X libraries. These models NVIDIA had been working on across multiple domains. With these models, it opens up new opportunities. And I have lots to say about that later on.

Well, every single generation, NVIDIA started with Pascal. The AI era started with the Pascal new, what is called DGX-1 system. DGX-1 brought to the world the world’s first AI supercomputer. I introduced it ten years ago. Delivered the first one to OpenAI. Started their computation journey.

From Pascal to Volta to Ampere, Then to many of you who heard about us for the very first time, Hopper, the current generation Blackwell, and then now Vera Rubin in full production. Notice six generations of architectures. With each generation, we advanced the technology. In the course of that last ten years, we’ve improved AI performance by 1 million times.

If you would take a step back and think about what would Moore’s Law have done, in ten years’ time, it would have… The first five years would have improved by a factor of 10, the second five years by another factor of 10, a 100 x improvement in ten years’ time, in the same period of time.

Because of the way NVIDIA extreme co-designs. We’re innovating across every single computer, across every single stack of the software, optimizing new algorithms, creating new algorithms while creating architectures to accelerate it. Working across the entire domain, what we call extreme co-design, we were able to accelerate AI by a million times.

And we’re going to continue to do so. This is what NVIDIA builds today. And this is the foundation by which we can enable Europe to be part of the AI ecosystem. The AI economy.

Now, each generation… In the beginning, Pascal was GPUs. Volta was GPUs. Ampere was GPUs. And Hopper was GPUs. We invented the new DGX form factor. Then it was Blackwell that introduced to the world the first rack scale computing. And I’ll show you that in just a moment. Rack scale computing turned one entire system into one giant computer, one GPU.

And now with Vera Rubin, we’re extending it to processing beyond models, beyond large language models to now processing the entire agent. Now, we’ll refer back to the slide later.

But remember this, the agent has to think. The agent has to think And that’s where the model comes in. Perception, reasoning, planning, and using tools. All of that runs on the large language model. And the models are designed to do that.

But the agent also has what is called a harness. Basically, this exoskeleton of the digital agent. Just last, we have the brain. The rest of our body represents our exoskeleton. The agent is the exoskeleton that orchestrates the work that allows the agent to actually engage memory, active memory, working memory, as well as long term memory, and activating tools.

Then the third part is a computer that processes knowledge designed for data processing. And then lastly, a computer dedicated to CUDA-X libraries. And its job is to run tools. And these computers now represent the entire agentic system.

Vera Rubin in Full Production

So Vera Rubin is in full production. It’s nice to kinda see it. This is the first computer in the world that we’ve created. No cables, no wires, completely a 100% liquid cooled, and it’s just really delightful. Whoops. Hang on.

Really delightful to be able to pull it out. And this is one Vera Rubin computer. Okay. So this has two GPUs. Yep.

What used to take, in the case of Grace Blackwell, two hours to assemble one node, and when you see it, you’ll know what I mean. Cables and wires everywhere, 80% liquid cooled. Now, Vera Rubin is a 100% liquid cooled, and it takes less than five minutes to assemble it. And so incredible, incredible achievement.

Of course, this translates to higher cycle times in manufacturing, more better real… Much much better reliability, and of course, a lot better performance.

This is the NVLink switch tray. And what this does, is it connects 72 of these GPUs. There are four of them in here right now. It connects 72 GPUs into one gigantic GPU. Okay? So, each GPU is 20 petaflops connecting 72 of them into one gigantic system. Okay. Thank you. Thank you.

And this is our new Vera computer. It’s okay. Be… Go easy. Let her… Give her a chance. Everybody has a chance to say something. But today, let me… No. No. Take just… I’ll talk to you later. I’ll talk to you guys later. Thank you. All important issues.

But let me get through my keynote. We have… So this auditorium is packed. We have a 100 other watch parties around the world. A 100 watch parties. 15,000 people are watching us right now. And so I’ll get through the keynote. There’s some really really fun things I want to show you in the end. Okay?

So you just have to… You have to… I have to torture you through some education first. And then the entertainment is coming.

This is our brand new Vera computer. Vera CPU is an extraordinary CPU beast. Eight instructions per clock. It has the world’s first LPDDR memory subs… Memory system, so that we can have high performance, very high bandwidth, as well as extremely excellent energy efficiency.

The on chip fabric is the fastest of any chip in the world. As a result, data processing is off the charts. Agentic sessions starting and processing is off the charts. And this is designed for doing exactly that. This is our Vera computer.

And this computer here, this computer here is the Spectrum-X. This is the world’s first co packaged optics. And you could see it here. I’ll show you. Some of this area falls into the ground, and some of this area you can walk on. You… I’m just not sure which. Alright. So if accidentally disappear, you guys know us. It’s not a stunt. It’s an accident.

And so you could see here, these are the lasers. And these lasers come up through optics and it directly connects to our chip. And it connects using a process technology we invented with TSMC called COUPE. And so silicon photonics are directly attached right onto our chip.

Otherwise, what you would see is all of this would be PC board and PC board’s transmission capability is much worse than optical cables. And so, now have extremely high data rate, extremely high quality data that comes directly onto the chip.

And this goes directly into, otherwise, transceivers without going through a PCB and a transceiver goes right into the optics. Okay? So this is the world’s first CPO. We’re in complete volume production and 512 port at 200 gigabits per second. The world’s most advanced Ethernet switch. And we’re really proud of it. Thank you.

And so what that gives us is the ability to process the thoughts, the large language model. Process the agent harness. Run data processing for its knowledge. And then run accelerated computing for all of the tools. That represents essentially the agent workload, and I’ll show you more in a second.

While Vera Rubin is in full production, and the performance is off the charts. Every single generation, as I mentioned, through extreme co-design, we’re able to achieve performances that are well beyond reason. Well beyond what Moore’s law would have dictated.

And as you know, Moore’s law has really largely ended. And so we need these extreme co-design ideas. Working from chips to systems, systems to system software, to algorithm, to the entire AI factory.

Well, Vera Rubin is another giant step on top of Blackwell. Blackwell was about 30 times faster than Hopper in doing large language model processing. And now, Vera Rubin is another 30 times greater than Blackwell in just two years time. Let’s take a look at how Vera Rubin turns into what is a DSX full stack AI factory.

NVIDIA DSX and DSX POD

VIDEO CLIP BEGINS:

UNIDENTIFIED SPEAKER: AI is creating the largest infrastructure build out in history. The factories are massive. The investments are significant. We have to build it right. But great chips alone don’t make a great AI factory.

NVIDIA DSX is a full stack AI factory reference architecture co innovated with the world’s leading infrastructure companies. Compute, networking, storage, software, power, and cooling, co designed as one system. Optimized from design to operation.

DSX starts in a digital twin. Performance, power, water, and scale validated virtually before construction begins. PTC turns equipment designs into simulation ready assets. Dassault Systèmes connects rack and facility design through model based systems engineering.

Schneider Electric simulates power distribution and liquid cooling. Jacobs brings it all together in the Omniverse Digital Twin. Siemens evaluates thermal, acoustic, and grid impacts. And Procore supports construction phase planning.

Power limits how much an AI factory can produce. DSX MaxLPS is an end to end control loop that dynamically steers power to match workload demand. It uses 30% less power for the same output.

DSX rethinks cooling. NVIDIA Vera Rubin systems are designed for closed loop liquid cooling with coolant temperatures up to 45 degrees Celsius. In suitable climates, outdoor dry coolers release heat into the air without evaporating water.

At two gigawatts, DSX can save 54 million liters of water a day. That’s the daily water use of 180,000 households.

But an AI factory can take several months to build. Until it is operational, it produces no tokens. Today, we are announcing NVIDIA DSX POD, bringing this full stack platform into a factory built, integrated system.

It starts with Vera Rubin NVL72. Seven liquid cooled compute racks connected with 17 kilometers of Spectrum-X cables form a Vera Rubin DSX POD. Vera Rubin DSX PODs are integrated, fully cabled, and factory tested, then delivered in transportable containers.

Dry coolers replace traditional evaporative cooling. 800 volts DC connects to the racks for efficient power delivery. Once site utilities are ready, estimated on-site deployment goes from several months to days.

NVIDIA DSX AI Factory is responsible by design. And now with DSX POD, AI…

VIDEO CLIP ENDS:

JENSEN HUANG: What do you think?

So in 2016, we introduced the world’s first AI supercomputer. It was in one node. One GPU node. Then three years ago, we introduced the Grace Blackwell NVLink 72 rack scale computer. Both times, we had to reinvent not just the computer, the software, and how it works, but also the supply chain.

In the case of NVLink 72, we took the entire rack, which would have otherwise been assembled on-site, and we assembled it inside the supply chain. We essentially manufactured supercomputers at scale.

It is the reason why we’re able to move so fast and be able to provide for literally several 100 data center build outs around the world at the same time. Every single computer manufacturer in the world now manufactures NVLink 72. The scale at which we are now producing computing is unheard of ever. Unimaginable.

Well, even then, when you bring these systems onto your sites, there are still months and months of assembly. And you just heard 17,000 kilometers of wiring to connect just Spectrum-X for the scale out computing. 17 kilometers. 17,000 kilometers. It’s crazy number. And so 17 kilometers. Incredible numbers.

Now suppose we were to do that all inside the supply chain. Run everything up, build everything up, make sure that it runs completely, get it qualified, burnt in. And by the time we ship it, it is fully functional. The amount of cabling and on-site assembly is going to be greatly reduced.

But there’s something much much more. In fact, what’s going to happen here is that by moving this into the supply chain, the amount of idle computing that’s going to be sitting inside at your data centers will reduce dramatically.

Remember this. Each gigawatt data center is $50 billion. The revenue generation you expect is tens of billion dollars per year. And so we just said, it’s $40 billion per year of revenue generation that comes out of that infrastructure. 40 billion if delayed by just three months, that’s $5 billion. Excuse me. What is that? $10 billion worth of revenues. $10 billion worth of revenues will make anybody anxious.

And so, by moving this into the supply chain, the speed by which we can turn on, go from installation, arrival of computing to production of tokens will go from many months. We think something along the lines of four to five months down to about a month. Okay.

So this is a complete transformation of the world supply chain. And this will change the way that people install and build data centers. And hopefully, as a result, velocity goes up. The wasted idle compute, the wasted idle CapEx will reduce and everybody becomes more profitable.

And so this is NVLink 72 now into a large cabinet system. We call it DSX POD. Okay? So this is the next click of NVIDIA computing.

One Architecture, Many Markets

Now, what you make and how you make it makes you special. Just as here in Europe, you… The work you know makes you special. You design machines. You operate factories. You discover medicine. You have extraordinary craft. That craft, that specialty makes you special.

Well, what makes NVIDIA special? Are really two things. The first thing is that our architecture is the only accelerated computing architecture that is general purpose. We don’t specialize on just one arc… One algorithm, for example, transformers.

We specialize across anything that could be parallelized. Could be fluids and particles. It could be quantum. It could be quantum chemistry. Could be image processing. Could be data processing. Structured vector data processing. Processing of all kinds.

And because we’ve been working at it for so long, and because NVIDIA’s architecture has remained consistent for literally two and a half decades, the number of algorithms, the number of domains, the number of industries we serve, we accelerate has grown tremendously.

Our install base, our general purpose architecture, our dedication to this architecture, and our collaboration with all of the world’s industries has made NVIDIA’s architecture literally everywhere.

The second thing is how we do it. Of course, we have the architecture that we developed, but we also build it now in a full stack complete AI factory. There are many companies in the world who want to design their own chips. But most of the companies in the world just wants to deploy the AI infrastructure so that they can run their business on top of it.

The ability to design a chip and build a supercomputer, write all the software, is really completely out of mind for most of the companies around the world. And so we now have the combination of these two ideas, CUDA, full stack AI factories, and it exposes us to markets that otherwise is impossible.

We’re able to partner with so many companies here for this very reason. If I were to go around this, starting with CSPs, this is where most people think the vast majority of AI is today. But as you can see in the beginning of our keynote, I mentioned a whole bunch of companies who are not those. So CSPs remain very large businesses for the world’s computing.

But there are AI natives, companies that are starting all over the world. There are so many that are here. Neo clouds, because every region wants to participate in the AI economy. Every region wants to ensure that the valuable resources that they have, the energy that they have is dedicated to companies in their region.

For example, Mistral here in Europe. Mistral in France. Nebius is here in Europe. So many companies around Europe that are building up Neo Clouds. OEMs and ODMs who are building computers for the world’s enterprise. AI labs.

Of course, we know about last year, all we spoke about was OpenAI. To… Today, we spoke about… We speak about OpenAI, Anthropic, but we also talk about the incredible Meta’s Muse, Grok, and of course, Gemini. So many different AI labs are now participating at the frontier.

And then very importantly, and this is ultimately the most important, for AI to succeed, for Europe to benefit from AI and to have really created an AI ecosystem and AI economy, you need all of the end markets, the end industries to adopt AI and to succeed with AI.

Whether it’s in the automotive industry because we’re building self driving cars. Or the way you manufacture with robotics. The way you discover new drugs in health care. Your financial services in industry. All of these different industries have to become AI companies. Every single one of them.

Well, we have a unique ability through what I just mentioned about CUDA and the way that we build computers. We have a unique ability to serve literally every single segment of AI. And so, one architecture, many markets.

And because these infrastructure are so significant and so expensive, we need to make sure that they are as productive as possible, as soon as possible, that they’re fungible, that you can use it for every workload.

From the beginning of your AI journey, data processing to training to post training, inference and agentic processing. Every single step of the AI journey to every type of AI, chemicals and physics to languages, to every single industry and the algorithms that they use.

And so, NVIDIA’s architecture is fungible. And as a result of that, continuous innovation, continuous updating of software. Because the installed base is so great, we have such an incredibly rich ecosystem of software, then these architectures also durable. Fungible, durable, productive. Well, that lays the foundation. That is the infrastructure. Now, what do we do with it?

The Five Elements of an Agent

This last year, we saw a transition in AI. AI is now productive. And it’s… One of the most important times was in the beginning of the year when we spoke about OpenClaw, which introduced the concept of an open source agent. And it really started what we all see across the world today.

Agents think, reason, iterate, use tools, and as a result, consume an enormous amount of tokens. But as a result, they become useful. So more agents are being developed. They’re more useful and there’s… And therefore they’re used by more people.

And because they have to use a lot of tokens to think, the amount of token consumption has gone through the roof. GPU demand has really gone through the roof. Now, what’s processing it is this. This is an agent. These are the five elements of an agent.

The first part are the language models. These are the tokens that are running on top of the GPUs and they’re generating tokens that are consumed by the agent. What is… I called the exoskeleton. That OpenClaw harness sits inside a sandbox, and that sandbox isolates it from the rest of the system.

This is a very important concept that we want to come back to. There are five things. There’s the model. There’s the harness. The sandbox that the harness sits in. And two other things, knowledge and tools. Models, sandbox, harness, knowledge and tools. These are the five elements of an agent.

OpenShell: Securing AI Agents

Now, before I talk about how to make the agents useful, the most important thing is we have to make sure that the agent is secure. The agent is autonomous. And it’s autonomous system that has access to files, has access to resources, networks, access to tools.

And we need to make sure that the privileges that we provide this autonomous system is minimal. And that we constantly monitor it. We introduced a fantastic new product. A new platform and it’s called OpenShell. OpenShell came just in time.

This is a secure runtime that provides for the things that I just mentioned. And has an additional hardware based monitor. Before I explain it to you, let’s watch a video.

VIDEO CLIP BEGINS:

UNIDENTIFIED SPEAKER: …Enterprises are preparing to deploy billions of AI agents. To be useful, they’ll reason, plan, use tools, access data, and work together. Today, the only thing stopping an agent from misusing all that is asking it nicely. “Please don’t delete anything. Please don’t push to main. Please don’t touch my SSH keys.”

That’s the “please don’t” approach. But hope isn’t security. Safety isn’t an aspiration. It’s technology and engineering. NVIDIA OpenShell and Sentry put that philosophy into practice.

You write a policy for what each agent can touch, and OpenShell enforces it at the granularity of system calls, file access, and network connections. Formal verification checks every policy change against your permission ceiling, so agents never gain more access than intended.

It’s like the early Internet. Viruses were everywhere until browsers enforced strict limits on what a web page could do. OpenShell does the same for agents. Enforcement happens outside the agent, not just in LLM guardrails, bringing agent native security into your infrastructure.

Every agent runs inside its own isolated sandbox with no direct route to the network. Every outbound network request goes through the supervisor outside the sandbox. It checks the action against your policy. If it’s not permitted, it’s denied.

Keys stay outside with the supervisor and are added only on the way to an approved endpoint. A compromised agent can’t leak a key it never had.

The gateway is the control plane. The supervisor makes the decisions. The sandbox holds the agent. And the policy prover checks proposed policies against your permission ceilings. OpenShell combines Linux kernel isolation with runtime policy enforcement.

NVIDIA Sentry adds an optional layer of protection outside the host itself, away from the agent. Running on NVIDIA BlueField-4, Sentry monitors agent activity and enforces security policies in a hardware isolated trust domain.

Because agent reasoning traffic passes through Sentry, if the agent is planning to do something unintended, it will update policies immediately and cut the connection to the model, quarantining it in milliseconds. And it keeps tamper evident records that operators can verify for detection, investigation, and response.

OpenShell is open source under Apache 2.0, runs everywhere your agents do, edge to cloud, and takes two commands to adopt. Install the CLI, create a sandbox, and run your favorite agent.

Agents are only as useful as the access you give them. OpenShell and Sentry put that access under your control. That’s how we advance AI capabilities and safety together.

VIDEO CLIP ENDS:

JENSEN HUANG: Agents are autonomous systems. We have to create extremely secure technology to make sure that we could develop these agents and deploy these agents safely.

Now it’s incredible. After announcing OpenShell, just about every major company in the world from the Frontier Labs to the operating system companies, cybersecurity companies, platform companies have come out to adopt OpenShell. It’s really fantastic to see.

And the reason for that is because this technology has arrived at exactly the right time. At a time when we want to deploy agents all over the world. Now remember, this isn’t just about the services using OpenShell as they deploy agents. This is really about all of the world’s companies also adopting OpenShell so that you could receive these agents and have them run autonomously inside your company.

Every single level of capability, every single task that you give the agent for every person that uses the agent. Every one of those combinations requires a different level of privilege that you would provide to the agent. And every single time an agent is launched, the privileges are different.

And so privileges, the policy enforcement, the monitoring, all has to be done dynamically. And so, OpenShell is really revolutionary.

I’m delighted to say that Anthropic and NVIDIA are partnering to create Claude OpenShell agents. And so you can go to the website. There’s a really great write up about that.

So the first part is how do you use these harnesses, OpenClaw, Hermes, OpenHands. You could create your own harnesses from LangChain, Deep Agents. There’s… Or just write it from Python. There’s a whole bunch of different ways you could create agent harnesses that sits inside now an OpenShell secure harness. Okay? And that’s the sandbox that it sits in the secure sandbox.

Open Models and Nemotron

And then next, let’s talk about models. There… You should use models of all kinds. In the future, it won’t just be one model that we run everything on. We’re going to run… We’re going to use the best models whether they’re closed, open, or custom. We’re going to use all the models to do your work.

Sometimes, the work that you do is fairly generic. Artificial general intelligence is going to do an incredibly good job. However, sometimes, the work that you want to do captures too much of your proprietary knowledge, too much of your specialty, and you want to continuously enhance that intelligence.

Every company in the future will use both closed models as well as open models that you customize. By the end of the session, you will see that in fact, building your own AIs is not that hard.

It starts with, of course, the model. We’ve been dedicated to advancing open models for… To open up new opportunities and enable people to build their own custom AIs.

Nemotron is for languages. Cosmos is our world foundation model. We are the frontier. GR00T for humanoid robotics. We are the frontier. Alpamayo for autonomous navigation. We are the frontier. Ising for quantum classical computing. We are the frontier. And of course, Earth-2, physics AI. We are the frontier.

So in all of these different areas, we offer you all of the models, and we’re continuously working on it. We’ve been working on it for a decade, and we’ll continue to work on this forever.

We offer you the model completely open, the data that goes along with it so that you know the prominence of the model that we created, as well as the scripts necessary for you to reproduce it. So not only do you have the open weights, you have from us the entire recipe and the data to recreate the open models So that you could know where it came from and have complete trust into it.

Well, these models are not necessarily the frontier of frontiers, especially for the language model. However, with customization and with partners that the ecosystem that we work with, you could turn these models into frontier. And as a result, turn your proprietary data into sovereign intelligence.

Here’s an example of AI21 helping Alice turn Nemotron into a model that exceeds frontier capabilities. Here’s Baseten helping CodeRabbit, turning a Nemotron model after fine tuning on your data turned into frontier capability, Baseten and CodeRabbit.

Here’s Trajectory in Harvey. Again, using Nemotron, post trained it and turned and created their own frontier level capability. Here’s Scale working with Skydance. Again, doing the same thing. It goes on and on and on.

The point being that you could start with an open model. Something that’s incredibly good like Nemotron. And with your own data, post train it for your use case and achieve, of course, super intelligence. That is a very sensible thing to do. And all the technology exists today.

Here’s an example of Palantir with Tampa General Hospital. The list goes on and on. We have a rich ecosystem of partners that can help you. And they’re incredible companies. And they can work with you from around the world. Okay?

So we work… It’s just… Anyways, I’ll mention one and I’ll offend everybody else. I’ll have to sit here and mention all of them. And I love them all.

The Agent Knowledge Foundation

And so first, agent harness running in a secure sandbox OpenShell. Second, Nemotron large language models. The first two components of what you need. But ultimately, what turns these super agents, these agents into super agents is that you give them access to your company’s knowledge.

And this is really special. This is called the agent knowledge foundation that we put together. Libraries and models that are completely world class. It allows you to turn your company’s data.

The data could be in… The assets could be in unstructured format and documents, so that you need an AI to understand how to read these PDFs and read all these unstructured documents. Or it could be in structured data inside databases. We have the libraries and the tools necessary to help you extract meaning and understanding from your company’s data. One at a time.

Nemotron Parse, completely world class top of leaderboard. Nemotron Embed, taking a query and associating with the knowledge that’s inside the vector database, top of leaderboard. Nemotron Rerank. It might find many different sources where the query is associated with different pieces of data. It brings it together, ranks them based on which one’s the most relevant to you.

Kumo, to be able to predict from your structured database what is likely to happen. And so this is, for example, if you were to receive a notice that one of your suppliers will be short on supply or delay on supply. That piece of information is likely unstructured data as a document.

The information about the orders in your company are likely inside a structured database. Your SQL database, for example. And in order to predict what’s going to happen to all of these orders, who are going to be short, the amount of delay, you use Kumo to make that prediction.

And then in order to figure out what to replace it with, you might use vector database search and to go find a engineering document of a replacement component. Okay?

So all of those different libraries and data processing capabilities now makes for you your knowledge foundation. You’ve now armed your AI with proprietary data, proprietary knowledge.

CUDA-X Libraries: The Tools of the Trade

Now, the last part that I’ll talk about is ultimately tools. Tools of the trade. NVIDIA has been working on CUDA-X. And for most of you, in the last… Before AI, all of our conversations have been about around how do we accelerate simulators and solvers, data processing engines, domain specific libraries of all kinds.

They’re used in scientific discovery, in all the different fields of science. They’re used in engineering. They’re used in research. They’re used in designing, building creative arts. At this core, our CUDA-X libraries and these CUDA-X libraries are algorithms. And algorithms, as you know, are beautiful. Let’s take a look at it.

Math is beautiful. Math is beautiful.

CUDA-X libraries, these are the crown jewels of our company. And the reason for that is this. When you’re building an AI assistant, it could be a lab assistant. It could be a research assistant. It could be a co designer. As somebody who’s co creator with you, building something with you. Ultimately, what makes those agents useful? What makes them able to do real work are the tools.

And so we now have models that allow these agents to reason, understand information, reason, to be able to plan and iteratively act and use tools until it could get the job done. But it still comes down to the tools. And these tools we’ve been working on for so many years.

And there was one particular computer I showed you on the right hand side that runs every single one of these tools. And it could be fluid dynamics and particle physics. It could be quantum chemistry. It could be data processing. It could be ray tracing. All of these tools run on that one computer.

And that one computer speeds up the processing of it compared to just running it natively on a CPU by 10, 50, a 100 x. And CUDA-X tools is what give the agents the tools of the trade. And it’s so cool to see how other people are using it. Let’s take a look.

So this is the magic behind the agents. Ultimately, the agents have to use tools. Now, of course, in one of the first fields of agentic use is software programming, and therefore, a compiler or a Python is all really it needs. But in the future, as we have these tools do more and more and more and more useful work, we’re going to need to give them more tools. And here’s some examples.

So here’s Lila Sciences. They’re creating a lab assistant, and they’re using NVIDIA’s BioNeMo models that are designed for understanding proteins and understanding cells and generating proteins and chemicals in drug discovery. They’re using BioNeMo and CUDA-X tools.

This is Siemens factory building a factory agent that uses NVIDIA Cosmos, our world foundation model that understands the physical world and CUDA-X tools. And here’s example of Dassault creating a virtual companion that uses NVIDIA’s Nemotron, our language model, and CUDA-X tools. Let’s see them in use.

VIDEO CLIP BEGINS:

UNIDENTIFIED SPEAKER: The most intelligent scientist still needs a laboratory. The best engineer still needs simulators and solvers. An AI agent is no different. An agent brings together five building blocks.

The model perceives reasons and plans. The harness coordinates the work. Knowledge foundation gives the agent context from your company’s structured and unstructured data. Tools let the agent simulate, experiment, and act. NVIDIA OpenShell controls what agents can access and execute.

Start with a frontier closed model or customize an open model. With NVIDIA’s agent toolkit, any company can turn its knowledge and expertise into a customized super agent. Let’s look at three examples.

For Lila Sciences, the goal is to design a therapy for B cell lymphoma. The Lila Iris agent built on NVIDIA Nemotron turns that goal into an experimental plan.

In the computational dry lab, Lila agents use NVIDIA BioNeMo agent toolkit to call models for protein design and RNA optimization, running millions of virtual experiments.

The most promising candidates move into Lila’s AI science factory, where autonomous robots and instruments developed with NVIDIA Omniverse and Cosmos make and test them. The results return to Iris, shaping the next designs and experiments. The loop continues.

At Siemens Electronics Factory Erlangen, the goal is to bring new products into production. Process Preparation X Agents use NVIDIA Cosmos to generate virtual prototypes and synthetic data, then train and validate inspection models before production starts.

Intelligent Center X Agents, built with NVIDIA’s factory operations blueprint, use orders, supplies, and line setups to recommend production plans. Digital Twin Composer uses NVIDIA Omniverse libraries to validate those plans with physics simulation and real time factory data before changes reach the factory floor.

With the Dassault Systèmes platform, an engineer needs to adapt a Volkswagen Golf door design for the Tiguan. Aura and Leo use NVIDIA Nemotron to reason over engineering requirements. The harness coordinates the specialized sub agents and preserves context across planning, design, and simulation.

Aura uses ENOVIA to plan the work around the vehicle’s dimensions, weight limits, and other engineering requirements. LEO Design Agent selects a compatible Golf door design and uses CATIA to reshape it for the Tiguan.

LEO simulation agent uses SIMULIA accelerated by NVIDIA CUDA-X libraries to test the variants, revealing a vibration problem. The results return to the design agent for refinement. The engineer reviews the refined designs and selects the final one.

Different goals. Different expertise. The same building blocks. Open models, CUDA-X accelerated tools, and the OpenShell secure runtime brought together with the NVIDIA agent toolkit.

VIDEO CLIP ENDS:

JENSEN HUANG: Isn’t that great?

So some people said that AI is here is going to be the end of software. It’s going to be end of tools company. It’s exactly the opposite. AI is here. There’ll be more users of these tools that we created over time than ever before.

Every single tools company will now become an AI company. They’ll be an agents company. Not only will they offer tools, but they’ll offer you specially optimized agents that are incredibly good at using their tools. Cadence and CrowdStrike and Dassault, IQVIA, and just about every enterprise company out there are all building agents. And we’re just delighted to be working with them.

Building agents is not that hard. It’s got four… Five components. It’s got the model, the harness, both of them have closed and open versions, but the OpenShell secure runtime, and then most importantly, the knowledge and tools. And those knowledge and tools belong to you. That’s your expertise.

And so all of this is now possible to do. And I’ll show you more in just a second.

Physical AI: Telecom, Radiotherapy and Robotaxis

But before I do that, let’s talk about physical AI. That’s really the next wave of AI. AI that understands the physical world can reason about it because the physical world is diverse. It is complicated. It’s got extremely long tails. And to reason about it and use appropriate tools to interact with the physical world.

Whenever people think about physical worlds, they think robotics. In fact, one of my favorite physical world applications is the radio cell phone cell towers. And the reason for that is this. It uses, instead of cameras, instead of speakers and microphones, it uses radio waves to sense the world. And to… Reasons about it.

It processes information, reasons about it, interacts with the world using beam forming. So in fact, this is an application, a technology that is just perfectly suited for AI. AI will revolutionize the way we do telecommunications. It will enhance our spectrum efficiency tremendously.

Just as we’ve done so with ray tracing. All of the graphics that you’ve seen, completely a 100% ray traced. The amount of computation we thought necessary in the past, and now that we apply AI to, is a million to one difference. And so we’re going to bring to the telecommunications network an enormous step up in spectrum efficiency.

But beyond that, we’re also going to revolutionize all of these cell towers. There’s a 100,000 sites around the world, around the Western world, About a 100 gigawatts of distributed data centers that will in the future become a distributed AI infrastructure.

We’re delighted to partner with many companies here in Europe. And just this last week, after working on this for nearly eight years, NVIDIA made the first commercial phone call on the T-Mobile network with Nokia. And so now, CUDA makes phone calls the traditional way. Then once it can do that, then we can layer AI on top of it. And so the revolution is going to begin. This is really exciting opportunity for us.

The other physical AI application that I’m super excited about is essentially robotic radiotherapy. Your tube… The tumor is changing in shape. The physical body is changing while you’re in therapy.

The robot, this radio… Adaptive radiotherapy system has to perceive the changing environment, understand where the tumor is, segment it just perfectly, and direct… Precisely direct the radiotherapy accordingly. Essentially, robotic… That’s right. Go ahead. This is a form of robotic surgery.

Now, of course, the highest volume, the most near term robotics application or physical AI application that everybody recognized has been something we’ve been working on for over a decade. And now, we recognize that every single car company will want to make their cars part of the ride hailing fleet.

Today, we have ride hailing, and as part of Uber, and here in Europe, we have Bolt, and we have Grab, and Lyft, and so on and so forth, and all these different ride hailing services. In the future, there’ll be robo taxis that will be connected to those ride hailing services as well. But it won’t be from just one brand.

And so we recognize that building these systems are incredibly expensive, particularly the software that goes on top of it. We decided to create one standard architecture. 14 different cameras, nine radar systems, three lidars, umpteen ultrasonics. This cocoon of a shell of sensors powered by dual NVIDIA GPU systems for the AV computing system.

We turned this into an industry standard. And now it’s being designed into one car fleet after another. On top of it, our AV software companies, and we are in fact one of them as well, to build an autonomous robo taxi capability, which then is connected into the ride hailing services. This ecosystem is being built out.

The world travels about a trillion miles per year around the world. Many of those miles in the future, a dollar or 2 per mile, are going to be on robo taxis. So this is one giant opportunity, but it’s incredibly fragmented.

And so creating an industry standard, Hyperion, and working with an entire ecosystem, we could create essentially a giant virtual fleet that everybody can connect into.

Today, I’m super proud to announce that four more brands, four more cars are going to… Lines are going to come in. The VW van, I just can’t imagine anything more iconic. The JLR brand new car. The Volvo fleet, as well as Mercedes next generation are going to be also Hyperion compatible and be part of this network, this ride hailing network, Robotaxi fleet. So welcome to Hyperion.

Humanoid Robotics

The next major industry. The next major industry, the next great industry, and this one is likely to be extremely large, is humanoid robotics. There’s a reason why robotics is hard to do. Because all the different work is rather diverse.

And so building a special configuration for each one of the diverse tasks segments the robotics industry into all of these different configurations, which prevents the technology flywheel to happen.

The reason why the cell phone has advanced so quickly, the reason why personal computers have advanced so quickly, the reason why NVIDIA’s NVLink 72 has advanced so quickly, is because those standardized configurations allows many customers to use it, the volume to therefore be derived. That volume drives the technology flywheel, and it gets faster, better, incredibly fast.

And so, we need a robotics configuration that is as standard as possible. Now, humanoid robotics obviously has the ability. It probably isn’t perfect for everything, but it’s likely to work for many things. As a result, this is likely going to be the largest volume robotics configuration.

Now, if we could just create the AI that allows it to learn very quickly, then for small medium businesses who don’t have an entire robotics team to teach and program the robot can now show the robot what needs to be done. The robot, because it has general capabilities and general intelligence, would be able to learn and adapt quickly. And so this is likely to be the next major platform, technology platform.

We’re working with so many different companies. You will recognize many of them here in Europe. This is your great opportunity. No country, no region has dedicated itself to making mechatronics, electronics, mechanical systems at the level of precision and craft as the European region.

So many companies here are incredibly good at manufacturing. This is going to be the next high volume industry. And you could see there’s so many companies that are now part of it, and we’re delighted to be partnered with you.

Now, in order to create a physical AI system, you need three computers. The first computer is the one we’ve been talking about and everybody knows quite well, which is training the AI model.

The second computer is where we have the AI learn how to do its skills. You can’t just have the AI program the AI and have the AI come out into the physical world and try to do whatever it wants to do. You do it inside a digital twin. We create a digital twin called Omniverse, and you’ve seen us talk about that repeatedly.

And in fact, just now, when you saw all the CUDA-X libraries, they’re all being done inside a digital twin world. That digital twin world is called Omniverse. Omniverse is a library that’s connected into all kinds of platforms and tools. Siemens and Dassault and Cadence and so on and so forth. And so you could see these digital twin systems being created. The second computer.

The third computer is the robotics computer. That computer integrates in the case of the self driving car, the robotaxi, we call it Hyperion. Hyperion consists of the computer itself, the functional safety system, the active safety system, an operating system designed for physical AI safety.

Well, we need to do exactly the same for robotic systems. So that every single time you start to develop research and start to work on your own robotic systems, you don’t have to start from scratch to cobble up all of the software and computing systems and the sensors and all the mechanical systems. That alone would take several years.

And so we decided to build a reference robot. This is the NVIDIA Agile ONE in partnership with Agile Robots. This robot is going… This reference robot will be available in the first quarter of next year, and it will come preassembled with computer inside, software integrated, and it should be able to perform quite well right out of the bat. And so this is a great starting point, and it’ll be the flywheel by which we can accelerate development of robotics.

RTX Spark: Reimagining the Personal Computer

Well, let me start… Let me end today’s speech from where we started. NVIDIA started in the PC revolution. It was in 1993 and the PC revolution was just about to start. Remember, Windows 95 kicked it into turbocharge.

We came since then, invented the GPU, programmable shading, real time ray tracing. We added CUDA to the computing model and helped ignite the modern AI revolution.

And so four years ago, we started to think, what can we do to reimagine the personal computer for this new era? Where the personal computer is not just your tool, your personal computer tool, but your personal agent or your personal assistant. And the way you interact with this computer would be essentially a companion, somebody who helps you do work.

And so this computer needs to have exactly the five things that I mentioned today. It needs to have model. It could be running locally or it could be running in the cloud. It needs a harness. It needs a secure runtime, OpenShell. And it needs access because it’s available on your PC. It needs access to knowledge and tools.

And the good news is because NVIDIA has been accelerating tools for so long, if we could just create a computer that can accelerate OpenGL and DirectX and everything CUDA, everything NVIDIA has ever worked on would run on this computer.

And so this is it. After four years of partnership with Microsoft, we have created the RTX Spark. And this is a brand new line of computer. This every… Literally thousands of years, thousands of engineering years went into this.

We’ve been secretly working on it for four years, created a chip called the N1X. And then we just recently announced it with the entire PC industry joining forces. We created desktops, laptops, and workstations.

This is one petaflops. One petaflops. Just to put that in perspective, in the year 2016, I delivered one petaflops and a 128 gigabytes of memory. One petaflops and a 128 gigabytes to a small company called OpenAI. This is one petaflops and a 128 gigabytes. The only difference, of course, is that I can hold this like this, and DGX-1 took two people to carry upstairs.

And so this is the brand new RTX Spark PC. And if you would like to have a companion workstation that runs CATIA and everything Dassault and everything Autodesk and Adobe, and Cadence, and Synopsys, you can always get yourself a DGX Station that runs Windows. And this is 20 petaflops and 754 gigabytes of memory. And so incredible.

And of course, there’s also the desktop version. All of these are now available. And I think the DGX Station with RTX Spark, with DGX Spark, DGX Station with Windows will be available at the end of this year.

The Microduck Demo

Now the question is, so far I just talked about the five components of an agent. And now the question is, how do you build it? How do you build an agent that runs, let’s say, on one of those computers or runs in the cloud or runs on prem somewhere?

Well, the answer is it’s really easy and you do it this way. The way you build an agent is you ask an agent to build it for you. So this is the prompt that is going to run the next application. And what you’re about to see is this. You’re about to see a robotics application.

And while you’re watching this demo, just remember, there’s an agent running on this personal computer. This demo runs across three computers. The first computer is to train. The second is a digital twin. And the third is to put that AI into the computer itself. Okay?

So two very… Two ideas. The first idea is agents, how do you build it and run it, and what you can use it to do. And the second is physical AI computing model, which consists of three computers. Let’s take a look.

VIDEO CLIP BEGINS:

UNIDENTIFIED SPEAKER: Look. Meet Microduck, Hugging Face’s robot. Let’s teach it to see and walk with NVIDIA RTX Spark.

I use Hermes as my agent harness, Nemotron and OpenAI’s ChatGPT Astra as my models, and CUDA-X accelerated tools FreeCAD, Blender, and Isaac Sim, Newton Physics Engine, and Isaac Lab. I’ll run everything on my RTX Spark, except when I need to burst larger workloads to cloud.

First, I describe the course and what my needs to learn. On RTX Spark, Hermes and Nemotron plan the course, listing every object required. Astra takes it from there, using FreeCAD to model 3D objects, Blender to apply materials, and Omniverse SimReady libraries to add physical properties. In Isaac Sim, Astra composes the scene.

Astra rigs Microduck’s joints, actuators, and links, and simulates how moves with Newton Physics Engine. Now training begins. In the cloud, thousands of Microducks train in parallel in Isaac Lab. Every fall, a lesson. Improving the policy every time.

Omniverse libraries simulate Microduck’s cameras, teaching it to see. With perception and physics working together, my Microduck can avoid obstacles, navigate tight spaces, and adapt to its surroundings.

I download the policy to RTX Spark and use Nemotron and Isaac Sim to evaluate my Microduck with multi robot orchestration. I deploy the validated policy to the real duck, and he’s on his way.

NVIDIA RTX Spark. Build and run AI agents and their accelerated tools right on your laptop.

VIDEO CLIP ENDS:

JENSEN HUANG: Hey, guys. Oh, yeah? Okay. Come on. Here. Come over here. Show people you can walk. And you have a camera. Oh, yeah? Don’t trip.

Can you guys sit down? Okay. Stand up. Take a nap. Take a nap. You need to take a nap. Okay. Stand up.

Did you know that you were created in here? You were created on RTX Spark. I know. And trained you in the cloud on a whole bunch of NVIDIA supercomputers, and you learned your skills right here on RTX Spark in a digital twin, the Omniverse digital twin. Isn’t that incredible? I know. I know. You guys are so cute.

And your parents are called Pollen, P-O-L-L-E-N. Is that right? Uh-huh. Pollen Robotics. And apparently, everybody loves you guys so much. You’re sold out. I know.

So if you go to the Pollen website, everybody watching this, there’s going to be 20 million viewers. Everybody watching this, go to the Pollen website and place an order. I think this startup company is like 10 people. They might blow up. Poof. It’s going to be so exciting. Everybody should have one. You guys are so cute.

And so we’ll just have an RTX Spark, and we’ll teach you new skills. Isn’t that right? Oi oi. Oi oi. Oh, look at that. That’s incredible. Can you do a somersault? That’s so good. That’s so good. Wow. Okay. Well, guess what? Okay.

Closing

So today, what do we talk about? We talked about Europe building its AI economy, and it starts with infrastructure. And we announced a brand new system called DSX POD. It makes it easy for you to build AI factories. And these AI factories don’t just run models anymore, pay attention. Stay focused, come here, come here. Come over here, stay focused.

And these AI factories don’t just run… There you go, sit down. Yeah. They don’t just run AI models anymore. They run entire agents. And these agents are consistent of five things.

There’s the large language model that does the perception, reasoning, planning and acting. And it runs in conjunction with a harness that turns it… It’s essentially exoskeleton. Once that harness sits inside, runs inside a secure runtime we call OpenShell, which has just so many people that have supported it. And you give it access to knowledge and tools.

And this agent runs on top of these computers, but we’re also extending into the next generation, which is where you guys come in, physical AI. So that’s it. That’s what we talked about.

Everybody can build agents. Every company in Europe will come… Will become an AI company. There’s no question everybody will become an AI company, and we’re going to have a thriving AI economy here in Europe.

So after thirty three years, what an incredible journey, and we made a really cool video. You guys want to watch it together? Okay. Come on. Come over here. Just come over here just a little bit. We’ll get out of the way. Come here. Top speed. Top speed.

VIDEO CLIP BEGINS:

UNIDENTIFIED SPEAKER: How the heck are we supposed to follow that? Well, there is one thing that can make this crowd go even crazier.

UNIDENTIFIED SPEAKER: What is it?

[Song recapping the keynote plays]

UNIDENTIFIED SPEAKER: Welcome to GTC.

VIDEO CLIP ENDS:

JENSEN HUANG: I want to thank all of you, all the sponsors. You can see all the names of companies that are here, all the partners we’ve been working with for thirty three years. Microduck, my new friends. Let’s work together to build Europe’s AI economy and put AI to work.

Alright, everybody. Have a great GTC.

Related Posts

• Transcript: Who’s Paying for the AI Boom—and Can It Last? –  WTF is AI? 
• Nikesh Arora: AI Cyber Panic, Ken Griffin’s Bubble Warning on All-In Podcast (Transcript) 
• All-In Interview: Arvind Krishna & Nikesh Arora – Quantum Threat (Transcript)
• Transcript: Tesla (TSLA) Q3 2026 Earnings Call
• Transcript: SpaceX (SPCX) Q3 2026 Earnings Call