NVIDIA CEO Jensen Huang delivers a live keynote at GTC Paris to kick off VivaTech 2025, revealing the next phase of AI computing—from agentic systems to AI factories. Discover the platform powering the industrialization of intelligence on June 11, 11:00 a.m. CEST
this is how intelligence is made a new kind of factory generator of tokens the building blocks of AI tokens have opened a new frontier the first step into an extraordinary world where endless possibilities are born tokens transform images into scientific data charting alien atmospheres and guiding the explorers of tomorrow they probe the earth’s depths to seek out hidden danger they turn potential into plenty [Music] and help us harvest our bounty tokens see disease before it takes hold cure with precision [Music] and learn what makes us tick tokens connect the dots so we can protect our most noble creatures tokens decode the laws of physics to move us faster and make our days more efficient tokens don’t just teach robots how to move but to bring joy and comfort hi Maroka hi Anna are you ready to see the doctor what’s that it’s my enchanted jewel [Music] tokens help us move forward one small step for man becomes one giant leap for mankind so we can boldly go where no one has gone before [Music] and here is where it all begins [Laughter] [Music] welcome to the stage Nvidia founder and CEO Jensen Wong [Music] hello Paris bonjour nvidia’s first GTC in Paris this is incredible thank you for all the partners who are here with us we have so many people that we work with over the years in fact we’ve been in Europe for a very long time even though this is my first GTC Paris I have a lot to tell you nvidia once upon a time wanted to create a new computing platform to do things that normal computers cannot we accelerated the CPU created a new type of computing called accelerated computing and one of our first applications was molecular dynamics we’ve come a long way since so many different libraries and in fact what makes accelerated computing special is it’s not just a new processor that you compile software to you have to reformulate how you do computing you have to reformulate your algorithm and it turns out to be incredibly hard for people to reformulate software and algorithms to be highly paralyzed and so we created libraries to help each market each domain of application become accelerated each one of these libraries opens up new opportunities for the developers and it opens up new opportunities for growth for us and our ecosystem partners computational lithography the probably the single most important applications in semiconductor design today runs in a factory at TSMC Samsung large semiconductor fabs before the chip is made it runs through an inverse physics algorithm called coup litho computational lithography direct sparse solvers algebraic multi-grid solvers coop we just open sourced incredibly exciting application library this library accelerates decision making to optimize problems with millions of variables for millions of constraints like traveling salespeople problems warp a pythonic framework for expressing geometry and physics solvers really important QDF QML structure databases data frames classical machine learning algorithms qdf accelerates Spark zero lines of code change qml accelerates Scikiticle Learn zero lines of code change dynamo and QDNN qdnn is probably the single most important library Nvidia has ever created it accelerates the primitives of deep neural networks and Dynamo is our brand new library that makes it possible to dispatch orchestrate distribute extremely complex inference workloads across an entire AI factory c equivariance and coupensor tensor contraction algorithms equal variance is for neuronet networks that obey the laws of geometry such as proteins molecules aerial and shiona really important framework to enable AI to run 6G Earth 2 our simulation environment for foundation models of weather and climate models kilometer square incredibly high resolution monai our framework for medical imaging incredibly popular parabicks our solver for genomics analysis incredibly successful coup I’ll talk about in just a second for quantum computing and coupler for numpy and scypi as you could see these are just a few of the examples of libraries there are 400 others each one of them accelerates a domain of application each one of them opens up new opportunities well one of the most exciting most exciting is CUDA Q cuda X is this suite of libraries a library suite for accelerating applications and algorithms on top of CUDA we now have CUDA Q cuda Q is for quantum computing for classical quantum quantum classical computing based on GPUs we’ve been working on CUDAQ now for several years and today I can tell you there’s an inflection point happening in quantum computing as you know the first physical cubit was demonstrated some nearly 30 years ago an error correction algorithm was invented in 1995 and in 2023 almost 30 years later the world’s first logical cubit was demonstrated by Google now since then a couple of years later the number of logical cubits which is represented by a whole lot of physical cubits with error correction the number of logical cubits are starting to grow well just like Moore’s law would I could I could totally expect 10 times more logical cubits every 5 years a 100 times more logical cubits every 10 years those logical cubits would become better error corrected more robust higher performance more resilient and of course will continue to be scalable quantum computing is reaching an inflection point we’ve been working with quantum computing companies all over the world in several different ways but here in Europe there’s a large community uh I saw Pascal last night i saw Barcelona supercomputing last night it is clear now we are within reach of being able to apply quantum computing quantum classical computing in areas that can solve some interesting problems in the coming years this is a really exciting time and so we’ve been working with all of the supercomputing centers it’s very clear now that over the next several years or at least the next generation of supercomputers every single one of them will have a QPU assigned and QPU connected to GPUs the QPU will do quantum computing of course and the GPUs would be used for pre-processing for control for error correction which would be intensely computationally intensive post-processing and such between the two architectures just as we accelerated the CPU now there’s QPU working with the GPU to enable the next generation of computing well we’ve today we’re announcing that our entire quantum algorithm stack is now accelerated on Grace Blackwell 200 and the speed up is utterly incredible we work with the comput quantum computing industry in several different ways one way is using coup quantum to simulate the cubits or simulate the algorithms that runs on top of uh these quantum computers essentially using a classical computer to simulate or emulate a quantum computer at the other extreme extremely important is CUDA Q basically inventing a new CUDA that extends CUDA into quantum classical so that applications that are developed on CUDA Q can run before the quantum computer arrives in an emulated way or after the quantum computer arrives in a collaborative way a quantum classical accelerated computing approach and so today we’re announcing CUDAQ is available for Grace Blackwell the ecosystem here is incredibly rich and of course Europe is deep with science and deep with supercomputing expertise and deep with heritage in this area and it’s not surprising to see quantum computing advance here in the next several years we’re going to see a really fantastic inflection point so anyways for all of the quantum computer industry that have been working on this for three decades now I congratulate you for just the incredible accomplishment and the milestones today thank you [Applause] let’s talk about AI i you might be surprised that I would have I would be talking to you about AI the same the same GPU that ran and enabled all of these applications that I mentioned that same GPU enabled artificial intelligence to come to the world our first contact was in 2012 just prior to that working with developers on a new type of algorithm called deep learning it enabled the AlexNet big bang of AI 2012 in the last 15 years or so AI has progressed incredibly fast the first wave of AI was perception for computers to recognize information understand it the second wave which most of us were talking about the five the last five years or so was generative AI it’s multimodal meaning that an AI was able to learn both images and language therefore you could prompt it with language and it could generate images the ability for AI to be multimodal as well as able to translate and generate content enabled generative AI revolution generative AI the ability to generate content is fundamentally vital for us to be productive well we’ve got a new we are starting a new wave of AI and this last couple of years we’ve seen enormous progress in AI’s ability fundamentally intelligence is about understanding perception reasoning planning a task how to solve a problem and then executing the task perception reasoning planning the fundamental cycles of intelligence it allows us to apply some previously learned rules to solve problems we’ve never seen before that’s why intelligent people are considered intelligent to be able to take a complicated problem break it down step by step reason about how to solve the problem maybe do research maybe go learn some new information get some help use tools and solve problems step by step well the words that I just described are fundamentally possible today with what is called agentic AI and I’ll show you more in just a second in the physical implementation of that the embodiment of that agentic AI and the motion now the generative capability is generating motion instead of generating videos and generating images or generating text this AI generates local motion the ability to walk or reach out and grab something use tools the ability for AI to be embodied in the physical form is basically robotics these capabilities the fundamental technology to enable agents which are basically information robots and embodied AI physical robots these two fundamental capabilities are now upon us really really exciting times for AI but it all started it all started with GeForce and GeForce brought computer graphics this is the first accelerated computing application we had ever worked on and it’s incredible how far computer graphics has come geforce brought CUDA to the world which enabled m machine learning researchers and AI researchers to advance deep learning then deep learning revolutionized computer graphics and made it possible for us to bring computer graphics to a whole new level everything I’m going to show you today everything I’m going to show you today I’m going to give you a preview of what I’m going to show you but everything I’m going to show you today is computer simulation not animation it’s photon simulation physics simulation particle simulations everything is fundamentally simulation not animation not art it just looks incredibly beautiful because it turns out the world is beautiful and it turns out math is beautiful so let’s take a look [Music] [Music] heat heat [Music] [Music] oh I’ve got a cramp coming on oh I can feel it [Music] heat heat [Music] heat [Music] up here [Music] what do you think numbers in action numbers in action that’s essentially what simulations are and it’s just incredibly beautiful to look at but because of the scale and the speed by which we can now simulate almost everything we can turn everything into a digital twin and because everything can be a digital twin it could be designed planned optimized and operated completely digitally before we put it into the physical world the idea that we would build everything in software is now upon us everything physical will be built digitally everything that’s built magnificently will be built digitally everything that’s operated at gigantic scale will be first built digitally and there will be digital twins that operate it and so today we’re going to talk a lot about digital twins well what started out as a GeForce graphics card anybody in here know what a GeForce is okay all right well what started out as GeForce looks like this now this is the new GeForce it is two tons two and a half tons 1.2 million parts about $3 million 120 kilowatts manufactured in 150 factories 200 technology partners working with us to do this probably something along the lines of $40 billion in R&D budget in order to create what is GB200 and now moving to GB300 it is completely in production and this the machine was designed to be a thinking machine a thinking machine in the sense that it reasons it plans it spends a lot of time talking to itself just like you do we spend most of our time generating words for our own mind generating images for our own mind before we produce it and so the thinking machine is really architecturally what Grace Blackwell was designed to do it was designed to be one giant GPU i compared it to GeForce for a good reason geforce is one GPU so is GB200 it is one giant virtual GPU now we had to disagregate it into a whole bunch of components create a a bunch of new networking technology and sides technology incredibly low low power high energy efficiency interconnects to connect all of these chips and systems together into one virtual GPU this is the Hopper version this is the world famous Hopper system this eight GPUs connected together on MVLink what’s not shown here is a CPU tray a CPU tray with dual CPU and system memory that sits on top together this represents one node of an AI supercomput about half a million dollars this is the Hopper system this is the the system that really put us on the face of face of the map of of uh AI and uh it was an it was under allocation for a very long time uh because the the market took off so quickly but this is the famous hopper system well this entire system including the CPU is replaced by this great blackwell node this is one compute tray right here will replace that entire system it is fully liquid cooled and the CPUs are integrated directly connected to the GPUs so you could see it here two CPUs four GPUs is more performant than that entire system but what’s amazing is this we wanted to connect a whole bunch of these systems together how would you connect all of these together was really hard for us to imagine so we disagregated it what we did was we took that entire motherboard we disagregated into this and this this is the revolutionary MVLink system scaling out computing is not that hard just connect more CPUs with Ethernet scaling out is not hard scaling up is incredibly hard you can only build as large of a computer as you can build the amount of technology and electronics that you could fit into one memory memory model is incredibly hard to do and so what we decided to do was we created a new interconnect called MVLink mvlink is a memory semantics interconnect it’s a compute fabric not a network it directly connects to the CPUs of all of these different MVLink systems compute nodes this is the switch nine of these nine of these stands on top nine of it stands sits on the bottom in the middle are the MVLink switches and what connects it together is this miracle this is the MVLink spine this is 100% copper copper coax it directly connects all of the MVLink chips to all of the GPUs directly connected over this entire spine so that every single 144 Blackwell dies or in 72 different packages are talking to each other at the same time on without blocking all across this MVLink spine the bandwidth of this is about 130 terabytes per second 132 I I know [Applause] no wait wait for it wait for it 130 terabytes per second if it’s in bits 130 terabytes per second it is more than the data rate of the peak traffic of the world’s entire internet traffic on this back plane and yeah so this is how this is how you shrink the internet into 60 pounds mvlink and so we did all that we did all that because the the way a computer is is considered the way you think about computers is going to be fundamentally different in the future i’ll spend more time on this but it was designed to give Blackwell a giant leap above Hopper remember Moore’s law semiconductor physics is only giving you about two times more performance every three to five years how could we achieve 30 40 times more performance in just one generation and we need a 30 40 times more performance because the reasoning models are talking to themselves instead of one shot chat GPT it’s now a reasoning model and it generates a ton more tokens when you’re thinking to yourself you’re breaking the problem down step by step you’re reasoning you’re trying a whole bunch of different paths maybe it’s chain of thoughts maybe it’s tree of thoughts best of end it’s reflecting on its own answers you probably seen these research models reflecting on the answers saying is this a good answer can you do better than that and they oh yeah I can do better than that goes back and thinks some more and so those thinking models reasoning models achieved c incredible performance but it requires a lot more computational capability and what net result MVLink 72 Blackwell’s architecture resulted in a giant leap in performance the way to read this is the X- axis is how fast it’s thinking the Y axis is how much the factory can output supporting a whole bunch of users at one time and so you want the throughput of the factory to be as high as possible so you could support as many people as possible so that the revenues of your factory is as high as possible you want this axis to be as large as possible because the AI is smart smarter here than it is here the more the faster it can think the more it can think before it answers your answer and so this has to do with the ASP of the the average selling price of the tokens and this has to be to do with the throughput of the factories these two combined in that corner is the revenues of the factory this factory based on Blackwell can generate a ton more revenues as a result of the architecture it is such an incredible thing what we built we made a movie for you just to give you a sense of the enormity of the engineering that went into building Grace Blackwell take a look [Music] blackwell is an engineering marvel it begins as a blank silicon wafer [Music] hundreds of chip processing and ultraviolet lithography steps build up each of the 200 billion transistors layer by layer on a 12in wafer the wafer is scribed into individual Blackwell dye tested and sorted separating the good dyes to move forward the chip on wafer on substrate process attaches 32 Blackwell dyes and 128 HPM stacks on a custom silicon interposer wafer metal interconnect traces are etched directly into it connecting Blackwell GPUs and HBM stacks into each system and package unit locking everything into place then the assembly is baked molded and cured creating the Blackwell B200 Super Chip each Blackwell is stress tested in ovens at 125° and pushed to its limits for several hours robots work around the clock to pick and place over 10,000 components onto the Grace Blackwell PCB meanwhile custom liquid cooling copper blocks are prepared to keep the chips at optimal temperatures at another facility Connect X7 Super Nix are built to enable scale out communications and Bluefield 3 DPUs to offload and accelerate networking storage and security tasks all these parts converge to be carefully integrated into GB200 compute trays [Music] mvlink is the breakthrough high-speed link that Nvidia invented to connect multiple GPUs and scale up into a massive virtual GPU the MVLink switch tray is constructed with MVLink switch chips providing 14.4 tab per second of all toall bandwidth mvlink spines form a custom blindmated back plane with 5,000 copper cables connecting all 72 black wells or 144 GPU dies into one giant GPU delivering 130 tab per second of all to all bandwidth more than the global internet’s peak traffic from around the world parts arrive to be assembled by skilled technicians into a rack scale AI supercomput [Music] in total 1.2 2 million components 2 miles of copper cable 130 trillion transistors weighing nearly 2 tons [Music] blackwell is more than a technological wonder it’s a testament to the power of global collaboration and innovation fueling the discoveries and solutions that will shape our future everywhere we are driven to enable the geniuses of our time to do their life’s work and we can’t wait to see the breakthroughs you deliver grace Blackwell Systems all in production it is really a miracle it’s a miracle from a technology perspective but the supply chain that came together to build these GB200 systems two tons each we’re producing them now a thousand systems a week no one has ever produced mass-produced supercomputers at this scale before each one of these racks is essentially an entire supercomput only in 2018 the largest Volta system the Sierra supercomputer in 2018 is less performant than one of these racks and that system was 10 megawws this is 100 kilowatts so the difference generationally between 2018 and now we’ve really taken supercomputing AI supercomputing to a whole new level and we’re now producing these machinery at enormous scales and this is just the beginning in fact what you’ve seen is just one system Grace Blackwell the entire world is talking about this one system clamoring for it to get deployed into into the world’s data centers for training and inferencing and generative AI however not everybody and not every data center can handle these liquid cool systems some data centers required enterprise stacks the ability to run Linux Red Hat or Newonix or VMware storage systems from Dell EMC Hitachi NetApp Vast Weta so many different storage systems so many different IT systems and the management of those has to be done in the way that’s consistent with traditional IT systems we have so many new computers to ramp into production and I’m so happy to tell you that every single one of these are now in production you haven’t seen them yet they’re all flying off the shelves flying off the ramps the manufacturing lines starting here dgx Spark enables you to have essentially the Grace Blackwell system on your desktop in the case of Spark Desktop in the case of DJX station desk side this way you don’t have to sit on a supercomput while you’re developing your software while you’re developing your AI but you want the architecture to be exactly the same these systems are identical from an architecture perspective from a software developer perspective it looks exactly the same the only difference is scale and speed and then on this side are all the x86 systems the world’s IT organization still prefers x86 and appreciates x86 wherever they can take advantage of the most advanced AI native systems they do where they can’t and they want to integrate into the enterprise IT systems we now offer them the ability to do so one of the most important systems and it’s taken us the longest to build because the software and the architecture is so complicated is how to bring the AI native architecture and infuse it into the traditional enterprise IT system this is our brand new RTX Pro server this is an incredible system the motherboard is completely redesigned ladies and gentlemen Janine Paul [Applause] This motherboard looks so simple and yet on top of this motherboard are eight Super Nix switches that connect eight GPUs across a 200 Gbits per second state-of-the-art networking chip that then connects eight of these GPUs and these Blackwell RTX Pro 6000 GPUs brand new just entered into production eight of these go into a server now what makes it special this server is the only server in the world that runs everything the world has ever written and everything Nvidia has ever developed it runs AI Omniverse RTX for video games it runs Windows it runs Linux it runs Kubernetes it runs Kubernetes in VMware it runs basically everything if you want to stream Windows desktop from a computer to your to your remote device no problem if you want to stream Omniverse no problem if you want to run your robotic stack no problem just a QA of this particular machine is insane the applications that it runs basically universal everything the world’s ever developed should run on here including if you’re a video gamer including Crisis and so if if you can run Crisis you can run anything okay this is the RTX Pro server brand new enterprise system so something is changing we know that AI is incredibly important technology we know for a fact now that AI is software that could revolutionize transform every industry it can do these amazing things that we know we also know that the way you process AI is fundamentally different than the way we used to process software written by hand machine learning software is developed differently and it runs differently the architecture of the systems the architecture of the software completely different the way the networking works completely different the way it acts as storage completely different so we know that the technology can do different things incredible things it’s intelligent we also know that it’s developed in a fundamentally different way needs new computers the thing that’s really interesting is what does this all mean to countries to companies to society and this is this is an observation that we made almost a decade ago that now everyone is awakening to that in fact these AI data centers are not data centers at all they’re not data centers in the classical sense of a data center storing your files that you retrieve these data centers are not storing our files it has one job and one job only to produce intelligent tokens the generation of AI these factories of AI are look like data centers in the sense that they have a lot of computers inside but that’s where everything breaks down how it’s designed the scale at which it’s manufactured or scaled designed and built and how it’s used and how it’s orchestrated and provisioned operated how you think about it for example nobody really thinks about their data center as a revenue generating facility i said something that everybody goes “Yeah I think you’re right nobody ever thinks about a data center as a revenue generating facility but they think of their factories their car factories as revenue generating facilities and they can’t wait to build another factory because whenever you build a factory revenue grows shortly after you could build more things for more people those ideas are exactly the same ideas in these AI factories they are revenue generating facilities and they are designed to manufacture tokens and these tokens can be reformulated into productive intelligence for so many industries that AI factories are now part of a country’s infrastructure which is the reason why you see me running around the world talking to heads of states because they all want to have AI factories they all want AI to be part of their infrastructure they want AI to be a growth manufacturing industry for them and this is genuinely profound and I think we’re talking about as a result of all that a new industrial revolution because every single industry is affected and a new industry just as electricity became a new industry at first when it was described as a technology and demonstrated as a technology it was understood as a technology but then we understood that it’s also a large industry then there’s the information industry which we now know as the internet and both of them because it affected so many industries became part of infrastructure we now have a new industry an AI industry and it’s now part of the new infrastructure called intelligence infrastructure every country every society every c company will depend on it and you could see its scale this is one that’s being talked about a lot this is Stargate this doesn’t look like a data center it looks like a factory this is 1 gawatt it will hold about 500,000 GPU dies and produce an enormous amount of intelligence that could be used by everybody well Europe has now awakened to the importance of these AI factories the importance of the AI infrastructure and I’m so delighted to see so much activity here this is um European Telos AI infrastructure with Nvidia this is the European cloud service providers building AI infrastructure with NVIDIA and this is the European supercomputing centers building next generation AI supercomputers and infrastructure with NVIDIA and this is just the beginning this is in addition to what will come in the public clouds this is in addition to the public clouds so indigenous built AI infrastructure here in Europe by European companies for the European market and then there’s 20 more being planned 20 more AI factories and several that are gigawatt gigafactories in total in just two years we will increase the amount of AI computing capacity in Europe by a factor of 10 and so the researchers the startups your AI shortage your GPU shortage will be resolved for you soon it’s coming for you now we’re partnering with each country to develop their ecosystem and so we’re building AI technology centers in seven seven different countries and the goal of these AI technology centers is one to do collaborative research to work with the startups and also to build the ecosystem let me show you what an ecosystem looks like in the UK i was just there yesterday the ecosystems are built on top of the NVIDIA stack so for example every single NVIDIA as you know NVIDIA is the only AI architecture that’s available in every cloud it’s the only computing architecture aside from x86 that’s available everywhere we’re avail partner with every cloud service provider we accelerate applications from the most important software developers in the world seammens here in Europe Cadence Red Hat Service Now we’ve reinvented the computing stack as you know computing is not just a computer but it’s compute networking and storage each one of those layers each one of those stacks has been reinvented great partnership with Cisco who announced a brand new model yesterday at their conference based on Nvidia Dell great partnerships NetApp Newtonics whole bunch of great partnerships as I mentioned earlier the way you develop software has been fundamentally changed it’s no longer just write C program compile C program de deliver C program it’s now DevOps MLOps AI ops so that entire ecosystem is being reinvented and we have ecosystem partners everywhere and then of course solution integrators and providers who could then help every company integrate these capabilities well here in the UK we have special companies that we work with really terrific companies from researchers to developers to partners to help us upskill the local economy and upskill the local talent enterprises that consume the technology and of course cloud service providers we have great partners in the UK we have great partners in Germany incredible incredible partnerships in Germany we have great partnerships in Italy and we of course have amazing partnerships here in France [Applause] that’s right go France [Applause] president Mcron’s going to be here later on we’re going to talk about some some new some new announcements so we have to show some enthusiasm for AI okay yeah there you go show him some enthusiasm so great partnerships here here here in France uh one particular one I want to highlight our partnership with Schneider building the even building these AI factories we build them digitally now we design them digitally we build them digitally we operate them or optimize them digitally and we will even eventually optimize them and operate them completely digitally in a digital twin these AI factories are so expensive $50 billion sometimes hundred billion dollars in the future if the utilization of that factory is not at its fullest the cost to the factory owner is going to be incredible and so we need to digitalize and use AI wherever we can put everything into Omniverse so that we have direct and constant telemetry we have a great partnership here that we’re announcing today a young company a CEO I really like and he’s trying to build a European AI company the name of the company is Mistral today today we’re announcing that we’re going to build an AI cloud together here to deliver their models as well as deliver AI applications for the ecosystem of a other AI startups so that they can use the MRO models or any model that they like and so Mrol and we’re going to be partnering to build a very sizable AI cloud here and we’ll tell we’ll talk about more of it later on today with President Mcronone ai technology is moving at light speed and what I’m showing you here proprietary models on the left moving at light speed however the open models are also moving at light speed only a few months behind whether it’s Mistral Llama Deep Seek R1 R2 coming Q1 these models are all exceptional every single one of them exceptional and so we’ve dedicated ourselves over the last several years to apply some of the world’s best AI researchers to make those AI models even better and we call that Neotron basically what we do is we take the models that are open sourced and of course they’re all built on NVIDIA anyhow and so we take those models open sourced we then post train it we might do neural architecture search we might do neuroarchchitecture search provide it with even better data use reinforcement learning techniques enhance those models give it reasoning capabilities extend the context so that it could learn and read more before it interacts with you most of these models have relatively short context and we want it to have enormous context capability because we want to use it in enterprise applications where the conversation we want to have with it is not available on the internet it’s available in our company and so we have to load it up with enormous amount of context all of that capability is then packaged together into a downloadable NIM you could come to Nvidia’s website and literally download an API a state-of-the-art AI model put it anywhere you like and we improve it tremendously this is an example of Neimotron improvement over llama so Llama 8B 70B 405B improved by our post-training capability extension of reason reasoning capability all the data that we provide enhanced it tremendously we’re going to do this generation after generation after generation and so for all of you who would uses Neotron you will know that there’s a whole slew of other models in the future and they’re open anyway so if you would like to start from the open model that’s terrific if you’d like to start with the Neotron model that’s terrific and the Neotron models the performance is excellent in benchmarks after benchmarks after benchmarks neotron performance has top of the leaderboard all over the place and so now you know that you have access to a enhanced open model that is still open that is top of the leader chart and you know that Nvidia is dedicated to this and so I will do this for as long as I shall live okay this strategy is so good this strategy is so good that the regional model makers the model builders across Europe have now recognized how wonderful the strategy is and we’re partnering together to adapt enhance each one of those models for regional languages your data belongs to you your data belongs to you it is the history of your people the knowledge of your people the culture of your people it belongs to you and for many companies in the case of Nvidia our data is largely inside 33 years of data i was looking up this morning Seammens 180 years of data some of it written down on papyrus roland Bush is here i I thought I’d pick on Roland Bush my good friend and so you’ll have to digitize that before the AI can learn and so you the data belongs to you you should use that data use an open model like Limotron and all the tool suites that we provide so that you can enhance it for your own use we’re also announcing that we have a great partnership with Perplexity perplexity is a reasoning search engine yep the three models I use I use Chat GPT Gemini Pro and Perplexity and these three models I use interchangeably and Perplexity is fantastic we’re announcing today that Perplexity will take these regional models and connect it right into Perplexity so that you could now ask and get questions in the language in the culture in the sensibility of your country okay so perplexity regional models agentic AI agentic AI agents is a very big deal as you know in the beginning with pre-trained models people said but it hallucinates it makes things up you’re absolutely right it doesn’t have access to the latest news and data information Absolutely right it gives up without reasoning through problems it’s as if every single answer has to be memorized from the past you’re absolutely right all of those things you know why is it trying to figure out how to add or count the count numbers and add numbers why doesn’t it use a calculator you’re absolutely right and so all of those capabilities associated with intelligence everybody was able to criticize but they was absolutely right because everybody largely understand how intelligence works but those technologies were being built all around the world and they were all coming together from retrieval augmented generation to web search to multimodal understanding so that you can read PDFs go to a website look at the images and the words listen to the videos watch the videos and then take all of that understanding into your context you could also now understand of course a prompt from almost anything you could even say I’m going to ask you a question but start from this image i could say start from this start from this text before you answer the question or do what I ask you to do it then goes off and reasons and plans and evaluates itself all of those capabilities are now integrated and you can see it coming out into the marketplace all over the place agentic AI is real agentic AI is a giant step function from oneshot AI the oneshot AI was necessary to lay the foundation so that we can teach the agents how to be agents you need some basic understanding of knowledge and basic understanding of reasoning to even be able to be teachable and so pre-training is about teachability of AI post-training reinforcement learning supervised learning human demonstration context provision generative AI all of that is coming together to formulate what is now a gentic AI let’s take a look at one example let me show you something it’s built on perplexity and it’s super cool ai agents are digital assistants based on a prompt they reason through and break down problems into multi-step plans they use the proper tools work with other agents and use context from memory to properly execute the job on NVIDIA accelerated systems it starts with a simple prompt let’s ask Perplexity to help start a food truck in Paris first the Perplexity agent reasons through the prompt and forms a plan then calls other agents to help tackle each step using many tools the market researcher reads reviews and reports to uncover trends and analyze the competitive market based on this research a concept designer explores local ingredients and proposes a menu complete with prep time estimates and researches pallets and generates a brand identity then the financial planner uses Monte Carlos simulations to forecast profitability and growth trajectory an operations planner builds a launch timeline with every detail from buying equipment to acquiring the right permits the marketing specialist builds a launch plan with a social media campaign and even codes an interactive website including a map menu and online ordering each agent’s work comes together in a final package proposal and it all started from a single prompt [Applause] one prompt one prompt like that on the in the original chatbot would have generated a few hundred tokens but now with that one single prompt into an agent to solve a problem it must have generated 10,000 times more tokens this is the reason why Grace Blackwell is necessary this is the reason why we need performance and the systems to be so much more performant generationally well this is how Perplexity builds their agents every company will have to build their own agents it’s terrific you’re going to be hiring agents from OpenAI and Gemini and Microsoft Copilot and Perplexity and Mistrol there’ll be agents that are built for you and they might help you plan a vacation or you know go do some research so on so forth however if you want to build a company you’re going to need specialized agents on specialized tools and using specialized tools and specialized skills and so the question is how do you build those agents and so we created a platform for you we created a framework and a set of tools that you can use and a whole bunch of partners to help you do it it starts with on the very bottom on the very bottom the ability to have reasoning models that I spoke about nvidia’s Nemo Neotron reasoning large language models are worldclass we have Nemo Retriever which is a multimodal search engine semantic search engine incredible and we built a blueprint a demonstration that is operational that is essentially a general agent we call it IQ AI AIQ and on top we have a suite of tools that allows you to onboard an agent a general agent curate data to teach it evaluate it guard rail it supervise train it use reinforcement learning all the way to deployment keep it secure keep it safe that suite of toolkits is integrated those libraries are integrated into the AI ops ecosystem you can come and download it from our website yourself as well but it’s largely integrated into AI ops ecosystem from that you could create your own special agents many companies are doing this this is Cisco they announced it yesterday we’re building AI platforms together for security now look at this ai agents is not one model that does all of these amazing things it’s a collection a system of models it’s a system of AI large language models some of them are optimized for certain type of things retrieval as I mentioned performing skills using a computer you don’t want to bundle all of that stuff up into one giant you know mass of AI but you break it up into small things that you could then deploy CI/CD over time this is an example of Cisco’s now the question is how do you now deploy this because as I mentioned earlier there are public clouds where Nvidia’s compute is there are regional clouds we call them NCPs here for example mistrol you might have something that is private cloud because of your security requirements and your data data privacy requirements you might even decide that you have something on your desk and so the question is how do you run all of these and sometimes they will run in different places because these are all microservices these are AIs that could talk to each other they could obviously talk to each other over networking and so how do you deploy all of these microservices well we now have a great system i’m so happy to announce this for you this is called our DGX Leptton dgx Leptton what you’re looking at here is a whole bunch of different clouds here’s the Lambda cloud the AWS cloud you know uh here’s your own developers machine your own system could be a DGX station NBS Yoda and Scale it could be AWS it could be GCP nvidia’s architecture is everywhere and so you could decide where you would like to run your models you deploy it using one super cloud so it’s a cloud of clouds once you get it to work once you get this NIMS deployed into Lepton it will go and be hosted and run on the various clouds that you decide one model architecture one deployment and you can run it everywhere you can even run it on this little tiny machine here you know this this DGX Spark it’s it’s Is this is it a cafe time look at this [Applause] this lift is 200 horsepower this is my favorite little machine DGX Spark the first the AI supercomput we built an AI supercomput in 2016 it’s called the DGX1 it was the first version of everything that I’ve been talking about eight Volta GPUs connected with MVLink it took us billions of dollars to build and on the day we announced it DGX1 there were no customers no interest no applause 100% confusion why would somebody build a computer like that does it run Windows nope and so we built it anyways well thank thankfully uh a young company a startup a nonprofit startup in San Francisco was so delighted to see the computer they said “Can we have one?” And I thought “Oh my gosh we sold one.” But then I discovered it was a nonprofit but I put a computer put a DGX1 in my car and I drove it up to San Francisco and the name of that company is Open AI i don’t know what the life lesson is there there are a lot of nonprofits you know so next time next time but the maybe the lesson is this if a developer reaches out to you need a need a GPU the answer is yes and so so that’s right so imagine you have leptton it’s in it’s in your browser and you have you have uh this this uh Helm chart an AI agent that you’ve developed and you want to run it here and parts of it you want to run in AWS and parts of it you want to run you know in a regional cloud somewhere you use leptton you deploy your Helmchart and it magically shows up here okay and so if you would like to run it here until you’re done with and ready to deploy it then deploy it into the cloud terrific but the beautiful thing is this architecture is based on Grace Blackwell Well GB10 versus GB200 versus GB2 300 and all of these different versions of but this architecture is exactly Grace Blackwell now this is amazing so we’re doing this for Lepon but next hugging face and Nvidia has connected Leptton together and so whenever you’re training a model on hugging face if you would like to deploy it into leptton and directly into Spark no problem it’s just one click so whether you’re training or inferencing we’re now connected to hugging face and leptton will help you decide where you want to deploy it let’s take a look at it developers need easy and reliable access to compute that keeps up with their work wherever they are whatever they’re building djx Cloud Leptin provides ondemand access to a global network of GPUs across clouds regions and partners like Yoda and Nebus multicloud GPU clusters are managed through a single unified interface provisioning is fast developers can scale up the number of nodes quickly without complex setups and start training right away with pre-integrated tools and training ready infrastructure progress is monitored in real time gpu performance convergence and throughput are at your fingertips you can test your fine-tuned models right within the console djx Cloud Leptin can deploy NIM endpoints or your models in multiple clouds or regions for fast distributed inference just like ride sharing apps connect riders to drivers DJX Cloud Leptin connects developers to GPU compute powering a virtual global AI factory dgx cloud leton okay so that’s Cisco this is the way SAP they’re building an AI platform in Nvidia sa is building an AI business application automation on NVIDIA deepl is building their uh language framework and platform on NVIDIA AI photo Room a video editing uh and AI editing uh platform building their platform on NVIDIA and this is Kodo used to be I think Kodium incredible coding agent built on Nvidia and this is Iola a voice platform built on Nvidia and this one is a uh a clinical trial platform the world’s largest uh automation platform for clinical trials built on Nvidia and so all of these all of these basically builds on the same idea nims that encapsulates and packages up in a virtual container that you could deploy anywhere the Neotron large language model or other large language models like Mistral or others uh we then integrate libraries that basically covers the entire life cycle of an AI an AI agent the way you treat an AI agents a little bit like a digital employee so your IT department would have to onboard them fine-tune them train them evaluate them keep them guard railed you know keep them secure and continuously improve them and that entire framework platform is called Nemo and all of that is now being integrated into one application framework after another all over the world this is just an example of a few of them and then now we make it possible for you to deploy them anywhere if you want to deploy it in the cloud you got DGx uh you got GB200’s in the cloud if you want to deploy it on prem because you’ve got uh uh u VMware or Red Hat Linux or uh Newonix and you want to deploy it in your virtual machines on prem you can do that if you want to deploy it as a private cloud you could do that you can deploy it all the way on your DGX Spark or DGX station no problem and so Lepton will help you do all of that let’s talk about industrial AI this is one of my favorite moments this is Roland Bush he just This is a really fun moment he wanted to remind me that neurocomputers neuronet network computers were invented in Europe that’s this whole slide look I just it is it was such a great moment this is the Synapse One this is incredible you guys synapse one this is Synapse One 1992 it runs neural networks 8,000 times faster than CPUs of that time isn’t it incredible so this is the world’s AI computer and and Roland just wants to just makes forget that Jensen never ever forget that i said “Okay all right good all right I’ll tell and I’ll even tell everybody Seammen’s 1992 seammen’s 1992 we have a great partnership with Seaman Seammens and and uh Roland Bush uh the CEO is uh supercharging the company so that they could leap completely leap the last IT industrial revolution and fuse the industrial capabilities of Europe the industrial capabilities of might of Seammens with artificial intelligence and create what is called the industrial AI revolution we’re partnering with Seammens on so many different fronts uh everything from design to simulation to digital twins of factories to operations of the AIs in the factories everything from end to end and it just reminds us it reminds me uh how how incredible the industrial capabilities of Europe is and what an extraordinary opportunity this is for you what an extraordinary opportunity because AI is unlike software ai is really really smart software and this smart software can finally do something that can revolutionize the very industries that you serve and so we made made a a love letter video if you will let’s play it it began here the first industrial revolution Watt’s steam engine and the mechanized loom introduced automation and the advent of factories and industry was born the age of electricity ampier unraveled electromagnetism [Music] faraday built the first electric generator and Maxwell laid the foundations for modern electrical engineering seamans and Wheatstone dynamo the engine of electricity bringing machines trains factories and cities to life electrifying the planet igniting modern manufacturing and today born out of the computing and information age the fourth industrial revolution the age of AI reimagining every part of industry across the continent industrial AI is taking hold from design to engineering you’re blazing new trails toward understanding and reinvention you brought the physical world into the virtual to plan and optimize the world’s modern factories you’re building the next frontier where everything that moves is robotic every car an intelligent autonomous agent and a new collaborative workforce to help close the global labor shortage gap developers across the continent are building every type of robot teaching them new skills in digital twin worlds and robot shows preparing them to work alongside us in our factories warehouses the operating room and at home the fourth industrial revolution is here right where the first began what do you think i love that video you made it that’s so great you made it well we’re um working on industrial AI with one company after another this is uh BMW doing building their next generation factory in Omniverse this is uh I don’t know how to say it can somebody teach me sounds good um exactly that’s exactly right good job good job that’s exactly right uh they’re they’re building of course their plants uh digital twins and omniverse this is Keon their uh uh uh their digital twin for um uh warehouse logistics this is uh Mercedes-Benz and their digital twins of their factories built in Omniverse this is Schaefer and their digital twin of their warehouse built in Omniverse this is your train station here in France building a digital twin of their train stations in Omniverse and this is Toyota building a digital twin of their warehouse in Omniverse and when when you build these warehouses and these factories in Omniverse then you could you could you could you could design it you can plan it you can change it in green field is wonderful in brown field is wonderful you could simulate its effectiveness before you go and physically lift and move things around to discover it wasn’t optimal and so the ability to do everything digitally in a digital twin is incredible but the question is why does the digital twin has to look photoreal and why does it has to obey the laws of physics the reason for that is because we wanted ultimately to be a digital twin where a robot could learn how to operate as a robot and robots rely on photons for their perception system and those photons are generated through omniverse and robots needs to interact with the physical world so that it could know whether it’s doing the right things and do learn how to do it properly and so these digital twins have to look real and behave realistically okay so that’s the reason why Omniverse was built this is uh this is fantastic this is a fusion reactor digital twin incredibly complicated piece of instrument as you know and without AI the next generation fusion reactor would not be possible well we’re we’re announcing today that we are going to build the world’s first industrial AI cloud here in Europe i’m going to announce Yep these industrial AI clouds are yes a lot whole lot of computers in the cloud however its requirement its performance its safety requirement is fundamentally different and so I’m going to tell you a lot more about it on Friday i’m only teasing you part of the story today but this industrial cloud will be used for design and simulation virtual wind tunnels that you just walk into virtual wind tunnels you just move a car into and you see it behave open doors open windows change the design all completely in real time design in real time simulate in a digital wind tunnel digital twin of a wind tunnel in real time build it in a factory of a digital factory digital twin in real time all of this and let robots learn how to be great robots and build the robots of our future self-driving cars and such we already have tremendous ecosystem here we’ve been here as you know for a very long time nvidia is 33 years old the first time we came to Europe was during the time when workstations and the digitalization of products CAD the CAD revolution started we were here during the CE revolution and now the digital twin revolution some two trillion dollars of ecosystem here in Europe that we part partner with and that we have the privilege of supporting what comes out of that is a new revolution that’s happening as you know everything that moves will be robotics everything that moves will be AIdriven and the car is the most obvious next one builds the AI supercomput computers to train the model the AI supercomputer for Omniverse digital twins we also build the AI supercomputers for the robots itself in every single case whether it’s in the cloud for Omniverse or in the car we offer the entire stack the computer itself the operating system that runs on top of this computer which is different in every single case this computer high-speed sensor rich must be functional safe in no in no circumstance could it fail completely and so the safety requirements incredibly high and now we have an incredible model that sits on top of it this model that sits on top of it is a transformer model it’s a reasoning model and it takes sensor in you tell it what you want it to do and it will drive you there takes pixels in and it generates path plans output so it’s a generative AI model based on transformers incredible technology nvidia’s AI team AV team is incredible this is the only team that I know that has won endtoend self-driving car challenge at CVPR two years in a row and so they’re the winner again this year let’s take a look at the video yep thank you like any driver autonomous vehicles operate in a world full of unpredictable and potentially safety critical scenarios nvidia drive built on the Halo safety system lets developers build safe autonomous vehicles with diverse software stacks and sensors and redundant computers it starts with training safe AVs need massive amounts of diverse data to be able to address edge cases but real world data is limited developers use NVIDIA Omniverse and Cosmos to reconstruct the real world and generate realistic synthetic training data to bring diversity to the AV model the model can perceive and reason about its environment predict future outcomes and generate a motion plan and for decision-making diversity an independent classical stack runs in parallel guardrails monitor safe performance and in cases of anomalies calls the arbitrator to make an emergency stop further diversity and redundancy are built into the sensor and compute architecture each sensor connects to redundant computers so even if a sensor or computer fails the vehicle stays safe and operational and in the event of a critical failure the system can execute a minimum risk maneuver like pulling over to the shoulder safety is foundational to autonomous driving nvidia Drive lets developers worldwide integrate halos into their own products to build the next generation of safe AVs a billion cars on the road 10,000 miles a year on average 10 trillion miles the future of autonomous driving is obviously gigantic and it’s going to be driven is going to be powered by AI this is this is the next gigantic opportunity and we’re working with enormous companies and really fantastic companies around the world to make this possible at the core of everything we do here with AV is safety and we’re really really proud of our Halo system it starts with the architecture of the chip and then the chip design and the systems design the operating system the AI models the methodology of developing the software the way we test it everything from the way we train the models the data we provide for the models all the way to the way we evaluate the models nvidia’s Halo system and our AV safety team and capabilities are absolutely worldrenowned this computer was the first one to be softwaredefined the world’s first softwaredefined completely 100% softwaredefined AIdriven software AIdriven stack for AVs we’ve been at this now for coming up on 10 years and so this capability is worldrenowned and I’m really proud of it the same thing that’s happening for cars is happening for a new industry as I mentioned earlier if you can generate video from prompts if AI can perceive it can reason and it can generate videos and words and images and just now with cars the path the steering wheel path why can’t it also generate local motion abilities and articulation abilities so that fundamental ability for AI to revolutionize one of the hardest robotics problems is around the corner humano or robots are going to be a thing we now know how to build these things train these things and operate these things human or robotics is going to potentially be one of the largest industries ever and it requires companies who know how to manufacture things manufacture things of extraordinary capabilities this speaks of the European countries so much of the world’s industries are based here i think this is going to be a giant opportunity well let’s say it’s a billion robots around the world the idea that there would be a billion robots is a very sensible thing now why hasn’t it happened well the reason for that is simple today’s robots are too hard to program only the largest companies can afford to install a robot get it to teach it program it to do exactly the right things keep it sufficiently surrounded so that it’s safe that’s the reason why the world’s largest car companies all have robots they’re large enough the work is sufficiently repetitive it is the industry is at a sufficient scale that you could deploy robots into those factories almost everybody who’s a middle or small medium medium companies or mom and pop restaurants or stores or warehouses it’s impossible to have that programming capability until now we’re going to give you essentially robots where you could teach them they’ll learn from you just as we were talking about agentic AI we now have humanoid AI that can learn from your teaching using toolkits that are very s very consistent with the Nemo toolkits I I spoke about nvidia here as well is built in three layer stack we build the computer the Thor the Thor computer devkit looks a little bit like this this is a robotic computer completely self-contained dev kit sits on your desk these are all the sensors and inside is a little supercomput Thor chip just really really incredible and these Yep i could I could imagine getting one of these inserted like that okay thank you Janine and so that’s the Thor processor on top is an operating system designed for robotics and on top of that transformer models that take sensor and instructions and transforms it and generates flight or paths and motor controls for arm articulation finger articulation and of course your legs articulation now the big challenge of human robotics is the amount of data necessary to train it is very very hard to get and so the question is how do you do that well the way you solve that problem is to back in omniverse a digital twin world that obeys the laws of physics and this is an incredible piece of work that we’re doing don’t do it don’t Oh my my fault okay these are robots [Applause] we have we develop computers to simulate to train them computers to simulate them and the computer that goes inside them there’s a whole bunch of human robotics companies being built around the world they all see the great opportunity to revolutionize this new new device if you will and uh the progress is going incredibly fast and the way that they all learn is they learn in a virtual world and this virtual world has to obey the laws of physics and recently we announced a big partnership with Disney Research and Deep Mind and we’re going to work together to create the world’s most sophisticated physics simulation and I’m just trying to figure out at this point how to go to that slide teach me who’s with me this is what happens when you only rehearse once okay so this this incredible system this incredible system is where an AI learns how to be an AI let me show it to you [Music] what’s up [Music] we have a special guest [Applause] your name is Gre are are you are you a petite Garson or petite Bill okay he’s Greck is a little girl now look at this gre learned how to walk inside Omniverse obeying the laws of physics and by inside Omniverse we created hundreds of thousands of scenarios then finally when Greck learned how to operate and walk and manipulate in those environments on sand and on you know on gravel on slippery floors on concrete on carpet then when it comes when Greg comes into the physical world the physical world is just 100,0001 version of the world and so you learn how to walk in the virtual world and look at you now can you can you jump wow can you dance well I think I think um I just want to let you know I am the keynote presenter so I need you I need you to behave i need you to behave for a few seconds i need you to behave for a few Could you sit sit hey you know what we should do let’s take a picture of everybody [Music] yeah bam bam would you like to come home with me would you like to come home with me i got Yeah I know yeah I have pets they would like to have you as a pet no no you’re so smart you’re so smart well incredible right [Applause] you are the world’s best robot and someday we’ll all have one like you and they’ll follow us around but if I need if I need a glass of whiskey you’re going to have to go tell somebody else to go get me a glass of whiskey because you have no arms yeah you’re so cute okay little girl you stay here for a second let’s wrap up all right it’s very clear it’s very clear an industrial revolution has started the next the next waves of AI has started greck is a perfect example of what’s possible now with robotics the technology necessary to teach a robot to manipulate to simulate and of of course the manifestation of an incredible robot is now right in front of us we have physical robots and we have information robots we call them agents so the next wave of AI has started it’s going to require inference workloads to explode it’s basically going to go exponential the number of people that are using inference has gone from 8 million to 800 million a 100 times in just a couple of years the number the amount of prompts that the tokens generate as I mentioned earlier from a few hundred tokens to thousands of tokens and of course we use AI even more than than ever today so we need a special computer designed for thinking designed for reasoning and that’s what blackwell is a thinking machine these black wells will go into new types of data centers essentially AI factories designed for one thing and one thing only and these AI factories are going to generate tokens and these tokens are going to become your food little Greck yeah I know i know and what’s really really incredible I’m so happy to see that Europe is going allin on AI the amount of AI infrastructure being built here will increase by an order of magnitude in the next couple years i want to thank all of you for your partnership have a great Vivate thank you say bye-bye say bye-bye take a bunch of pictures take a bunch of pictures take a bunch of pictures yeah [Applause]